From 5393b281ffd25409d3a9b12598e03d1f94057121 Mon Sep 17 00:00:00 2001 From: s-mousmita Date: Sat, 23 Jul 2022 08:48:27 -0400 Subject: [PATCH 1/6] Adding fisher swbd recipe training with pruned_transducer_stateless2 --- egs/fisher_swbd/ASR/local/compile_hlg.py | 159 ++ .../compute_fbank_fisher_swbd_eval2000.py | 107 ++ .../ASR/local/compute_fbank_musan.py | 96 + .../ASR/local/extract_json_cuts.py | 72 + .../ASR/local/extract_json_supervision.py | 40 + .../ASR/local/extract_list_of_sph.py | 38 + .../normalize_and_filter_supervisions.py | 189 ++ .../ASR/local/normalize_eval2000.py | 208 +++ egs/fisher_swbd/ASR/local/prepare_lang_bpe.py | 254 +++ .../ASR/local/prepare_lang_g2pen.py | 487 +++++ egs/fisher_swbd/ASR/local/train_bpe_model.py | 98 + egs/fisher_swbd/ASR/prepare.sh | 306 ++++ .../pruned_transducer_stateless2/__init__.py | 0 .../asr_datamodule.py | 436 +++++ .../beam_search.py | 1352 ++++++++++++++ .../pruned_transducer_stateless2/conformer.py | 1594 +++++++++++++++++ .../pruned_transducer_stateless2/decode.py | 815 +++++++++ .../decode_stream.py | 126 ++ .../pruned_transducer_stateless2/decoder.py | 106 ++ .../encoder_interface.py | 43 + .../pruned_transducer_stateless2/export.py | 230 +++ .../pruned_transducer_stateless2/joiner.py | 69 + .../ASR/pruned_transducer_stateless2/model.py | 194 ++ .../ASR/pruned_transducer_stateless2/optim.py | 331 ++++ .../pretrained.py | 390 ++++ .../pruned_transducer_stateless2/scaling.py | 733 ++++++++ .../streaming_decode.py | 693 +++++++ .../test_model.py | 76 + .../ASR/pruned_transducer_stateless2/train.py | 1085 +++++++++++ egs/fisher_swbd/ASR/shared | 1 + 30 files changed, 10328 insertions(+) create mode 100755 egs/fisher_swbd/ASR/local/compile_hlg.py create mode 100755 egs/fisher_swbd/ASR/local/compute_fbank_fisher_swbd_eval2000.py create mode 100755 egs/fisher_swbd/ASR/local/compute_fbank_musan.py create mode 100644 egs/fisher_swbd/ASR/local/extract_json_cuts.py create mode 100644 egs/fisher_swbd/ASR/local/extract_json_supervision.py create mode 100644 egs/fisher_swbd/ASR/local/extract_list_of_sph.py create mode 100644 egs/fisher_swbd/ASR/local/normalize_and_filter_supervisions.py create mode 100644 egs/fisher_swbd/ASR/local/normalize_eval2000.py create mode 100755 egs/fisher_swbd/ASR/local/prepare_lang_bpe.py create mode 100755 egs/fisher_swbd/ASR/local/prepare_lang_g2pen.py create mode 100755 egs/fisher_swbd/ASR/local/train_bpe_model.py create mode 100755 egs/fisher_swbd/ASR/prepare.sh create mode 100644 egs/fisher_swbd/ASR/pruned_transducer_stateless2/__init__.py create mode 100644 egs/fisher_swbd/ASR/pruned_transducer_stateless2/asr_datamodule.py create mode 100644 egs/fisher_swbd/ASR/pruned_transducer_stateless2/beam_search.py create mode 100644 egs/fisher_swbd/ASR/pruned_transducer_stateless2/conformer.py create mode 100755 egs/fisher_swbd/ASR/pruned_transducer_stateless2/decode.py create mode 100644 egs/fisher_swbd/ASR/pruned_transducer_stateless2/decode_stream.py create mode 100644 egs/fisher_swbd/ASR/pruned_transducer_stateless2/decoder.py create mode 100644 egs/fisher_swbd/ASR/pruned_transducer_stateless2/encoder_interface.py create mode 100755 egs/fisher_swbd/ASR/pruned_transducer_stateless2/export.py create mode 100644 egs/fisher_swbd/ASR/pruned_transducer_stateless2/joiner.py create mode 100644 egs/fisher_swbd/ASR/pruned_transducer_stateless2/model.py create mode 100644 egs/fisher_swbd/ASR/pruned_transducer_stateless2/optim.py create mode 100755 egs/fisher_swbd/ASR/pruned_transducer_stateless2/pretrained.py create mode 100644 egs/fisher_swbd/ASR/pruned_transducer_stateless2/scaling.py create mode 100755 egs/fisher_swbd/ASR/pruned_transducer_stateless2/streaming_decode.py create mode 100755 egs/fisher_swbd/ASR/pruned_transducer_stateless2/test_model.py create mode 100755 egs/fisher_swbd/ASR/pruned_transducer_stateless2/train.py create mode 120000 egs/fisher_swbd/ASR/shared diff --git a/egs/fisher_swbd/ASR/local/compile_hlg.py b/egs/fisher_swbd/ASR/local/compile_hlg.py new file mode 100755 index 000000000..9a35750e0 --- /dev/null +++ b/egs/fisher_swbd/ASR/local/compile_hlg.py @@ -0,0 +1,159 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This script takes as input lang_dir and generates HLG from + + - H, the ctc topology, built from tokens contained in lang_dir/lexicon.txt + - L, the lexicon, built from lang_dir/L_disambig.pt + + Caution: We use a lexicon that contains disambiguation symbols + + - G, the LM, built from data/lm/G_3_gram.fst.txt + +The generated HLG is saved in $lang_dir/HLG.pt +""" +import argparse +import logging +from pathlib import Path + +import k2 +import torch + +from icefall.lexicon import Lexicon + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--lang-dir", + type=str, + help="""Input and output directory. + """, + ) + + return parser.parse_args() + + +def compile_HLG(lang_dir: str) -> k2.Fsa: + """ + Args: + lang_dir: + The language directory, e.g., data/lang_phone or data/lang_bpe_5000. + + Return: + An FSA representing HLG. + """ + lexicon = Lexicon(lang_dir) + max_token_id = max(lexicon.tokens) + logging.info(f"Building ctc_topo. max_token_id: {max_token_id}") + H = k2.ctc_topo(max_token_id) + L = k2.Fsa.from_dict(torch.load(f"{lang_dir}/L_disambig.pt")) + + if Path("data/lm/G_3_gram.pt").is_file(): + logging.info("Loading pre-compiled G_3_gram") + d = torch.load("data/lm/G_3_gram.pt") + G = k2.Fsa.from_dict(d) + else: + logging.info("Loading G_3_gram.fst.txt") + with open("data/lm/G_3_gram.fst.txt") as f: + G = k2.Fsa.from_openfst(f.read(), acceptor=False) + torch.save(G.as_dict(), "data/lm/G_3_gram.pt") + + first_token_disambig_id = lexicon.token_table["#0"] + first_word_disambig_id = lexicon.word_table["#0"] + + L = k2.arc_sort(L) + G = k2.arc_sort(G) + + logging.info("Intersecting L and G") + LG = k2.compose(L, G) + logging.info(f"LG shape: {LG.shape}") + + logging.info("Connecting LG") + LG = k2.connect(LG) + logging.info(f"LG shape after k2.connect: {LG.shape}") + + logging.info(type(LG.aux_labels)) + logging.info("Determinizing LG") + + LG = k2.determinize(LG) + logging.info(type(LG.aux_labels)) + + logging.info("Connecting LG after k2.determinize") + LG = k2.connect(LG) + + logging.info("Removing disambiguation symbols on LG") + + LG.labels[LG.labels >= first_token_disambig_id] = 0 + # See https://github.com/k2-fsa/k2/issues/874 + # for why we need to set LG.properties to None + LG.__dict__["_properties"] = None + + assert isinstance(LG.aux_labels, k2.RaggedTensor) + LG.aux_labels.values[LG.aux_labels.values >= first_word_disambig_id] = 0 + + LG = k2.remove_epsilon(LG) + logging.info(f"LG shape after k2.remove_epsilon: {LG.shape}") + + LG = k2.connect(LG) + LG.aux_labels = LG.aux_labels.remove_values_eq(0) + + logging.info("Arc sorting LG") + LG = k2.arc_sort(LG) + + logging.info("Composing H and LG") + # CAUTION: The name of the inner_labels is fixed + # to `tokens`. If you want to change it, please + # also change other places in icefall that are using + # it. + HLG = k2.compose(H, LG, inner_labels="tokens") + + logging.info("Connecting LG") + HLG = k2.connect(HLG) + + logging.info("Arc sorting LG") + HLG = k2.arc_sort(HLG) + logging.info(f"HLG.shape: {HLG.shape}") + + return HLG + + +def main(): + args = get_args() + lang_dir = Path(args.lang_dir) + + if (lang_dir / "HLG.pt").is_file(): + logging.info(f"{lang_dir}/HLG.pt already exists - skipping") + return + + logging.info(f"Processing {lang_dir}") + + HLG = compile_HLG(lang_dir) + logging.info(f"Saving HLG.pt to {lang_dir}") + torch.save(HLG.as_dict(), f"{lang_dir}/HLG.pt") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + + main() diff --git a/egs/fisher_swbd/ASR/local/compute_fbank_fisher_swbd_eval2000.py b/egs/fisher_swbd/ASR/local/compute_fbank_fisher_swbd_eval2000.py new file mode 100755 index 000000000..f37cabc49 --- /dev/null +++ b/egs/fisher_swbd/ASR/local/compute_fbank_fisher_swbd_eval2000.py @@ -0,0 +1,107 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the Fisher, Swbd and Eval2000 dataset. +It looks for manifests in the directory data/manifests. + +The generated fbank features are saved in data/fbank. +""" + +import logging +import os +from pathlib import Path + +import torch +from lhotse import CutSet, Fbank, FbankConfig +from lhotse.recipes.utils import read_manifests_if_cached + +from icefall.utils import get_executor + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + + +def compute_fbank_fisher_swbd_eval2000(): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + num_jobs = min(25, os.cpu_count()) + num_mel_bins = 80 + sampling_rate = 8000 + dataset_parts = ( + "swbd", + ) + + """ + dataset_parts = ( + "eval2000", + "fisher", + "swbd", + ) + """ + + test_dataset=("eval2000",) + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, output_dir=src_dir, lazy=True, suffix="jsonl" + ) + assert manifests is not None + + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins, sampling_rate=sampling_rate)) + + with get_executor() as ex: # Initialize the executor only once. + for partition, m in manifests.items(): + if (output_dir / f"cuts_{partition}.json.gz").is_file(): + logging.info(f"{partition} already exists - skipping.") + continue + logging.info(f"Processing {partition}") + cut_set = CutSet.from_manifests( + recordings=m["recordings"], + supervisions=m["supervisions"], + ) + #if "train" in partition: + if partition not in test_dataset: + logging.info(f"Adding speed perturbations to : {partition}") + cut_set = ( + cut_set + + cut_set.perturb_speed(0.9) + + cut_set.perturb_speed(1.1) + ) + cut_set = cut_set.trim_to_supervisions(keep_overlapping=False) + cut_set = cut_set.filter(lambda c: c.duration > 0.5) + cut_set = cut_set.compute_and_store_features( + extractor=extractor, + storage_path=f"{output_dir}/feats_{partition}", + # when an executor is specified, make more partitions + num_jobs=num_jobs if ex is None else 80, + executor=ex, + ) + cut_set.to_json(output_dir / f"cuts_{partition}.json.gz") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + + compute_fbank_fisher_swbd_eval2000() diff --git a/egs/fisher_swbd/ASR/local/compute_fbank_musan.py b/egs/fisher_swbd/ASR/local/compute_fbank_musan.py new file mode 100755 index 000000000..f97002b82 --- /dev/null +++ b/egs/fisher_swbd/ASR/local/compute_fbank_musan.py @@ -0,0 +1,96 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the musan dataset. +It looks for manifests in the directory data/manifests. + +The generated fbank features are saved in data/fbank. +""" + +import logging +import os +from pathlib import Path + +import torch +from lhotse import CutSet, Fbank, FbankConfig, combine +from lhotse.recipes.utils import read_manifests_if_cached + +from icefall.utils import get_executor + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + + +def compute_fbank_musan(): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + num_jobs = min(15, os.cpu_count()) + num_mel_bins = 80 + sampling_rate = 8000 + dataset_parts = ( + "music", + "speech", + "noise", + ) + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, output_dir=src_dir, lazy=True, suffix="jsonl" + ) + assert manifests is not None + + musan_cuts_path = output_dir / "cuts_musan.json.gz" + + if musan_cuts_path.is_file(): + logging.info(f"{musan_cuts_path} already exists - skipping") + return + + logging.info("Extracting features for Musan") + + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins, sampling_rate=sampling_rate)) + + with get_executor() as ex: # Initialize the executor only once. + # create chunks of Musan with duration 5 - 10 seconds + musan_cuts = ( + CutSet.from_manifests( + recordings=combine( + part["recordings"] for part in manifests.values() + ) + ) + .cut_into_windows(10.0) + .filter(lambda c: c.duration > 5) + .compute_and_store_features( + extractor=extractor, + storage_path=f"{output_dir}/feats_musan", + num_jobs=num_jobs if ex is None else 80, + executor=ex, + ) + ) + musan_cuts.to_json(musan_cuts_path) + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + compute_fbank_musan() diff --git a/egs/fisher_swbd/ASR/local/extract_json_cuts.py b/egs/fisher_swbd/ASR/local/extract_json_cuts.py new file mode 100644 index 000000000..932f14714 --- /dev/null +++ b/egs/fisher_swbd/ASR/local/extract_json_cuts.py @@ -0,0 +1,72 @@ +#!/usr/bin/env python3 + +# + + +import sys, json ; +import ntpath; + +list_of_sph = sys.argv[1]; +jsonfile = sys.argv[2]; +out_partition_json = sys.argv[3]; + + +list_of_sph=[line.rstrip('\n') for line in open(list_of_sph)] + +sph_basename_list=[] + +for f in list_of_sph: + bsname=ntpath.basename(f) + #print(bsname) + sph_basename_list.append(ntpath.basename(f)) + + +json_str=[line.rstrip('\n') for line in open(jsonfile)] +num_json = len(json_str) + +#cutid2sph=dict() +out_partition=open(out_partition_json,'w',encoding='utf-8') + +for i in range(num_json): + if json_str[i] != '': + #print(json_str[i]) + cur_json = json.loads(json_str[i]) + #print(cur_json) + cur_cutid= cur_json['id'] + cur_rec = cur_json['recording'] + cur_sources = cur_rec['sources'] + #print(cur_cutid) + #print(cur_rec) + #print(cur_sources) + for s in cur_sources: + cur_sph = s['source'] + cur_sph_basename=ntpath.basename(cur_sph) + #print(cur_sph) + #print(cur_sph_basename) + if cur_sph_basename in sph_basename_list : + out_json_line = json_str[i] + out_partition.write(out_json_line) + out_partition.write("\n") + #for keys in cur_json: + #cur_cutid= cur_json['id'] + #cur_rec = cur_json['recording_id'] + #print(cur_cutid) + + +""" + for keys in cur_json: + #print(keys) + cur_cutid= cur_json['id'] + cur_rec = cur_json['recording_id'] + print(cur_rec) + if cur_sph_basename in sph_basename_list : + out_json_line = json_str[i] + out_partition.write(out_json_line) + out_partition.write("\n") +""" + + + + + + diff --git a/egs/fisher_swbd/ASR/local/extract_json_supervision.py b/egs/fisher_swbd/ASR/local/extract_json_supervision.py new file mode 100644 index 000000000..8dd714f23 --- /dev/null +++ b/egs/fisher_swbd/ASR/local/extract_json_supervision.py @@ -0,0 +1,40 @@ +#!/usr/bin/env python3 +# + + +import sys, json ; +import ntpath; + +list_of_sph = sys.argv[1]; +jsonfile = sys.argv[2]; +out_partition_json = sys.argv[3]; + + +list_of_sph=[line.rstrip('\n') for line in open(list_of_sph)] + +sph_basename_list=[] + +for f in list_of_sph: + bsname=ntpath.basename(f) + #print(bsname) + sph_basename_list.append(ntpath.basename(f)) + + +json_str=[line.rstrip('\n') for line in open(jsonfile)] +num_json = len(json_str) + +out_partition=open(out_partition_json,'w',encoding='utf-8') + +for i in range(num_json): + if json_str[i] != '': + #print(json_str[i]) + cur_json = json.loads(json_str[i]) + #print(cur_json) + cur_rec = cur_json['recording_id'] + #print(cur_rec) + cur_sph_basename = cur_rec + ".sph" + #print(cur_sph_basename) + if cur_sph_basename in sph_basename_list : + out_json_line = json_str[i] + out_partition.write(out_json_line) + out_partition.write("\n") diff --git a/egs/fisher_swbd/ASR/local/extract_list_of_sph.py b/egs/fisher_swbd/ASR/local/extract_list_of_sph.py new file mode 100644 index 000000000..708291820 --- /dev/null +++ b/egs/fisher_swbd/ASR/local/extract_list_of_sph.py @@ -0,0 +1,38 @@ +#!/usr/bin/env python3 + +# python3 extract_list_of_sph.py dev_cuts_swbd.jsonl > data/fbank/dev_swbd_sph.list + + +import sys, json ; +inputfile = sys.argv[1] +json_str=[line.rstrip('\n') for line in open(inputfile)] +num_json = len(json_str) + +#print(num_json) +#with open(inputfile, 'r',encoding='utf-8') as Jsonfile: +# print("Converting JSON encoded data into Python dictionary") +# json_dict = json.load(Jsonfile) +# for k,v in json_dict: +# print(k,v) + + + +for i in range(num_json): + if json_str[i] != '': + #print(json_str[i]) + cur_json = json.loads(json_str[i]) + # print(cur_json) + for keys in cur_json: + #print(keys) + cur_rec = cur_json['recording'] + cur_sources = cur_rec['sources'] + #print(cur_sources) + for s in cur_sources: + cur_sph = s['source'] + print(cur_sph) + #cur_sph = cur_sources[2] + #print(cur_sph) + + + +#print(json.load(sys.stdin)['source']) diff --git a/egs/fisher_swbd/ASR/local/normalize_and_filter_supervisions.py b/egs/fisher_swbd/ASR/local/normalize_and_filter_supervisions.py new file mode 100644 index 000000000..cd65f1c86 --- /dev/null +++ b/egs/fisher_swbd/ASR/local/normalize_and_filter_supervisions.py @@ -0,0 +1,189 @@ +#!/usr/bin/env python3 + +import argparse +import re +from typing import Tuple + +from tqdm import tqdm + +from lhotse import SupervisionSet, SupervisionSegment +from lhotse.serialization import load_manifest_lazy_or_eager + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument("input_sups") + parser.add_argument("output_sups") + return parser.parse_args() + + +# fmt: off +class FisherSwbdNormalizer: + """ + Note: the functions "normalize" and "keep" implement the logic similar to + Kaldi's data prep scripts for Fisher: + https://github.com/kaldi-asr/kaldi/blob/master/egs/fisher_swbd/s5/local/fisher_data_prep.sh + and for SWBD: + https://github.com/kaldi-asr/kaldi/blob/master/egs/fisher_swbd/s5/local/swbd1_data_prep.sh + + One notable difference is that we don't change [cough], [lipsmack], etc. to [noise]. + We also don't implement all the edge cases of normalization from Kaldi + (hopefully won't make too much difference). + """ + + + def __init__(self) -> None: + + self.remove_regexp_before = re.compile( + r"|".join([ + # special symbols + r"\[\[SKIP.*\]\]", + r"\[SKIP.*\]", + r"\[PAUSE.*\]", + r"\[SILENCE\]", + r"", + r"", + ]) + ) + + # tuples of (pattern, replacement) + # note: Kaldi replaces sighs, coughs, etc with [noise]. + # We don't do that here. + # We also uppercase the text as the first operation. + self.replace_regexps: Tuple[re.Pattern, str] = [ + # SWBD: + # [LAUGHTER-STORY] -> STORY + (re.compile(r"\[LAUGHTER-(.*?)\]"), r"\1"), + # [WEA[SONABLE]-/REASONABLE] + (re.compile(r"\[\S+/(\S+)\]"), r"\1"), + # -[ADV]AN[TAGE]- -> AN + (re.compile(r"-?\[.*?\](\w+)\[.*?\]-?"), r"\1-"), + # ABSOLUTE[LY]- -> ABSOLUTE- + (re.compile(r"(\w+)\[.*?\]-?"), r"\1-"), + # [AN]Y- -> Y- + # -[AN]Y- -> Y- + (re.compile(r"-?\[.*?\](\w+)-?"), r"\1-"), + # special tokens + (re.compile(r"\[LAUGH.*?\]"), r"[LAUGHTER]"), + (re.compile(r"\[SIGH.*?\]"), r"[SIGH]"), + (re.compile(r"\[COUGH.*?\]"), r"[COUGH]"), + (re.compile(r"\[MN.*?\]"), r"[VOCALIZED-NOISE]"), + (re.compile(r"\[BREATH.*?\]"), r"[BREATH]"), + (re.compile(r"\[LIPSMACK.*?\]"), r"[LIPSMACK]"), + (re.compile(r"\[SNEEZE.*?\]"), r"[SNEEZE]"), + # abbreviations + (re.compile(r"(\w)\.(\w)\.(\w)",), r"\1 \2 \3"), + (re.compile(r"(\w)\.(\w)",), r"\1 \2"), + (re.compile(r"\._",), r" "), + (re.compile(r"_(\w)",), r"\1"), + (re.compile(r"(\w)\.s",), r"\1's"), + # words between apostrophes + (re.compile(r"'(\S*?)'"), r"\1"), + # dangling dashes (2 passes) + (re.compile(r"\s-\s"), r" "), + (re.compile(r"\s-\s"), r" "), + # special symbol with trailing dash + (re.compile(r"(\[.*?\])-"), r"\1"), + ] + + # unwanted symbols in the transcripts + self.remove_regexp_after = re.compile( + r"|".join([ + # remaining punctuation + r"\.", + r",", + r"\?", + r"{", + r"}", + r"~", + r"_\d", + ]) + ) + + self.whitespace_regexp = re.compile(r"\s+") + + def normalize(self, text: str) -> str: + text = text.upper() + + # first remove + text = self.remove_regexp_before.sub("", text) + + # then replace + for pattern, sub in self.replace_regexps: + text = pattern.sub(sub, text) + + # then remove + text = self.remove_regexp_after.sub("", text) + + # then clean up whitespace + text = self.whitespace_regexp.sub(" ", text).strip() + + return text +# fmt: on + + +def keep(sup: SupervisionSegment) -> bool: + if "((" in sup.text: + return False + + if " yes", + "[laugh] oh this is [laught] this is great [silence] yes", + "i don't kn- - know a.b.c's", + "'absolutely yes", + "absolutely' yes", + "'absolutely' yes", + "'absolutely' yes 'aight", + "ABSOLUTE[LY]", + "ABSOLUTE[LY]-", + "[AN]Y", + "[AN]Y-", + "[ADV]AN[TAGE]", + "[ADV]AN[TAGE]-", + "-[ADV]AN[TAGE]", + "-[ADV]AN[TAGE]-", + "[WEA[SONABLE]-/REASONABLE]", + "[VOCALIZED-NOISE]-", + "~BULL", + ]: + print(text) + print(normalizer.normalize(text)) + print() + + +if __name__ == "__main__": + # test() + main() diff --git a/egs/fisher_swbd/ASR/local/normalize_eval2000.py b/egs/fisher_swbd/ASR/local/normalize_eval2000.py new file mode 100644 index 000000000..244f20ea4 --- /dev/null +++ b/egs/fisher_swbd/ASR/local/normalize_eval2000.py @@ -0,0 +1,208 @@ +#!/usr/bin/env python3 + +import argparse +import re +from typing import Tuple + +from tqdm import tqdm + +from lhotse import SupervisionSet, SupervisionSegment +from lhotse.serialization import load_manifest_lazy_or_eager + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument("input_sups") + parser.add_argument("output_sups") + return parser.parse_args() + +def remove_punctutation_and_other_symbol(text:str) -> str: + text = text.replace("--"," ") + text = text.replace("//"," ") + text = text.replace("."," ") + text = text.replace("?"," ") + text = text.replace("~"," ") + text = text.replace(","," ") + text = text.replace(";"," ") + text = text.replace("("," ") + text = text.replace(")"," ") + text = text.replace("&"," ") + text = text.replace("%"," ") + text = text.replace("*"," ") + text = text.replace("{"," ") + text = text.replace("}"," ") + return text + +def eval2000_clean_eform(text: str, eform_count) -> str: + string_to_remove = [] + piece=text.split("\">") + for i in range(0,len(piece)): + s=piece[i]+"\">" + res = re.search(r'', s) + if res is not None: + res_rm= res.group(1) + string_to_remove.append(res_rm) + for p in string_to_remove: + eform_string = p + text = text.replace(eform_string, " ") + eform_1 = "" + text = text.replace(eform_2," ") + #print("TEXT final: ", text) + return text + +def replace_silphone(text: str) -> str: + text = text.replace("[/BABY CRYING]", " ") + text = text.replace("[/CHILD]" , " ") + text = text.replace("[[DISTORTED]]" , " ") + text = text.replace("[/DISTORTION]" , " ") + text = text.replace("[[DRAWN OUT]]" , " ") + text = text.replace("[[DRAWN-OUT]]" , " ") + text = text.replace("[[FAINT]]" , " ") + text = text.replace("[SMACK]" , " ") + text = text.replace("[[MUMBLES]]" , " ") + text = text.replace("[[HIGH PITCHED SQUEAKY VOICE]]" , " ") + text = text.replace("[[IN THE LAUGH]]" , "[LAUGHTER]") + text = text.replace("[[LAST WORD SPOKEN WITH A LAUGH]]" , "[LAUGHTER]") + text = text.replace("[[PART OF FIRST SYLLABLE OF PREVIOUS WORD CUT OFF]]" , " ") + text = text.replace("[[PREVIOUS WORD SPOKEN WITH A LAUGH]]" , " ") + text = text.replace("[[PREVIOUS TWO WORDS SPOKEN WHILE LAUGHING]]" , " ") + text = text.replace("[[PROLONGED]]" , " ") + text = text.replace("[/RUNNING WATER]" , " ") + text = text.replace("[[SAYS LAUGHING]]" , "[LAUGHTER]") + text = text.replace("[[SINGING]]" , " ") + text = text.replace("[[SPOKEN WHILE LAUGHING]]" , "[LAUGHTER]") + text = text.replace("[/STATIC]" , " ") + text = text.replace("['THIRTIETH' DRAWN OUT]" , " ") + text = text.replace("[/VOICES]" , " ") + text = text.replace("[[WHISPERED]]" , " ") + text = text.replace("[DISTORTION]", " ") + text = text.replace("[DISTORTION, HIGH VOLUME ON WAVES]", " ") + text = text.replace("[BACKGROUND LAUGHTER]", "[LAUGHTER]") + text = text.replace("[CHILD'S VOICE]", " ") + text = text.replace("[CHILD SCREAMS]", " ") + text = text.replace("[CHILD VOICE]", " ") + text = text.replace("[CHILD YELLING]", " ") + text = text.replace("[CHILD SCREAMING]", " ") + text = text.replace("[CHILD'S VOICE IN BACKGROUND]", " ") + text = text.replace("[CHANNEL NOISE]", " ") + text = text.replace("[CHANNEL ECHO]", " ") + text = text.replace("[ECHO FROM OTHER CHANNEL]", " ") + text = text.replace("[ECHO OF OTHER CHANNEL]", " ") + text = text.replace("[CLICK]", " ") + text = text.replace("[DISTORTED]", " ") + text = text.replace("[BABY CRYING]", " ") + text = text.replace("[METALLIC KNOCKING SOUND]", " ") + text = text.replace("[METALLIC SOUND]", " ") + + text = text.replace("[PHONE JIGGLING]", " ") + text = text.replace("[BACKGROUND SOUND]", " ") + text = text.replace("[BACKGROUND VOICE]", " ") + text = text.replace("[BACKGROUND VOICES]", " ") + text = text.replace("[BACKGROUND NOISE]", " ") + text = text.replace("[CAR HORNS IN BACKGROUND]", " ") + text = text.replace("[CAR HORNS]", " ") + text = text.replace("[CARNATING]", " ") + text = text.replace("[CRYING CHILD]", " ") + text = text.replace("[CHOPPING SOUND]", " ") + text = text.replace("[BANGING]", " ") + text = text.replace("[CLICKING NOISE]", " ") + text = text.replace("[CLATTERING]", " ") + text = text.replace("[ECHO]", " ") + text = text.replace("[KNOCK]", " ") + text = text.replace("[NOISE-GOOD]", "[NOISE]") + text = text.replace("[RIGHT]", " ") + text = text.replace("[SOUND]", " ") + text = text.replace("[SQUEAK]", " ") + text = text.replace("[STATIC]", " ") + text = text.replace("[[SAYS WITH HIGH-PITCHED SCREAMING LAUGHTER]]", " ") + text = text.replace("[UH]", "UH") + text = text.replace("[MN]", "[VOCALIZED-NOISE]") + text = text.replace("[VOICES]", " ") + text = text.replace("[WATER RUNNING]", " ") + text = text.replace("[SOUND OF TWISTING PHONE CORD]", " ") + text = text.replace("[SOUND OF SOMETHING FALLING]", " ") + text = text.replace("[SOUND]", " ") + text = text.replace("[NOISE OF MOVING PHONE]", " ") + text = text.replace("[SOUND OF RUNNING WATER]", " ") + text = text.replace("[CHANNEL]", " ") + text = text.replace("-[W]HERE", "WHERE") + text = text.replace("Y[OU]I-", "YOU I") + text = text.replace("-[A]ND", "AND") + text = text.replace("JU[ST]", "JUST") + text = text.replace("{BREATH}" , " ") + text = text.replace("{BREATHY}" , " ") + text = text.replace("{CHANNEL NOISE}" , " ") + text = text.replace("{CLEAR THROAT}" , " ") + + text = text.replace("{CLEARING THROAT}" , " ") + text = text.replace("{CLEARS THROAT}" , " ") + text = text.replace("{COUGH}" , " ") + text = text.replace("{DRAWN OUT}" , " ") + text = text.replace("{EXHALATION}" , " ") + text = text.replace("{EXHALE}" , " ") + text = text.replace("{GASP}" , " ") + text = text.replace("{HIGH SQUEAL}" , " ") + text = text.replace("{INHALE}" , " ") + text = text.replace("{LAUGH}" , "[LAUGHTER]") + text = text.replace("{LAUGH}" , "[LAUGHTER]") + text = text.replace("{LAUGH}" , "[LAUGHTER]") + text = text.replace("{LIPSMACK}" , " ") + text = text.replace("{LIPSMACK}" , " ") + + text = text.replace("{NOISE OF DISGUST}" , " ") + text = text.replace("{SIGH}" , " ") + text = text.replace("{SNIFF}" , " ") + text = text.replace("{SNORT}" , " ") + text = text.replace("{SHARP EXHALATION}" , " ") + text = text.replace("{BREATH LAUGH}" , " ") + return text + +def remove_languagetag(text:str) -> str: + langtag = re.findall(r'<(.*?)>', text) + for t in langtag: + text = text.replace(t, " ") + text = text.replace("<"," ") + text = text.replace(">"," ") + return text + +def eval2000_normalizer(text: str) -> str: + #print("TEXT original: ",text) + eform_count=text.count("contraction e_form") + #print("eform corunt:", eform_count) + if eform_count>0: + text = eval2000_clean_eform(text,eform_count) + text = text.upper() + text = remove_languagetag(text) + text = replace_silphone(text) + text = remove_punctutation_and_other_symbol(text) + text = text.replace("IGNORE_TIME_SEGMENT_IN_SCORING", " ") + text = text.replace("IGNORE_TIME_SEGMENT_SCORING", " ") + spaces = re.findall(r'\s+', text) + for sp in spaces: + text = text.replace(sp," ") + text = text.strip() + #text = self.whitespace_regexp.sub(" ", text).strip() + #print(text) + return text + + + +def main(): + args = get_args() + sups = load_manifest_lazy_or_eager(args.input_sups) + assert isinstance(sups, SupervisionSet) + + tot, skip = 0, 0 + with SupervisionSet.open_writer(args.output_sups) as writer: + for sup in tqdm(sups, desc="Normalizing supervisions"): + tot += 1 + sup.text = eval2000_normalizer(sup.text) + if not sup.text: + skip += 1 + continue + writer.write(sup) + +if __name__ == "__main__": + main() diff --git a/egs/fisher_swbd/ASR/local/prepare_lang_bpe.py b/egs/fisher_swbd/ASR/local/prepare_lang_bpe.py new file mode 100755 index 000000000..b3f012e84 --- /dev/null +++ b/egs/fisher_swbd/ASR/local/prepare_lang_bpe.py @@ -0,0 +1,254 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +# Copyright (c) 2021 Xiaomi Corporation (authors: Fangjun Kuang) + +""" + +This script takes as input `lang_dir`, which should contain:: + + - lang_dir/bpe.model, + - lang_dir/words.txt + +and generates the following files in the directory `lang_dir`: + + - lexicon.txt + - lexicon_disambig.txt + - L.pt + - L_disambig.pt + - tokens.txt +""" + +import argparse +from pathlib import Path +from typing import Dict, List, Tuple + +import k2 +import sentencepiece as spm +import torch +from prepare_lang_g2pen import ( + Lexicon, + add_disambig_symbols, + add_self_loops, + write_lexicon, + write_mapping, +) + +from icefall.utils import str2bool + + +def lexicon_to_fst_no_sil( + lexicon: Lexicon, + token2id: Dict[str, int], + word2id: Dict[str, int], + need_self_loops: bool = False, +) -> k2.Fsa: + """Convert a lexicon to an FST (in k2 format). + + Args: + lexicon: + The input lexicon. See also :func:`read_lexicon` + token2id: + A dict mapping tokens to IDs. + word2id: + A dict mapping words to IDs. + need_self_loops: + If True, add self-loop to states with non-epsilon output symbols + on at least one arc out of the state. The input label for this + self loop is `token2id["#0"]` and the output label is `word2id["#0"]`. + Returns: + Return an instance of `k2.Fsa` representing the given lexicon. + """ + loop_state = 0 # words enter and leave from here + next_state = 1 # the next un-allocated state, will be incremented as we go + + arcs = [] + + # The blank symbol is defined in local/train_bpe_model.py + assert token2id[""] == 0 + assert word2id[""] == 0 + + eps = 0 + + for word, pieces in lexicon: + assert len(pieces) > 0, f"{word} has no pronunciations" + cur_state = loop_state + + word = word2id[word] + pieces = [token2id[i] for i in pieces] + + for i in range(len(pieces) - 1): + w = word if i == 0 else eps + arcs.append([cur_state, next_state, pieces[i], w, 0]) + + cur_state = next_state + next_state += 1 + + # now for the last piece of this word + i = len(pieces) - 1 + w = word if i == 0 else eps + arcs.append([cur_state, loop_state, pieces[i], w, 0]) + + if need_self_loops: + disambig_token = token2id["#0"] + disambig_word = word2id["#0"] + arcs = add_self_loops( + arcs, + disambig_token=disambig_token, + disambig_word=disambig_word, + ) + + final_state = next_state + arcs.append([loop_state, final_state, -1, -1, 0]) + arcs.append([final_state]) + + arcs = sorted(arcs, key=lambda arc: arc[0]) + arcs = [[str(i) for i in arc] for arc in arcs] + arcs = [" ".join(arc) for arc in arcs] + arcs = "\n".join(arcs) + + fsa = k2.Fsa.from_str(arcs, acceptor=False) + return fsa + + +def generate_lexicon( + model_file: str, words: List[str] +) -> Tuple[Lexicon, Dict[str, int]]: + """Generate a lexicon from a BPE model. + + Args: + model_file: + Path to a sentencepiece model. + words: + A list of strings representing words. + Returns: + Return a tuple with two elements: + - A dict whose keys are words and values are the corresponding + word pieces. + - A dict representing the token symbol, mapping from tokens to IDs. + """ + sp = spm.SentencePieceProcessor() + sp.load(str(model_file)) + + words_pieces: List[List[str]] = sp.encode(words, out_type=str) + + lexicon = [] + for word, pieces in zip(words, words_pieces): + lexicon.append((word, pieces)) + + # The OOV word is + lexicon.append(("[UNK]", [sp.id_to_piece(sp.unk_id())])) + + token2id: Dict[str, int] = dict() + for i in range(sp.vocab_size()): + token2id[sp.id_to_piece(i)] = i + + return lexicon, token2id + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--lang-dir", + type=str, + help="""Input and output directory. + It should contain the bpe.model and words.txt + """, + ) + + parser.add_argument( + "--debug", + type=str2bool, + default=False, + help="""True for debugging, which will generate + a visualization of the lexicon FST. + + Caution: If your lexicon contains hundreds of thousands + of lines, please set it to False! + + See "test/test_bpe_lexicon.py" for usage. + """, + ) + + return parser.parse_args() + + +def main(): + args = get_args() + lang_dir = Path(args.lang_dir) + model_file = lang_dir / "bpe.model" + + word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt") + + words = word_sym_table.symbols + + excluded = ["", "!SIL", "", "[UNK]", "#0", "", ""] + for w in excluded: + if w in words: + words.remove(w) + + lexicon, token_sym_table = generate_lexicon(model_file, words) + + lexicon_disambig, max_disambig = add_disambig_symbols(lexicon) + + next_token_id = max(token_sym_table.values()) + 1 + for i in range(max_disambig + 1): + disambig = f"#{i}" + assert disambig not in token_sym_table + token_sym_table[disambig] = next_token_id + next_token_id += 1 + + word_sym_table.add("#0") + word_sym_table.add("") + word_sym_table.add("") + + write_mapping(lang_dir / "tokens.txt", token_sym_table) + + write_lexicon(lang_dir / "lexicon.txt", lexicon) + write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig) + + L = lexicon_to_fst_no_sil( + lexicon, + token2id=token_sym_table, + word2id=word_sym_table, + ) + + L_disambig = lexicon_to_fst_no_sil( + lexicon_disambig, + token2id=token_sym_table, + word2id=word_sym_table, + need_self_loops=True, + ) + torch.save(L.as_dict(), lang_dir / "L.pt") + torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt") + + if args.debug: + labels_sym = k2.SymbolTable.from_file(lang_dir / "tokens.txt") + aux_labels_sym = k2.SymbolTable.from_file(lang_dir / "words.txt") + + L.labels_sym = labels_sym + L.aux_labels_sym = aux_labels_sym + L.draw(f"{lang_dir / 'L.svg'}", title="L.pt") + + L_disambig.labels_sym = labels_sym + L_disambig.aux_labels_sym = aux_labels_sym + L_disambig.draw(f"{lang_dir / 'L_disambig.svg'}", title="L_disambig.pt") + + +if __name__ == "__main__": + main() diff --git a/egs/fisher_swbd/ASR/local/prepare_lang_g2pen.py b/egs/fisher_swbd/ASR/local/prepare_lang_g2pen.py new file mode 100755 index 000000000..0549d7306 --- /dev/null +++ b/egs/fisher_swbd/ASR/local/prepare_lang_g2pen.py @@ -0,0 +1,487 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This script takes as input a wors.txt file "data/lang_phone/words.txt" +consisting of words and their IDs and creates a lexicon with g2p_en python package +(it's CMUdict based). It also creates rest of the files typically expected in a lang +dir, including L.pt and Linv.pt. +""" +import argparse +import math +from collections import defaultdict +from pathlib import Path +from typing import Any, Dict, List, Tuple + +import k2 +import torch +from g2p_en import G2p +from tqdm import tqdm + +from icefall.lexicon import read_lexicon, write_lexicon +from icefall.utils import str2bool + +Lexicon = List[Tuple[str, List[str]]] + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--lang-dir", + type=str, + help="""Input and output directory. + It should contain a file words.txt. + Generated files by this script are saved into this directory. + """, + ) + + parser.add_argument( + "--debug", + type=str2bool, + default=False, + help="""True for debugging, which will generate + a visualization of the lexicon FST. + + Caution: If your lexicon contains hundreds of thousands + of lines, please set it to False! + """, + ) + + return parser.parse_args() + + +def get_g2p_sym2int(): + + # These symbols are removed from from g2p_en's vocabulary + excluded_symbols = [ + "", + "", + "", + "", + ] + + symbols = [p for p in sorted(G2p().phonemes) if p not in excluded_symbols] + # reserve 0 and 1 for blank and sos/eos/pad tokens + # symbols start at index 2 + sym2int = { + "": 0, + "SIL": 1, + "UNK": 2, + "LAUGHTER": 3, + "SIGH": 4, + "COUGH": 5, + "VOCALIZED-NOISE": 6, + "BREATH": 7, + "LIPSMACK": 8, + "SNEEZE": 9, + "NOISE": 10, + **{sym: idx for idx, sym in enumerate(symbols, start=11)}, + } + return sym2int + + +def write_mapping(filename: str, sym2id: Dict[str, int]) -> None: + """Write a symbol to ID mapping to a file. + + Note: + No need to implement `read_mapping` as it can be done + through :func:`k2.SymbolTable.from_file`. + + Args: + filename: + Filename to save the mapping. + sym2id: + A dict mapping symbols to IDs. + Returns: + Return None. + """ + with open(filename, "w", encoding="utf-8") as f: + for sym, i in sym2id.items(): + f.write(f"{sym} {i}\n") + + +def get_tokens(lexicon: Lexicon) -> List[str]: + """Get tokens from a lexicon. + + Args: + lexicon: + It is the return value of :func:`read_lexicon`. + Returns: + Return a list of unique tokens. + """ + ans = set() + for _, tokens in lexicon: + ans.update(tokens) + sorted_ans = sorted(list(ans)) + return sorted_ans + + +def get_words(lexicon: Lexicon) -> List[str]: + """Get words from a lexicon. + + Args: + lexicon: + It is the return value of :func:`read_lexicon`. + Returns: + Return a list of unique words. + """ + ans = set() + for word, _ in lexicon: + ans.add(word) + sorted_ans = sorted(list(ans)) + return sorted_ans + + +def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]: + """It adds pseudo-token disambiguation symbols #1, #2 and so on + at the ends of tokens to ensure that all pronunciations are different, + and that none is a prefix of another. + + See also add_lex_disambig.pl from kaldi. + + Args: + lexicon: + It is returned by :func:`read_lexicon`. + Returns: + Return a tuple with two elements: + + - The output lexicon with disambiguation symbols + - The ID of the max disambiguation symbol that appears + in the lexicon + """ + + # (1) Work out the count of each token-sequence in the + # lexicon. + count = defaultdict(int) + for _, tokens in lexicon: + count[" ".join(tokens)] += 1 + + # (2) For each left sub-sequence of each token-sequence, note down + # that it exists (for identifying prefixes of longer strings). + issubseq = defaultdict(int) + for _, tokens in lexicon: + tokens = tokens.copy() + tokens.pop() + while tokens: + issubseq[" ".join(tokens)] = 1 + tokens.pop() + + # (3) For each entry in the lexicon: + # if the token sequence is unique and is not a + # prefix of another word, no disambig symbol. + # Else output #1, or #2, #3, ... if the same token-seq + # has already been assigned a disambig symbol. + ans = [] + + # We start with #1 since #0 has its own purpose + first_allowed_disambig = 1 + max_disambig = first_allowed_disambig - 1 + last_used_disambig_symbol_of = defaultdict(int) + + for word, tokens in lexicon: + tokenseq = " ".join(tokens) + assert tokenseq != "" + if issubseq[tokenseq] == 0 and count[tokenseq] == 1: + ans.append((word, tokens)) + continue + + cur_disambig = last_used_disambig_symbol_of[tokenseq] + if cur_disambig == 0: + cur_disambig = first_allowed_disambig + else: + cur_disambig += 1 + + if cur_disambig > max_disambig: + max_disambig = cur_disambig + last_used_disambig_symbol_of[tokenseq] = cur_disambig + tokenseq += f" #{cur_disambig}" + ans.append((word, tokenseq.split())) + return ans, max_disambig + + +def generate_id_map(symbols: List[str]) -> Dict[str, int]: + """Generate ID maps, i.e., map a symbol to a unique ID. + + Args: + symbols: + A list of unique symbols. + Returns: + A dict containing the mapping between symbols and IDs. + """ + return {sym: i for i, sym in enumerate(symbols)} + + +def add_self_loops( + arcs: List[List[Any]], disambig_token: int, disambig_word: int +) -> List[List[Any]]: + """Adds self-loops to states of an FST to propagate disambiguation symbols + through it. They are added on each state with non-epsilon output symbols + on at least one arc out of the state. + + See also fstaddselfloops.pl from Kaldi. One difference is that + Kaldi uses OpenFst style FSTs and it has multiple final states. + This function uses k2 style FSTs and it does not need to add self-loops + to the final state. + + The input label of a self-loop is `disambig_token`, while the output + label is `disambig_word`. + + Args: + arcs: + A list-of-list. The sublist contains + `[src_state, dest_state, label, aux_label, score]` + disambig_token: + It is the token ID of the symbol `#0`. + disambig_word: + It is the word ID of the symbol `#0`. + + Return: + Return new `arcs` containing self-loops. + """ + states_needs_self_loops = set() + for arc in arcs: + src, dst, ilabel, olabel, score = arc + if olabel != 0: + states_needs_self_loops.add(src) + + ans = [] + for s in states_needs_self_loops: + ans.append([s, s, disambig_token, disambig_word, 0]) + + return arcs + ans + + +def lexicon_to_fst( + lexicon: Lexicon, + token2id: Dict[str, int], + word2id: Dict[str, int], + sil_token: str = "SIL", + sil_prob: float = 0.5, + need_self_loops: bool = False, +) -> k2.Fsa: + """Convert a lexicon to an FST (in k2 format) with optional silence at + the beginning and end of each word. + + Args: + lexicon: + The input lexicon. See also :func:`read_lexicon` + token2id: + A dict mapping tokens to IDs. + word2id: + A dict mapping words to IDs. + sil_token: + The silence token. + sil_prob: + The probability for adding a silence at the beginning and end + of the word. + need_self_loops: + If True, add self-loop to states with non-epsilon output symbols + on at least one arc out of the state. The input label for this + self loop is `token2id["#0"]` and the output label is `word2id["#0"]`. + Returns: + Return an instance of `k2.Fsa` representing the given lexicon. + """ + assert sil_prob > 0.0 and sil_prob < 1.0 + # CAUTION: we use score, i.e, negative cost. + sil_score = math.log(sil_prob) + no_sil_score = math.log(1.0 - sil_prob) + + start_state = 0 + loop_state = 1 # words enter and leave from here + sil_state = 2 # words terminate here when followed by silence; this state + # has a silence transition to loop_state. + next_state = 3 # the next un-allocated state, will be incremented as we go. + arcs = [] + + assert token2id[""] == 0 + assert word2id[""] == 0 + + eps = 0 + + sil_token = token2id[sil_token] + + arcs.append([start_state, loop_state, eps, eps, no_sil_score]) + arcs.append([start_state, sil_state, eps, eps, sil_score]) + arcs.append([sil_state, loop_state, sil_token, eps, 0]) + + for word, tokens in lexicon: + assert len(tokens) > 0, f"{word} has no pronunciations" + cur_state = loop_state + + word = word2id[word] + tokens = [token2id[i] for i in tokens] + + for i in range(len(tokens) - 1): + w = word if i == 0 else eps + arcs.append([cur_state, next_state, tokens[i], w, 0]) + + cur_state = next_state + next_state += 1 + + # now for the last token of this word + # It has two out-going arcs, one to the loop state, + # the other one to the sil_state. + i = len(tokens) - 1 + w = word if i == 0 else eps + arcs.append([cur_state, loop_state, tokens[i], w, no_sil_score]) + arcs.append([cur_state, sil_state, tokens[i], w, sil_score]) + + if need_self_loops: + disambig_token = token2id["#0"] + disambig_word = word2id["#0"] + arcs = add_self_loops( + arcs, + disambig_token=disambig_token, + disambig_word=disambig_word, + ) + + final_state = next_state + arcs.append([loop_state, final_state, -1, -1, 0]) + arcs.append([final_state]) + + arcs = sorted(arcs, key=lambda arc: arc[0]) + arcs = [[str(i) for i in arc] for arc in arcs] + arcs = [" ".join(arc) for arc in arcs] + arcs = "\n".join(arcs) + + fsa = k2.Fsa.from_str(arcs, acceptor=False) + return fsa + + +def main(): + args = get_args() + lang_dir = Path(args.lang_dir) + vocab_filename = lang_dir / "words.txt" + lexicon_filename = lang_dir / "lexicon.txt" + sil_token = "SIL" + sil_prob = 0.5 + special_symbols = [ + "[UNK]", + "[BREATH]", + "[COUGH]", + "[LAUGHTER]", + "[LIPSMACK]", + "[NOISE]", + "[SIGH]", + "[SNEEZE]", + "[VOCALIZED-NOISE]", + ] + + g2p = G2p() + token2id = get_g2p_sym2int() + + vocab = sorted( + [ + l.split()[0] + for l in vocab_filename.read_text().splitlines() + if l.strip() and not l.startswith(("!", "[", "<", "#")) + ] + ) + print("First ten words from the vocabulary:") + print(vocab[:10]) + + if not lexicon_filename.is_file(): + lexicon = [ + ("!SIL", [sil_token]), + ] + for symbol in special_symbols: + lexicon.append((symbol, [symbol[1:-1]])) + lexicon += [ + ( + word, + [ + phn + for phn in g2p(word) + if phn + not in ( + "'", + " ", + "-", + ",", + ) # g2p_en has these symbols as phones + ], + ) + for word in tqdm(vocab, desc="Processing vocab with G2P") + ] + lexicon = [entry for entry in lexicon if entry[1]] # filter empty prons + print(lexicon[:10]) + + write_lexicon(lexicon_filename, lexicon) + else: + lexicon = read_lexicon(lexicon_filename) + + tokens = get_tokens(lexicon) + + lexicon_disambig, max_disambig = add_disambig_symbols(lexicon) + + for i in range(max_disambig + 1): + disambig = f"#{i}" + assert disambig not in tokens + tokens.append(disambig) + token2id[disambig] = max(token2id.values()) + 1 + + print("Tokens in the lexicon:") + print(tokens) + + # sort by ID + token2id = dict(sorted(token2id.items(), key=lambda tpl: tpl[1])) + print(token2id) + word2id = {"": 0} + word2id.update( + {word: int(id_) for id_, (word, pron) in enumerate(lexicon, start=1)} + ) + for symbol in ["", "", "#0"]: + word2id[symbol] = len(word2id) + + write_mapping(lang_dir / "tokens.txt", token2id) + write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig) + + L = lexicon_to_fst( + lexicon, + token2id=token2id, + word2id=word2id, + sil_token=sil_token, + sil_prob=sil_prob, + ) + + L_disambig = lexicon_to_fst( + lexicon_disambig, + token2id=token2id, + word2id=word2id, + sil_token=sil_token, + sil_prob=sil_prob, + need_self_loops=True, + ) + torch.save(L.as_dict(), lang_dir / "L.pt") + torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt") + + if args.debug: + labels_sym = k2.SymbolTable.from_file(lang_dir / "tokens.txt") + aux_labels_sym = k2.SymbolTable.from_file(lang_dir / "words.txt") + + L.labels_sym = labels_sym + L.aux_labels_sym = aux_labels_sym + L.draw(f"{lang_dir / 'L.svg'}", title="L.pt") + + L_disambig.labels_sym = labels_sym + L_disambig.aux_labels_sym = aux_labels_sym + L_disambig.draw(f"{lang_dir / 'L_disambig.svg'}", title="L_disambig.pt") + + +if __name__ == "__main__": + main() diff --git a/egs/fisher_swbd/ASR/local/train_bpe_model.py b/egs/fisher_swbd/ASR/local/train_bpe_model.py new file mode 100755 index 000000000..bc5812810 --- /dev/null +++ b/egs/fisher_swbd/ASR/local/train_bpe_model.py @@ -0,0 +1,98 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +# You can install sentencepiece via: +# +# pip install sentencepiece +# +# Due to an issue reported in +# https://github.com/google/sentencepiece/pull/642#issuecomment-857972030 +# +# Please install a version >=0.1.96 + +import argparse +import shutil +from pathlib import Path + +import sentencepiece as spm + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--lang-dir", + type=str, + help="""Input and output directory. + It should contain the training corpus: transcript_words.txt. + The generated bpe.model is saved to this directory. + """, + ) + + parser.add_argument( + "--transcript", + type=str, + help="Training transcript.", + ) + + parser.add_argument( + "--vocab-size", + type=int, + help="Vocabulary size for BPE training", + ) + + return parser.parse_args() + + +def main(): + args = get_args() + vocab_size = args.vocab_size + lang_dir = Path(args.lang_dir) + + model_type = "unigram" + + model_prefix = f"{lang_dir}/{model_type}_{vocab_size}" + train_text = args.transcript + character_coverage = 1.0 + input_sentence_size = 100000000 + + user_defined_symbols = ["", ""] + unk_id = len(user_defined_symbols) + # Note: unk_id is fixed to 2. + # If you change it, you should also change other + # places that are using it. + + model_file = Path(model_prefix + ".model") + if not model_file.is_file(): + spm.SentencePieceTrainer.train( + input=train_text, + vocab_size=vocab_size, + model_type=model_type, + model_prefix=model_prefix, + input_sentence_size=input_sentence_size, + character_coverage=character_coverage, + user_defined_symbols=user_defined_symbols, + unk_id=unk_id, + bos_id=-1, + eos_id=-1, + ) + + shutil.copyfile(model_file, f"{lang_dir}/bpe.model") + + +if __name__ == "__main__": + main() diff --git a/egs/fisher_swbd/ASR/prepare.sh b/egs/fisher_swbd/ASR/prepare.sh new file mode 100755 index 000000000..604eda754 --- /dev/null +++ b/egs/fisher_swbd/ASR/prepare.sh @@ -0,0 +1,306 @@ +#!/usr/bin/env bash +. ./path.sh +set -eou pipefail + +nj=15 +stage=0 +stop_stage=500 + +# We assume dl_dir (download dir) contains the following +# directories and files. Most of them can't be downloaded automatically +# as they are not publically available and require a license purchased +# from the LDC. +# +# - $dl_dir/{LDC2004S13,LDC2004T19,LDC2005S13,LDC2005T19} +# Fisher LDC packages. +# +# - $dl_dir/LDC97S62 +# Switchboard LDC audio package (transcripts are auto-downloaded) +# +# - $dl_dir/{LDC2002S09,LDC2002T43} +# Eval2000 audio and transcript +# +# - $dl_dir/musan +# This directory contains the following directories downloaded from +# http://www.openslr.org/17/ +# +# - music +# - noise +# - speech + +dl_dir=$PWD/download +mkdir -p $dl_dir + +. shared/parse_options.sh || exit 1 + +# vocab size for sentence piece models. +# It will generate data/lang_bpe_xxx, +# data/lang_bpe_yyy if the array contains xxx, yyy +vocab_sizes=( + 500 +) + +# All files generated by this script are saved in "data". +# You can safely remove "data" and rerun this script to regenerate it. +mkdir -p data + +log() { + # This function is from espnet + local fname=${BASH_SOURCE[1]##*/} + echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*" +} + +log "dl_dir: $dl_dir" + +if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then + log "Stage 0: Download data" + + # If you have pre-downloaded it to /path/to/fisher and /path/to/swbd, + # you can create a symlink + # + # ln -sfv /path/to/fisher $dl_dir/fisher + # + + # TODO: remove + LDC_ROOT=/nas/data4/DATA + for pkg in LDC2004S13 LDC2004T19 LDC2005S13 LDC2005T19 LDC97S62 LDC2002S09 LDC2002T43; do + ln -sfv $LDC_ROOT/$pkg $dl_dir/ + done + + # If you have pre-downloaded it to /path/to/musan, + # you can create a symlink + # + # ln -sfv /path/to/musan $dl_dir/ + # + if [ ! -d $dl_dir/musan ]; then + lhotse download musan $dl_dir + fi +fi + +if [ $stage -le 1 ] && [ $stop_stage -ge 1 ] ; then + log "Stage 1: Prepare Fisher manifests" + mkdir -p data/manifests/fisher + lhotse prepare fisher-english --absolute-paths 1 $dl_dir data/manifests/fisher + local/normalize_and_filter_supervisions.py data/manifests/fisher/supervisions.jsonl.gz data/manifests/supervisions_fisher.jsonl.gz + cp data/manifests/fisher/recordings.jsonl.gz data/manifests/recordings_fisher.jsonl.gz + gzip -d data/manifests/supervisions_fisher.jsonl.gz + gzip -d data/manifests/recordings_fisher.jsonl.gz +fi + +if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then + log "Stage 2: Prepare SWBD manifests" + mkdir -p data/manifests/swbd + lhotse prepare switchboard --absolute-paths 1 --omit-silence $dl_dir/LDC97S62 data/manifests/swbd + gzip -d data/manifests/swbd/swbd_supervisions.jsonl > data/manifests/swbd/swbd_supervisions.jsonl.gz + python3 local/normalize_and_filter_supervisions.py data/manifests/swbd/swbd_supervisions.jsonl.gz data/manifests/supervisions_swbd.jsonl.gz + gzip -d data/manifests/supervisions_swbd.jsonl.gz + cp data/manifests/swbd/swbd_recordings.jsonl data/manifests/recordings_swbd.jsonl +fi + +if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then + mkdir -p data/manifests/eval2000 + lhotse prepare eval2000 --absolute-paths 1 $dl_dir data/manifests/eval2000 + gzip -c data/manifests/eval2000/supervisions_unnorm_eval2000.jsonl > data/manifests/eval2000/supervisions_unnorm_eval2000.jsonl.gz + python3 local/normalize_eval2000.py data/manifests/eval2000/supervisions_unnorm_eval2000.jsonl.gz data/manifests/eval2000/supervisions_eval2000.jsonl.gz + gzip -c data/manifests/eval2000/recordings_eval2000.jsonl >data/manifests/eval2000/recordings_eval2000.jsonl.gz + lhotse fix data/manifests/eval2000/recordings_eval2000.jsonl.gz data/manifests/eval2000/supervisions_eval2000.jsonl.gz data/manifests/ + gzip -d data/manifests/recordings_eval2000.jsonl.gz + gzip -d data/manifests/supervisions_eval2000.jsonl.gz +fi + +if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then + mkdir -p data/fbank + python3 local/compute_fbank_fisher_swbd_eval2000.py +fi + +if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then + ##################################### + #fisher + ##################################### + + gzip -d data/fbank/cuts_fisher.json.gz + jq -c '.[]' data/fbank/cuts_fisher.json > data/fbank/cuts_fisher.jsonl + gzip -c data/fbank/cuts_fisher.jsonl > data/fbank/cuts_fisher.jsonl.gz + + # extract list of sph + python3 local/extract_list_of_sph.py data/fbank/cuts_fisher.jsonl | sort | uniq > data/fbank/cuts_fisher_sph.list + + num_fisher_total_session=$(wc -l data/fbank/cuts_fisher_sph_dev.list + tail -n $num_fisher_train_session data/fbank/cuts_fisher_sph.list >data/fbank/cuts_fisher_sph_train.list + + # extarct dev json + python3 local/extract_json_cuts.py data/fbank/cuts_fisher_sph_dev.list data/fbank/cuts_fisher.jsonl data/fbank/dev_cuts_fisher.jsonl + gzip -c data/fbank/dev_cuts_fisher.jsonl > data/fbank/dev_cuts_fisher.jsonl.gz + + # extract train json + python3 local/extract_json_cuts.py data/fbank/cuts_fisher_sph_train.list data/fbank/cuts_fisher.jsonl data/fbank/train_cuts_fisher.jsonl + gzip -c data/fbank/train_cuts_fisher.jsonl > data/fbank/train_cuts_fisher.jsonl.gz + + # describe cut + lhotse cut describe data/fbank/train_cuts_fisher.jsonl.gz + lhotse cut describe data/fbank/dev_cuts_fisher.jsonl.gz + + # extract dev supervision + python local/extract_json_supervision.py data/fbank/cuts_fisher_sph_dev.list data/manifests/supervisions_fisher.jsonl data/manifests/dev_supervisions_fisher.jsonl + python local/extract_json_supervision.py data/fbank/cuts_fisher_sph_train.list data/manifests/supervisions_fisher.jsonl data/manifests/train_supervisions_fisher.jsonl + + ###################################### + #swbd + ###################################### + + gzip -d data/fbank/cuts_swbd.json.gz + jq -c '.[]' data/fbank/cuts_swbd.json > data/fbank/cuts_swbd.jsonl + gzip -c data/fbank/cuts_swbd.jsonl > data/fbank/cuts_swbd.jsonl.gz + + python3 local/extract_list_of_sph.py data/fbank/cuts_swbd.jsonl| sort | uniq > data/fbank/cuts_swbd_sph.list + num_swbd_total_session=$(wc -l data/fbank/cuts_swbd_sph_dev.list + tail -n $num_swbd_train_session data/fbank/cuts_swbd_sph.list >data/fbank/cuts_swbd_sph_train.list + + # extarct dev json + python3 local/extract_json_cuts.py data/fbank/cuts_swbd_sph_dev.list data/fbank/cuts_swbd.jsonl data/fbank/dev_cuts_swbd.jsonl + gzip -c data/fbank/dev_cuts_swbd.jsonl > data/fbank/dev_cuts_swbd.jsonl.gz + + python3 local/extract_json_cuts.py data/fbank/cuts_swbd_sph_train.list data/fbank/cuts_swbd.jsonl data/fbank/train_cuts_swbd.jsonl + gzip -c data/fbank/train_cuts_swbd.jsonl > data/fbank/train_cuts_swbd.jsonl.gz + + # describe cut + lhotse cut describe data/fbank/train_cuts_swbd.jsonl.gz + lhotse cut describe data/fbank/dev_cuts_swbd.jsonl.gz + + # extract dev supervision + python local/extract_json_supervision.py data/fbank/cuts_swbd_sph_dev.list data/manifests/supervisions_swbd.jsonl data/manifests/dev_supervisions_swbd.jsonl + python local/extract_json_supervision.py data/fbank/cuts_swbd_sph_train.list data/manifests/supervisions_swbd.jsonl data/manifests/train_supervisions_swbd.jsonl +fi + +if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then + log "Stage 3: Prepare musan manifest" + # We assume that you have downloaded the musan corpus + # to data/musan + mkdir -p data/manifests/musan + lhotse prepare musan $dl_dir/musan data/manifests/musan + + for name in music noise speech ; do + jq -c '.[]' data/manifests/musan/recordings_${name}.json > data/manifests/recordings_${name}.jsonl + gzip -c data/manifests/recordings_${name}.jsonl > data/manifests/recordings_${name}.jsonl.gz + done +fi + +if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then + python3 local/compute_fbank_musan.py +fi + + +if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then + log "Stage 6: Dump transcripts for LM training" + mkdir -p data/lm + cat data/manifests/supervisions_fisher.jsonl data/manifests/supervisions_swbd.jsonl \ + | jq '.text' \ + | sed 's:"::g' \ + > data/lm/transcript_words.txt + + cat data/manifests/train_supervisions_fisher.jsonl data/manifests/train_supervisions_swbd.jsonl \ + | jq '.text' \ + | sed 's:"::g' \ + > data/lm/train_transcript_words.txt + + cat data/manifests/dev_supervisions_fisher.jsonl data/manifests/dev_supervisions_swbd.jsonl \ + | jq '.text' \ + | sed 's:"::g' \ + > data/lm/dev_transcript_words.txt +fi + +if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then + log "Stage 7: Prepare lexicon using g2p_en" + lang_dir=data/lang_phone + mkdir -p $lang_dir + + # Add special words to words.txt + echo " 0" > $lang_dir/words.txt + echo "!SIL 1" >> $lang_dir/words.txt + echo "[UNK] 2" >> $lang_dir/words.txt + + # Add regular words to words.txt + #gunzip -c data/manifests/fisher-swbd_supervisions_norm.jsonl.gz \ + cat data/manifests/supervisions_fisher.jsonl data/manifests/supervisions_swbd.jsonl \ + | jq '.text' \ + | sed 's:"::g' \ + | sed 's: :\n:g' \ + | sort \ + | uniq \ + | awk '{print $0,NR+2}' \ + >> $lang_dir/words.txt + + # Add remaining special word symbols expected by LM scripts. + num_words=$(cat $lang_dir/words.txt | wc -l) + echo " ${num_words}" >> $lang_dir/words.txt + num_words=$(cat $lang_dir/words.txt | wc -l) + echo " ${num_words}" >> $lang_dir/words.txt + num_words=$(cat $lang_dir/words.txt | wc -l) + echo "#0 ${num_words}" >> $lang_dir/words.txt + + if [ ! -f $lang_dir/L_disambig.pt ]; then + # We discard SWBD's lexicon and just use g2p_en + # It was trained on CMUdict and looks it up before + # resorting to an LSTM G2P model. + pip install g2p_en + ./local/prepare_lang_g2pen.py --lang-dir $lang_dir + fi +fi + +if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then + log "Stage 8: Prepare BPE based lang" + + for vocab_size in ${vocab_sizes[@]}; do + lang_dir=data/lang_bpe_${vocab_size} + mkdir -p $lang_dir + # We reuse words.txt from phone based lexicon + # so that the two can share G.pt later. + cp data/lang_phone/words.txt $lang_dir + + ./local/train_bpe_model.py \ + --lang-dir $lang_dir \ + --vocab-size $vocab_size \ + --transcript data/lm/transcript_words.txt + + if [ ! -f $lang_dir/L_disambig.pt ]; then + ./local/prepare_lang_bpe.py --lang-dir $lang_dir + fi + done +fi + +if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then + log "Stage 9: Train LM" + lm_dir=data/lm + + if [ ! -f $lm_dir/G.arpa ]; then + ./shared/make_kn_lm.py \ + -ngram-order 3 \ + -text $lm_dir/transcript_words.txt \ + -lm $lm_dir/G.arpa + fi + + if [ ! -f $lm_dir/G_3_gram.fst.txt ]; then + python3 -m kaldilm \ + --read-symbol-table="data/lang_phone/words.txt" \ + --disambig-symbol='#0' \ + --max-order=3 \ + $lm_dir/G.arpa > $lm_dir/G_3_gram.fst.txt + fi +fi + +if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then + log "Stage 10: Compile HLG" + ./local/compile_hlg.py --lang-dir data/lang_phone + + for vocab_size in ${vocab_sizes[@]}; do + lang_dir=data/lang_bpe_${vocab_size} + ./local/compile_hlg.py --lang-dir $lang_dir + done +fi diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/__init__.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/asr_datamodule.py new file mode 100644 index 000000000..706cdf34c --- /dev/null +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/asr_datamodule.py @@ -0,0 +1,436 @@ +# Copyright 2021 Piotr Żelasko +# Copyright 2022 Xiaomi Corporation (Author: Mingshuang Luo) +# +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import argparse +import inspect +import logging +from functools import lru_cache +from pathlib import Path +from typing import Any, Dict, Optional + +import torch +from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy +from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures + CutConcatenate, + CutMix, + DynamicBucketingSampler, + K2SpeechRecognitionDataset, + PrecomputedFeatures, + SingleCutSampler, + SpecAugment, +) +from lhotse.dataset.input_strategies import ( # noqa F401 For AudioSamples + AudioSamples, + OnTheFlyFeatures, +) +from lhotse.utils import fix_random_seed +from torch.utils.data import DataLoader + +from icefall.utils import str2bool + + +class _SeedWorkers: + def __init__(self, seed: int): + self.seed = seed + + def __call__(self, worker_id: int): + fix_random_seed(self.seed + worker_id) + + +class FisherSwbdSpeechAsrDataModule: + """ + DataModule for k2 ASR experiments. + It assumes there is always one train and valid dataloader for fisher and swbd, + but there are eval2000, eval2000 swbd and eval2000 callhome partition. + + It contains all the common data pipeline modules used in ASR + experiments, e.g.: + - dynamic batch size, + - bucketing samplers, + - cut concatenation, + - augmentation, + - on-the-fly feature extraction + + This class should be derived for specific corpora used in ASR tasks. + """ + + def __init__(self, args: argparse.Namespace): + self.args = args + + @classmethod + def add_arguments(cls, parser: argparse.ArgumentParser): + group = parser.add_argument_group( + title="ASR data related options", + description="These options are used for the preparation of " + "PyTorch DataLoaders from Lhotse CutSet's -- they control the " + "effective batch sizes, sampling strategies, applied data " + "augmentations, etc.", + ) + group.add_argument( + "--manifest-dir", + type=Path, + default=Path("data/fbank"), + help="Path to directory with train/valid/test cuts.", + ) + group.add_argument( + "--max-duration", + type=int, + default=200.0, + help="Maximum pooled recordings duration (seconds) in a " + "single batch. You can reduce it if it causes CUDA OOM.", + ) + group.add_argument( + "--bucketing-sampler", + type=str2bool, + default=True, + help="When enabled, the batches will come from buckets of " + "similar duration (saves padding frames).", + ) + group.add_argument( + "--num-buckets", + type=int, + default=30, + help="The number of buckets for the DynamicBucketingSampler" + "(you might want to increase it for larger datasets).", + ) + group.add_argument( + "--concatenate-cuts", + type=str2bool, + default=False, + help="When enabled, utterances (cuts) will be concatenated " + "to minimize the amount of padding.", + ) + group.add_argument( + "--duration-factor", + type=float, + default=1.0, + help="Determines the maximum duration of a concatenated cut " + "relative to the duration of the longest cut in a batch.", + ) + group.add_argument( + "--gap", + type=float, + default=1.0, + help="The amount of padding (in seconds) inserted between " + "concatenated cuts. This padding is filled with noise when " + "noise augmentation is used.", + ) + group.add_argument( + "--on-the-fly-feats", + type=str2bool, + default=False, + help="When enabled, use on-the-fly cut mixing and feature " + "extraction. Will drop existing precomputed feature manifests " + "if available.", + ) + group.add_argument( + "--shuffle", + type=str2bool, + default=True, + help="When enabled (=default), the examples will be " + "shuffled for each epoch.", + ) + group.add_argument( + "--drop-last", + type=str2bool, + default=True, + help="Whether to drop last batch. Used by sampler.", + ) + group.add_argument( + "--return-cuts", + type=str2bool, + default=True, + help="When enabled, each batch will have the " + "field: batch['supervisions']['cut'] with the cuts that " + "were used to construct it.", + ) + + group.add_argument( + "--num-workers", + type=int, + default=2, + help="The number of training dataloader workers that " + "collect the batches.", + ) + + group.add_argument( + "--enable-spec-aug", + type=str2bool, + default=True, + help="When enabled, use SpecAugment for training dataset.", + ) + + group.add_argument( + "--spec-aug-time-warp-factor", + type=int, + default=80, + help="Used only when --enable-spec-aug is True. " + "It specifies the factor for time warping in SpecAugment. " + "Larger values mean more warping. " + "A value less than 1 means to disable time warp.", + ) + + group.add_argument( + "--enable-musan", + type=str2bool, + default=True, + help="When enabled, select noise from MUSAN and mix it" + "with training dataset. ", + ) + + group.add_argument( + "--input-strategy", + type=str, + default="PrecomputedFeatures", + help="AudioSamples or PrecomputedFeatures", + ) + + def train_dataloaders( + self, + cuts_train: CutSet, + sampler_state_dict: Optional[Dict[str, Any]] = None, + ) -> DataLoader: + """ + Args: + cuts_train: + CutSet for training. + sampler_state_dict: + The state dict for the training sampler. + """ + transforms = [] + if self.args.enable_musan: + logging.info("Enable MUSAN") + logging.info("About to get Musan cuts") + cuts_musan = load_manifest( + self.args.manifest_dir / "musan_cuts.jsonl.gz" + ) + transforms.append( + CutMix( + cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True + ) + ) + else: + logging.info("Disable MUSAN") + + if self.args.concatenate_cuts: + logging.info( + f"Using cut concatenation with duration factor " + f"{self.args.duration_factor} and gap {self.args.gap}." + ) + # Cut concatenation should be the first transform in the list, + # so that if we e.g. mix noise in, it will fill the gaps between + # different utterances. + transforms = [ + CutConcatenate( + duration_factor=self.args.duration_factor, gap=self.args.gap + ) + ] + transforms + + input_transforms = [] + if self.args.enable_spec_aug: + logging.info("Enable SpecAugment") + logging.info( + f"Time warp factor: {self.args.spec_aug_time_warp_factor}" + ) + # Set the value of num_frame_masks according to Lhotse's version. + # In different Lhotse's versions, the default of num_frame_masks is + # different. + num_frame_masks = 10 + num_frame_masks_parameter = inspect.signature( + SpecAugment.__init__ + ).parameters["num_frame_masks"] + if num_frame_masks_parameter.default == 1: + num_frame_masks = 2 + logging.info(f"Num frame mask: {num_frame_masks}") + input_transforms.append( + SpecAugment( + time_warp_factor=self.args.spec_aug_time_warp_factor, + num_frame_masks=num_frame_masks, + features_mask_size=27, + num_feature_masks=2, + frames_mask_size=100, + ) + ) + else: + logging.info("Disable SpecAugment") + + logging.info("About to create train dataset") + train = K2SpeechRecognitionDataset( + input_strategy=eval(self.args.input_strategy)(), + cut_transforms=transforms, + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + ) + + if self.args.on_the_fly_feats: + # NOTE: the PerturbSpeed transform should be added only if we + # remove it from data prep stage. + # Add on-the-fly speed perturbation; since originally it would + # have increased epoch size by 3, we will apply prob 2/3 and use + # 3x more epochs. + # Speed perturbation probably should come first before + # concatenation, but in principle the transforms order doesn't have + # to be strict (e.g. could be randomized) + # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms # noqa + # Drop feats to be on the safe side. + train = K2SpeechRecognitionDataset( + cut_transforms=transforms, + input_strategy=OnTheFlyFeatures( + Fbank(FbankConfig(num_mel_bins=80)) + ), + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + ) + + if self.args.bucketing_sampler: + logging.info("Using DynamicBucketingSampler.") + train_sampler = DynamicBucketingSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + num_buckets=self.args.num_buckets, + drop_last=self.args.drop_last, + ) + else: + logging.info("Using SingleCutSampler.") + train_sampler = SingleCutSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + ) + logging.info("About to create train dataloader") + + if sampler_state_dict is not None: + logging.info("Loading sampler state dict") + train_sampler.load_state_dict(sampler_state_dict) + + # 'seed' is derived from the current random state, which will have + # previously been set in the main process. + seed = torch.randint(0, 100000, ()).item() + worker_init_fn = _SeedWorkers(seed) + + train_dl = DataLoader( + train, + sampler=train_sampler, + batch_size=None, + num_workers=self.args.num_workers, + persistent_workers=False, + worker_init_fn=worker_init_fn, + ) + + return train_dl + + def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader: + transforms = [] + if self.args.concatenate_cuts: + transforms = [ + CutConcatenate( + duration_factor=self.args.duration_factor, gap=self.args.gap + ) + ] + transforms + + logging.info("About to create dev dataset") + if self.args.on_the_fly_feats: + validate = K2SpeechRecognitionDataset( + cut_transforms=transforms, + input_strategy=OnTheFlyFeatures( + Fbank(FbankConfig(num_mel_bins=80)) + ), + return_cuts=self.args.return_cuts, + ) + else: + validate = K2SpeechRecognitionDataset( + cut_transforms=transforms, + return_cuts=self.args.return_cuts, + ) + valid_sampler = DynamicBucketingSampler( + cuts_valid, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.info("About to create dev dataloader") + valid_dl = DataLoader( + validate, + sampler=valid_sampler, + batch_size=None, + num_workers=2, + persistent_workers=False, + ) + + return valid_dl + + def test_dataloaders(self, cuts: CutSet) -> DataLoader: + logging.debug("About to create test dataset") + test = K2SpeechRecognitionDataset( + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + if self.args.on_the_fly_feats + else eval(self.args.input_strategy)(), + return_cuts=self.args.return_cuts, + ) + sampler = DynamicBucketingSampler( + cuts, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.debug("About to create test dataloader") + test_dl = DataLoader( + test, + batch_size=None, + sampler=sampler, + num_workers=self.args.num_workers, + ) + return test_dl + + @lru_cache() + def train_fisher_cuts(self) -> CutSet: + logging.info("About to get fisher cuts") + return load_manifest_lazy( + self.args.manifest_dir / "train_cuts_fisher.jsonl.gz" + ) + @lru_cache() + def train_swbd_cuts(self) -> CutSet: + logging.info("About to get train swbd cuts") + return load_manifest_lazy( + self.args.manifest_dir / "train_cuts_swbd.jsonl.gz" + ) + @lru_cache() + def dev_fisher_cuts(self) -> CutSet: + logging.info("About to get dev fisher cuts") + return load_manifest_lazy(self.args.manifest_dir / "dev_cuts_fisher.jsonl.gz" + ) + @lru_cache() + def dev_swbd_cuts(self) -> CutSet: + logging.info("About to get dev swbd cuts") + return load_manifest_lazy(self.args.manifest_dir / "dev_cuts_swbd.jsonl.gz" + ) + @lru_cache() + def test_eval2000_cuts(self) -> CutSet: + logging.info("About to get test eval2000 cuts") + return load_manifest_lazy(self.args.manifest_dir / "cuts_eval2000.jsonl.gz" + ) + @lru_cache() + def test_swbd_cuts(self) -> CutSet: + logging.info("About to get test eval2000 swbd cuts") + return load_manifest_lazy(self.args.manifest_dir / "cuts_eval2000_swbd.jsonl.gz" + ) + @lru_cache() + def test_callhome_cuts(self) -> CutSet: + logging.info("About to get test eval2000 callhome cuts") + return load_manifest_lazy(self.args.manifest_dir / "cuts_eval2000_callhome.jsonl.gz" + ) diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/beam_search.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/beam_search.py new file mode 100644 index 000000000..ed6a6ea82 --- /dev/null +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/beam_search.py @@ -0,0 +1,1352 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import warnings +from dataclasses import dataclass +from typing import Dict, List, Optional + +import k2 +import sentencepiece as spm +import torch +from model import Transducer + +from icefall.decode import Nbest, one_best_decoding +from icefall.utils import get_texts + + +def fast_beam_search_one_best( + model: Transducer, + decoding_graph: k2.Fsa, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, + temperature: float = 1.0, +) -> List[List[int]]: + """It limits the maximum number of symbols per frame to 1. + + A lattice is first obtained using fast beam search, and then + the shortest path within the lattice is used as the final output. + + Args: + model: + An instance of `Transducer`. + decoding_graph: + Decoding graph used for decoding, may be a TrivialGraph or a HLG. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + encoder_out_lens: + A tensor of shape (N,) containing the number of frames in `encoder_out` + before padding. + beam: + Beam value, similar to the beam used in Kaldi.. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + temperature: + Softmax temperature. + Returns: + Return the decoded result. + """ + lattice = fast_beam_search( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=beam, + max_states=max_states, + max_contexts=max_contexts, + temperature=temperature, + ) + + best_path = one_best_decoding(lattice) + hyps = get_texts(best_path) + return hyps + + +def fast_beam_search_nbest_LG( + model: Transducer, + decoding_graph: k2.Fsa, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, + num_paths: int, + nbest_scale: float = 0.5, + use_double_scores: bool = True, + temperature: float = 1.0, +) -> List[List[int]]: + """It limits the maximum number of symbols per frame to 1. + + The process to get the results is: + - (1) Use fast beam search to get a lattice + - (2) Select `num_paths` paths from the lattice using k2.random_paths() + - (3) Unique the selected paths + - (4) Intersect the selected paths with the lattice and compute the + shortest path from the intersection result + - (5) The path with the largest score is used as the decoding output. + + Args: + model: + An instance of `Transducer`. + decoding_graph: + Decoding graph used for decoding, may be a TrivialGraph or a HLG. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + encoder_out_lens: + A tensor of shape (N,) containing the number of frames in `encoder_out` + before padding. + beam: + Beam value, similar to the beam used in Kaldi.. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + num_paths: + Number of paths to extract from the decoded lattice. + nbest_scale: + It's the scale applied to the lattice.scores. A smaller value + yields more unique paths. + use_double_scores: + True to use double precision for computation. False to use + single precision. + temperature: + Softmax temperature. + Returns: + Return the decoded result. + """ + lattice = fast_beam_search( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=beam, + max_states=max_states, + max_contexts=max_contexts, + temperature=temperature, + ) + + nbest = Nbest.from_lattice( + lattice=lattice, + num_paths=num_paths, + use_double_scores=use_double_scores, + nbest_scale=nbest_scale, + ) + + # The following code is modified from nbest.intersect() + word_fsa = k2.invert(nbest.fsa) + if hasattr(lattice, "aux_labels"): + # delete token IDs as it is not needed + del word_fsa.aux_labels + word_fsa.scores.zero_() + word_fsa_with_epsilon_loops = k2.linear_fsa_with_self_loops(word_fsa) + path_to_utt_map = nbest.shape.row_ids(1) + + if hasattr(lattice, "aux_labels"): + # lattice has token IDs as labels and word IDs as aux_labels. + # inv_lattice has word IDs as labels and token IDs as aux_labels + inv_lattice = k2.invert(lattice) + inv_lattice = k2.arc_sort(inv_lattice) + else: + inv_lattice = k2.arc_sort(lattice) + + if inv_lattice.shape[0] == 1: + path_lattice = k2.intersect_device( + inv_lattice, + word_fsa_with_epsilon_loops, + b_to_a_map=torch.zeros_like(path_to_utt_map), + sorted_match_a=True, + ) + else: + path_lattice = k2.intersect_device( + inv_lattice, + word_fsa_with_epsilon_loops, + b_to_a_map=path_to_utt_map, + sorted_match_a=True, + ) + + # path_lattice has word IDs as labels and token IDs as aux_labels + path_lattice = k2.top_sort(k2.connect(path_lattice)) + tot_scores = path_lattice.get_tot_scores( + use_double_scores=use_double_scores, + log_semiring=True, # Note: we always use True + ) + # See https://github.com/k2-fsa/icefall/pull/420 for why + # we always use log_semiring=True + + ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores) + best_hyp_indexes = ragged_tot_scores.argmax() + best_path = k2.index_fsa(nbest.fsa, best_hyp_indexes) + + hyps = get_texts(best_path) + + return hyps + + +def fast_beam_search_nbest( + model: Transducer, + decoding_graph: k2.Fsa, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, + num_paths: int, + nbest_scale: float = 0.5, + use_double_scores: bool = True, + temperature: float = 1.0, +) -> List[List[int]]: + """It limits the maximum number of symbols per frame to 1. + + The process to get the results is: + - (1) Use fast beam search to get a lattice + - (2) Select `num_paths` paths from the lattice using k2.random_paths() + - (3) Unique the selected paths + - (4) Intersect the selected paths with the lattice and compute the + shortest path from the intersection result + - (5) The path with the largest score is used as the decoding output. + + Args: + model: + An instance of `Transducer`. + decoding_graph: + Decoding graph used for decoding, may be a TrivialGraph or a HLG. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + encoder_out_lens: + A tensor of shape (N,) containing the number of frames in `encoder_out` + before padding. + beam: + Beam value, similar to the beam used in Kaldi.. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + num_paths: + Number of paths to extract from the decoded lattice. + nbest_scale: + It's the scale applied to the lattice.scores. A smaller value + yields more unique paths. + use_double_scores: + True to use double precision for computation. False to use + single precision. + temperature: + Softmax temperature. + Returns: + Return the decoded result. + """ + lattice = fast_beam_search( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=beam, + max_states=max_states, + max_contexts=max_contexts, + temperature=temperature, + ) + + nbest = Nbest.from_lattice( + lattice=lattice, + num_paths=num_paths, + use_double_scores=use_double_scores, + nbest_scale=nbest_scale, + ) + + # at this point, nbest.fsa.scores are all zeros. + + nbest = nbest.intersect(lattice) + # Now nbest.fsa.scores contains acoustic scores + + max_indexes = nbest.tot_scores().argmax() + + best_path = k2.index_fsa(nbest.fsa, max_indexes) + + hyps = get_texts(best_path) + + return hyps + + +def fast_beam_search_nbest_oracle( + model: Transducer, + decoding_graph: k2.Fsa, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, + num_paths: int, + ref_texts: List[List[int]], + use_double_scores: bool = True, + nbest_scale: float = 0.5, + temperature: float = 1.0, +) -> List[List[int]]: + """It limits the maximum number of symbols per frame to 1. + + A lattice is first obtained using fast beam search, and then + we select `num_paths` linear paths from the lattice. The path + that has the minimum edit distance with the given reference transcript + is used as the output. + + This is the best result we can achieve for any nbest based rescoring + methods. + + Args: + model: + An instance of `Transducer`. + decoding_graph: + Decoding graph used for decoding, may be a TrivialGraph or a HLG. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + encoder_out_lens: + A tensor of shape (N,) containing the number of frames in `encoder_out` + before padding. + beam: + Beam value, similar to the beam used in Kaldi.. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + num_paths: + Number of paths to extract from the decoded lattice. + ref_texts: + A list-of-list of integers containing the reference transcripts. + If the decoding_graph is a trivial_graph, the integer ID is the + BPE token ID. + use_double_scores: + True to use double precision for computation. False to use + single precision. + nbest_scale: + It's the scale applied to the lattice.scores. A smaller value + yields more unique paths. + temperature: + Softmax temperature. + Returns: + Return the decoded result. + """ + lattice = fast_beam_search( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=beam, + max_states=max_states, + max_contexts=max_contexts, + temperature=temperature, + ) + + nbest = Nbest.from_lattice( + lattice=lattice, + num_paths=num_paths, + use_double_scores=use_double_scores, + nbest_scale=nbest_scale, + ) + + hyps = nbest.build_levenshtein_graphs() + refs = k2.levenshtein_graph(ref_texts, device=hyps.device) + + levenshtein_alignment = k2.levenshtein_alignment( + refs=refs, + hyps=hyps, + hyp_to_ref_map=nbest.shape.row_ids(1), + sorted_match_ref=True, + ) + + tot_scores = levenshtein_alignment.get_tot_scores( + use_double_scores=False, log_semiring=False + ) + ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores) + + max_indexes = ragged_tot_scores.argmax() + + best_path = k2.index_fsa(nbest.fsa, max_indexes) + + hyps = get_texts(best_path) + return hyps + + +def fast_beam_search( + model: Transducer, + decoding_graph: k2.Fsa, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, + temperature: float = 1.0, +) -> k2.Fsa: + """It limits the maximum number of symbols per frame to 1. + + Args: + model: + An instance of `Transducer`. + decoding_graph: + Decoding graph used for decoding, may be a TrivialGraph or a HLG. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + encoder_out_lens: + A tensor of shape (N,) containing the number of frames in `encoder_out` + before padding. + beam: + Beam value, similar to the beam used in Kaldi.. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + temperature: + Softmax temperature. + Returns: + Return an FsaVec with axes [utt][state][arc] containing the decoded + lattice. Note: When the input graph is a TrivialGraph, the returned + lattice is actually an acceptor. + """ + assert encoder_out.ndim == 3 + + context_size = model.decoder.context_size + vocab_size = model.decoder.vocab_size + + B, T, C = encoder_out.shape + + config = k2.RnntDecodingConfig( + vocab_size=vocab_size, + decoder_history_len=context_size, + beam=beam, + max_contexts=max_contexts, + max_states=max_states, + ) + individual_streams = [] + for i in range(B): + individual_streams.append(k2.RnntDecodingStream(decoding_graph)) + decoding_streams = k2.RnntDecodingStreams(individual_streams, config) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + for t in range(T): + # shape is a RaggedShape of shape (B, context) + # contexts is a Tensor of shape (shape.NumElements(), context_size) + shape, contexts = decoding_streams.get_contexts() + # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64 + contexts = contexts.to(torch.int64) + # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim) + decoder_out = model.decoder(contexts, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + # current_encoder_out is of shape + # (shape.NumElements(), 1, joiner_dim) + # fmt: off + current_encoder_out = torch.index_select( + encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64) + ) + # fmt: on + logits = model.joiner( + current_encoder_out.unsqueeze(2), + decoder_out.unsqueeze(1), + project_input=False, + ) + logits = logits.squeeze(1).squeeze(1) + log_probs = (logits / temperature).log_softmax(dim=-1) + decoding_streams.advance(log_probs) + decoding_streams.terminate_and_flush_to_streams() + lattice = decoding_streams.format_output(encoder_out_lens.tolist()) + + return lattice + + +def greedy_search( + model: Transducer, encoder_out: torch.Tensor, max_sym_per_frame: int +) -> List[int]: + """Greedy search for a single utterance. + Args: + model: + An instance of `Transducer`. + encoder_out: + A tensor of shape (N, T, C) from the encoder. Support only N==1 for now. + max_sym_per_frame: + Maximum number of symbols per frame. If it is set to 0, the WER + would be 100%. + Returns: + Return the decoded result. + """ + assert encoder_out.ndim == 3 + + # support only batch_size == 1 for now + assert encoder_out.size(0) == 1, encoder_out.size(0) + + blank_id = model.decoder.blank_id + context_size = model.decoder.context_size + unk_id = getattr(model, "unk_id", blank_id) + + device = next(model.parameters()).device + + decoder_input = torch.tensor( + [blank_id] * context_size, device=device, dtype=torch.int64 + ).reshape(1, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + T = encoder_out.size(1) + t = 0 + hyp = [blank_id] * context_size + + # Maximum symbols per utterance. + max_sym_per_utt = 1000 + + # symbols per frame + sym_per_frame = 0 + + # symbols per utterance decoded so far + sym_per_utt = 0 + + while t < T and sym_per_utt < max_sym_per_utt: + if sym_per_frame >= max_sym_per_frame: + sym_per_frame = 0 + t += 1 + continue + + # fmt: off + current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2) + # fmt: on + logits = model.joiner( + current_encoder_out, decoder_out.unsqueeze(1), project_input=False + ) + # logits is (1, 1, 1, vocab_size) + + y = logits.argmax().item() + if y not in (blank_id, unk_id): + hyp.append(y) + decoder_input = torch.tensor( + [hyp[-context_size:]], device=device + ).reshape(1, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + + sym_per_utt += 1 + sym_per_frame += 1 + else: + sym_per_frame = 0 + t += 1 + hyp = hyp[context_size:] # remove blanks + + return hyp + + +def greedy_search_batch( + model: Transducer, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, +) -> List[List[int]]: + """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1. + Args: + model: + The transducer model. + encoder_out: + Output from the encoder. Its shape is (N, T, C), where N >= 1. + encoder_out_lens: + A 1-D tensor of shape (N,), containing number of valid frames in + encoder_out before padding. + Returns: + Return a list-of-list of token IDs containing the decoded results. + len(ans) equals to encoder_out.size(0). + """ + assert encoder_out.ndim == 3 + assert encoder_out.size(0) >= 1, encoder_out.size(0) + + packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence( + input=encoder_out, + lengths=encoder_out_lens.cpu(), + batch_first=True, + enforce_sorted=False, + ) + + device = next(model.parameters()).device + + blank_id = model.decoder.blank_id + unk_id = getattr(model, "unk_id", blank_id) + context_size = model.decoder.context_size + + batch_size_list = packed_encoder_out.batch_sizes.tolist() + N = encoder_out.size(0) + assert torch.all(encoder_out_lens > 0), encoder_out_lens + assert N == batch_size_list[0], (N, batch_size_list) + + hyps = [[blank_id] * context_size for _ in range(N)] + + decoder_input = torch.tensor( + hyps, + device=device, + dtype=torch.int64, + ) # (N, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + # decoder_out: (N, 1, decoder_out_dim) + + encoder_out = model.joiner.encoder_proj(packed_encoder_out.data) + + offset = 0 + for batch_size in batch_size_list: + start = offset + end = offset + batch_size + current_encoder_out = encoder_out.data[start:end] + current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1) + # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim) + offset = end + + decoder_out = decoder_out[:batch_size] + + logits = model.joiner( + current_encoder_out, decoder_out.unsqueeze(1), project_input=False + ) + # logits'shape (batch_size, 1, 1, vocab_size) + + logits = logits.squeeze(1).squeeze(1) # (batch_size, vocab_size) + assert logits.ndim == 2, logits.shape + y = logits.argmax(dim=1).tolist() + emitted = False + for i, v in enumerate(y): + if v not in (blank_id, unk_id): + hyps[i].append(v) + emitted = True + if emitted: + # update decoder output + decoder_input = [h[-context_size:] for h in hyps[:batch_size]] + decoder_input = torch.tensor( + decoder_input, + device=device, + dtype=torch.int64, + ) + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + + sorted_ans = [h[context_size:] for h in hyps] + ans = [] + unsorted_indices = packed_encoder_out.unsorted_indices.tolist() + for i in range(N): + ans.append(sorted_ans[unsorted_indices[i]]) + + return ans + + +@dataclass +class Hypothesis: + # The predicted tokens so far. + # Newly predicted tokens are appended to `ys`. + ys: List[int] + + # The log prob of ys. + # It contains only one entry. + log_prob: torch.Tensor + + @property + def key(self) -> str: + """Return a string representation of self.ys""" + return "_".join(map(str, self.ys)) + + +class HypothesisList(object): + def __init__(self, data: Optional[Dict[str, Hypothesis]] = None) -> None: + """ + Args: + data: + A dict of Hypotheses. Its key is its `value.key`. + """ + if data is None: + self._data = {} + else: + self._data = data + + @property + def data(self) -> Dict[str, Hypothesis]: + return self._data + + def add(self, hyp: Hypothesis) -> None: + """Add a Hypothesis to `self`. + + If `hyp` already exists in `self`, its probability is updated using + `log-sum-exp` with the existed one. + + Args: + hyp: + The hypothesis to be added. + """ + key = hyp.key + if key in self: + old_hyp = self._data[key] # shallow copy + torch.logaddexp( + old_hyp.log_prob, hyp.log_prob, out=old_hyp.log_prob + ) + else: + self._data[key] = hyp + + def get_most_probable(self, length_norm: bool = False) -> Hypothesis: + """Get the most probable hypothesis, i.e., the one with + the largest `log_prob`. + + Args: + length_norm: + If True, the `log_prob` of a hypothesis is normalized by the + number of tokens in it. + Returns: + Return the hypothesis that has the largest `log_prob`. + """ + if length_norm: + return max( + self._data.values(), key=lambda hyp: hyp.log_prob / len(hyp.ys) + ) + else: + return max(self._data.values(), key=lambda hyp: hyp.log_prob) + + def remove(self, hyp: Hypothesis) -> None: + """Remove a given hypothesis. + + Caution: + `self` is modified **in-place**. + + Args: + hyp: + The hypothesis to be removed from `self`. + Note: It must be contained in `self`. Otherwise, + an exception is raised. + """ + key = hyp.key + assert key in self, f"{key} does not exist" + del self._data[key] + + def filter(self, threshold: torch.Tensor) -> "HypothesisList": + """Remove all Hypotheses whose log_prob is less than threshold. + + Caution: + `self` is not modified. Instead, a new HypothesisList is returned. + + Returns: + Return a new HypothesisList containing all hypotheses from `self` + with `log_prob` being greater than the given `threshold`. + """ + ans = HypothesisList() + for _, hyp in self._data.items(): + if hyp.log_prob > threshold: + ans.add(hyp) # shallow copy + return ans + + def topk(self, k: int) -> "HypothesisList": + """Return the top-k hypothesis.""" + hyps = list(self._data.items()) + + hyps = sorted(hyps, key=lambda h: h[1].log_prob, reverse=True)[:k] + + ans = HypothesisList(dict(hyps)) + return ans + + def __contains__(self, key: str): + return key in self._data + + def __iter__(self): + return iter(self._data.values()) + + def __len__(self) -> int: + return len(self._data) + + def __str__(self) -> str: + s = [] + for key in self: + s.append(key) + return ", ".join(s) + + +def get_hyps_shape(hyps: List[HypothesisList]) -> k2.RaggedShape: + """Return a ragged shape with axes [utt][num_hyps]. + + Args: + hyps: + len(hyps) == batch_size. It contains the current hypothesis for + each utterance in the batch. + Returns: + Return a ragged shape with 2 axes [utt][num_hyps]. Note that + the shape is on CPU. + """ + num_hyps = [len(h) for h in hyps] + + # torch.cumsum() is inclusive sum, so we put a 0 at the beginning + # to get exclusive sum later. + num_hyps.insert(0, 0) + + num_hyps = torch.tensor(num_hyps) + row_splits = torch.cumsum(num_hyps, dim=0, dtype=torch.int32) + ans = k2.ragged.create_ragged_shape2( + row_splits=row_splits, cached_tot_size=row_splits[-1].item() + ) + return ans + + +def modified_beam_search( + model: Transducer, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: int = 4, + temperature: float = 1.0, +) -> List[List[int]]: + """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded. + + Args: + model: + The transducer model. + encoder_out: + Output from the encoder. Its shape is (N, T, C). + encoder_out_lens: + A 1-D tensor of shape (N,), containing number of valid frames in + encoder_out before padding. + beam: + Number of active paths during the beam search. + temperature: + Softmax temperature. + Returns: + Return a list-of-list of token IDs. ans[i] is the decoding results + for the i-th utterance. + """ + assert encoder_out.ndim == 3, encoder_out.shape + assert encoder_out.size(0) >= 1, encoder_out.size(0) + + packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence( + input=encoder_out, + lengths=encoder_out_lens.cpu(), + batch_first=True, + enforce_sorted=False, + ) + + blank_id = model.decoder.blank_id + unk_id = getattr(model, "unk_id", blank_id) + context_size = model.decoder.context_size + device = next(model.parameters()).device + + batch_size_list = packed_encoder_out.batch_sizes.tolist() + N = encoder_out.size(0) + assert torch.all(encoder_out_lens > 0), encoder_out_lens + assert N == batch_size_list[0], (N, batch_size_list) + + B = [HypothesisList() for _ in range(N)] + for i in range(N): + B[i].add( + Hypothesis( + ys=[blank_id] * context_size, + log_prob=torch.zeros(1, dtype=torch.float32, device=device), + ) + ) + + encoder_out = model.joiner.encoder_proj(packed_encoder_out.data) + + offset = 0 + finalized_B = [] + for batch_size in batch_size_list: + start = offset + end = offset + batch_size + current_encoder_out = encoder_out.data[start:end] + current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1) + # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim) + offset = end + + finalized_B = B[batch_size:] + finalized_B + B = B[:batch_size] + + hyps_shape = get_hyps_shape(B).to(device) + + A = [list(b) for b in B] + B = [HypothesisList() for _ in range(batch_size)] + + ys_log_probs = torch.cat( + [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps] + ) # (num_hyps, 1) + + decoder_input = torch.tensor( + [hyp.ys[-context_size:] for hyps in A for hyp in hyps], + device=device, + dtype=torch.int64, + ) # (num_hyps, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1) + decoder_out = model.joiner.decoder_proj(decoder_out) + # decoder_out is of shape (num_hyps, 1, 1, joiner_dim) + + # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor + # as index, so we use `to(torch.int64)` below. + current_encoder_out = torch.index_select( + current_encoder_out, + dim=0, + index=hyps_shape.row_ids(1).to(torch.int64), + ) # (num_hyps, 1, 1, encoder_out_dim) + + logits = model.joiner( + current_encoder_out, + decoder_out, + project_input=False, + ) # (num_hyps, 1, 1, vocab_size) + + logits = logits.squeeze(1).squeeze(1) # (num_hyps, vocab_size) + + log_probs = (logits / temperature).log_softmax( + dim=-1 + ) # (num_hyps, vocab_size) + + log_probs.add_(ys_log_probs) + + vocab_size = log_probs.size(-1) + + log_probs = log_probs.reshape(-1) + + row_splits = hyps_shape.row_splits(1) * vocab_size + log_probs_shape = k2.ragged.create_ragged_shape2( + row_splits=row_splits, cached_tot_size=log_probs.numel() + ) + ragged_log_probs = k2.RaggedTensor( + shape=log_probs_shape, value=log_probs + ) + + for i in range(batch_size): + topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + topk_hyp_indexes = (topk_indexes // vocab_size).tolist() + topk_token_indexes = (topk_indexes % vocab_size).tolist() + + for k in range(len(topk_hyp_indexes)): + hyp_idx = topk_hyp_indexes[k] + hyp = A[i][hyp_idx] + + new_ys = hyp.ys[:] + new_token = topk_token_indexes[k] + if new_token not in (blank_id, unk_id): + new_ys.append(new_token) + + new_log_prob = topk_log_probs[k] + new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob) + B[i].add(new_hyp) + + B = B + finalized_B + best_hyps = [b.get_most_probable(length_norm=True) for b in B] + + sorted_ans = [h.ys[context_size:] for h in best_hyps] + ans = [] + unsorted_indices = packed_encoder_out.unsorted_indices.tolist() + for i in range(N): + ans.append(sorted_ans[unsorted_indices[i]]) + + return ans + + +def _deprecated_modified_beam_search( + model: Transducer, + encoder_out: torch.Tensor, + beam: int = 4, +) -> List[int]: + """It limits the maximum number of symbols per frame to 1. + + It decodes only one utterance at a time. We keep it only for reference. + The function :func:`modified_beam_search` should be preferred as it + supports batch decoding. + + + Args: + model: + An instance of `Transducer`. + encoder_out: + A tensor of shape (N, T, C) from the encoder. Support only N==1 for now. + beam: + Beam size. + Returns: + Return the decoded result. + """ + + assert encoder_out.ndim == 3 + + # support only batch_size == 1 for now + assert encoder_out.size(0) == 1, encoder_out.size(0) + blank_id = model.decoder.blank_id + unk_id = getattr(model, "unk_id", blank_id) + context_size = model.decoder.context_size + + device = next(model.parameters()).device + + T = encoder_out.size(1) + + B = HypothesisList() + B.add( + Hypothesis( + ys=[blank_id] * context_size, + log_prob=torch.zeros(1, dtype=torch.float32, device=device), + ) + ) + encoder_out = model.joiner.encoder_proj(encoder_out) + + for t in range(T): + # fmt: off + current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2) + # current_encoder_out is of shape (1, 1, 1, encoder_out_dim) + # fmt: on + A = list(B) + B = HypothesisList() + + ys_log_probs = torch.cat([hyp.log_prob.reshape(1, 1) for hyp in A]) + # ys_log_probs is of shape (num_hyps, 1) + + decoder_input = torch.tensor( + [hyp.ys[-context_size:] for hyp in A], + device=device, + dtype=torch.int64, + ) + # decoder_input is of shape (num_hyps, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1) + decoder_out = model.joiner.decoder_proj(decoder_out) + # decoder_output is of shape (num_hyps, 1, 1, joiner_dim) + + current_encoder_out = current_encoder_out.expand( + decoder_out.size(0), 1, 1, -1 + ) # (num_hyps, 1, 1, encoder_out_dim) + + logits = model.joiner( + current_encoder_out, + decoder_out, + project_input=False, + ) + # logits is of shape (num_hyps, 1, 1, vocab_size) + logits = logits.squeeze(1).squeeze(1) + + # now logits is of shape (num_hyps, vocab_size) + log_probs = logits.log_softmax(dim=-1) + + log_probs.add_(ys_log_probs) + + log_probs = log_probs.reshape(-1) + topk_log_probs, topk_indexes = log_probs.topk(beam) + + # topk_hyp_indexes are indexes into `A` + topk_hyp_indexes = topk_indexes // logits.size(-1) + topk_token_indexes = topk_indexes % logits.size(-1) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + topk_hyp_indexes = topk_hyp_indexes.tolist() + topk_token_indexes = topk_token_indexes.tolist() + + for i in range(len(topk_hyp_indexes)): + hyp = A[topk_hyp_indexes[i]] + new_ys = hyp.ys[:] + new_token = topk_token_indexes[i] + if new_token not in (blank_id, unk_id): + new_ys.append(new_token) + new_log_prob = topk_log_probs[i] + new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob) + B.add(new_hyp) + + best_hyp = B.get_most_probable(length_norm=True) + ys = best_hyp.ys[context_size:] # [context_size:] to remove blanks + + return ys + + +def beam_search( + model: Transducer, + encoder_out: torch.Tensor, + beam: int = 4, + temperature: float = 1.0, +) -> List[int]: + """ + It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf + + espnet/nets/beam_search_transducer.py#L247 is used as a reference. + + Args: + model: + An instance of `Transducer`. + encoder_out: + A tensor of shape (N, T, C) from the encoder. Support only N==1 for now. + beam: + Beam size. + temperature: + Softmax temperature. + Returns: + Return the decoded result. + """ + assert encoder_out.ndim == 3 + + # support only batch_size == 1 for now + assert encoder_out.size(0) == 1, encoder_out.size(0) + blank_id = model.decoder.blank_id + unk_id = getattr(model, "unk_id", blank_id) + context_size = model.decoder.context_size + + device = next(model.parameters()).device + + decoder_input = torch.tensor( + [blank_id] * context_size, + device=device, + dtype=torch.int64, + ).reshape(1, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + T = encoder_out.size(1) + t = 0 + + B = HypothesisList() + B.add(Hypothesis(ys=[blank_id] * context_size, log_prob=0.0)) + + max_sym_per_utt = 20000 + + sym_per_utt = 0 + + decoder_cache: Dict[str, torch.Tensor] = {} + + while t < T and sym_per_utt < max_sym_per_utt: + # fmt: off + current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2) + # fmt: on + A = B + B = HypothesisList() + + joint_cache: Dict[str, torch.Tensor] = {} + + # TODO(fangjun): Implement prefix search to update the `log_prob` + # of hypotheses in A + + while True: + y_star = A.get_most_probable() + A.remove(y_star) + + cached_key = y_star.key + + if cached_key not in decoder_cache: + decoder_input = torch.tensor( + [y_star.ys[-context_size:]], + device=device, + dtype=torch.int64, + ).reshape(1, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + decoder_cache[cached_key] = decoder_out + else: + decoder_out = decoder_cache[cached_key] + + cached_key += f"-t-{t}" + if cached_key not in joint_cache: + logits = model.joiner( + current_encoder_out, + decoder_out.unsqueeze(1), + project_input=False, + ) + + # TODO(fangjun): Scale the blank posterior + log_prob = (logits / temperature).log_softmax(dim=-1) + # log_prob is (1, 1, 1, vocab_size) + log_prob = log_prob.squeeze() + # Now log_prob is (vocab_size,) + joint_cache[cached_key] = log_prob + else: + log_prob = joint_cache[cached_key] + + # First, process the blank symbol + skip_log_prob = log_prob[blank_id] + new_y_star_log_prob = y_star.log_prob + skip_log_prob + + # ys[:] returns a copy of ys + B.add(Hypothesis(ys=y_star.ys[:], log_prob=new_y_star_log_prob)) + + # Second, process other non-blank labels + values, indices = log_prob.topk(beam + 1) + for i, v in zip(indices.tolist(), values.tolist()): + if i in (blank_id, unk_id): + continue + new_ys = y_star.ys + [i] + new_log_prob = y_star.log_prob + v + A.add(Hypothesis(ys=new_ys, log_prob=new_log_prob)) + + # Check whether B contains more than "beam" elements more probable + # than the most probable in A + A_most_probable = A.get_most_probable() + + kept_B = B.filter(A_most_probable.log_prob) + + if len(kept_B) >= beam: + B = kept_B.topk(beam) + break + + t += 1 + + best_hyp = B.get_most_probable(length_norm=True) + ys = best_hyp.ys[context_size:] # [context_size:] to remove blanks + return ys + + +def fast_beam_search_with_nbest_rescoring( + model: Transducer, + decoding_graph: k2.Fsa, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, + ngram_lm_scale_list: List[float], + num_paths: int, + G: k2.Fsa, + sp: spm.SentencePieceProcessor, + word_table: k2.SymbolTable, + oov_word: str = "", + use_double_scores: bool = True, + nbest_scale: float = 0.5, + temperature: float = 1.0, +) -> Dict[str, List[List[int]]]: + """It limits the maximum number of symbols per frame to 1. + A lattice is first obtained using modified beam search, and then + the shortest path within the lattice is used as the final output. + Args: + model: + An instance of `Transducer`. + decoding_graph: + Decoding graph used for decoding, may be a TrivialGraph or a HLG. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + encoder_out_lens: + A tensor of shape (N,) containing the number of frames in `encoder_out` + before padding. + beam: + Beam value, similar to the beam used in Kaldi. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + ngram_lm_scale_list: + A list of floats representing LM score scales. + num_paths: + Number of paths to extract from the decoded lattice. + G: + An FsaVec containing only a single FSA. It is an n-gram LM. + sp: + The BPE model. + word_table: + The word symbol table. + oov_word: + OOV words are replaced with this word. + use_double_scores: + True to use double precision for computation. False to use + single precision. + nbest_scale: + It's the scale applied to the lattice.scores. A smaller value + yields more unique paths. + temperature: + Softmax temperature. + Returns: + Return the decoded result in a dict, where the key has the form + 'ngram_lm_scale_xx' and the value is the decoded results. `xx` is the + ngram LM scale value used during decoding, i.e., 0.1. + """ + lattice = fast_beam_search( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=beam, + max_states=max_states, + max_contexts=max_contexts, + temperature=temperature, + ) + + nbest = Nbest.from_lattice( + lattice=lattice, + num_paths=num_paths, + use_double_scores=use_double_scores, + nbest_scale=nbest_scale, + ) + # at this point, nbest.fsa.scores are all zeros. + + nbest = nbest.intersect(lattice) + # Now nbest.fsa.scores contains acoustic scores + + am_scores = nbest.tot_scores() + + # Now we need to compute the LM scores of each path. + # (1) Get the token IDs of each Path. We assume the decoding_graph + # is an acceptor, i.e., lattice is also an acceptor + tokens_shape = nbest.fsa.arcs.shape().remove_axis(1) # [path][arc] + + tokens = k2.RaggedTensor(tokens_shape, nbest.fsa.labels.contiguous()) + tokens = tokens.remove_values_leq(0) # remove -1 and 0 + + token_list: List[List[int]] = tokens.tolist() + word_list: List[List[str]] = sp.decode(token_list) + + assert isinstance(oov_word, str), oov_word + assert oov_word in word_table, oov_word + oov_word_id = word_table[oov_word] + + word_ids_list: List[List[int]] = [] + + for words in word_list: + this_word_ids = [] + for w in words.split(): + if w in word_table: + this_word_ids.append(word_table[w]) + else: + this_word_ids.append(oov_word_id) + word_ids_list.append(this_word_ids) + + word_fsas = k2.linear_fsa(word_ids_list, device=lattice.device) + word_fsas_with_self_loops = k2.add_epsilon_self_loops(word_fsas) + + num_unique_paths = len(word_ids_list) + + b_to_a_map = torch.zeros( + num_unique_paths, + dtype=torch.int32, + device=lattice.device, + ) + + rescored_word_fsas = k2.intersect_device( + a_fsas=G, + b_fsas=word_fsas_with_self_loops, + b_to_a_map=b_to_a_map, + sorted_match_a=True, + ret_arc_maps=False, + ) + + rescored_word_fsas = k2.remove_epsilon_self_loops(rescored_word_fsas) + rescored_word_fsas = k2.top_sort(k2.connect(rescored_word_fsas)) + ngram_lm_scores = rescored_word_fsas.get_tot_scores( + use_double_scores=True, + log_semiring=False, + ) + + ans: Dict[str, List[List[int]]] = {} + for s in ngram_lm_scale_list: + key = f"ngram_lm_scale_{s}" + tot_scores = am_scores.values + s * ngram_lm_scores + ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores) + max_indexes = ragged_tot_scores.argmax() + best_path = k2.index_fsa(nbest.fsa, max_indexes) + hyps = get_texts(best_path) + + ans[key] = hyps + + return ans diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/conformer.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/conformer.py new file mode 100644 index 000000000..fb8123838 --- /dev/null +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/conformer.py @@ -0,0 +1,1594 @@ +#!/usr/bin/env python3 +# Copyright (c) 2021 University of Chinese Academy of Sciences (author: Han Zhu) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import copy +import math +import warnings +from typing import List, Optional, Tuple + +import torch +from encoder_interface import EncoderInterface +from scaling import ( + ActivationBalancer, + BasicNorm, + DoubleSwish, + ScaledConv1d, + ScaledConv2d, + ScaledLinear, +) +from torch import Tensor, nn + +from icefall.utils import make_pad_mask, subsequent_chunk_mask + + +class Conformer(EncoderInterface): + """ + Args: + num_features (int): Number of input features + subsampling_factor (int): subsampling factor of encoder (the convolution layers before transformers) + d_model (int): attention dimension, also the output dimension + nhead (int): number of head + dim_feedforward (int): feedforward dimention + num_encoder_layers (int): number of encoder layers + dropout (float): dropout rate + layer_dropout (float): layer-dropout rate. + cnn_module_kernel (int): Kernel size of convolution module + vgg_frontend (bool): whether to use vgg frontend. + dynamic_chunk_training (bool): whether to use dynamic chunk training, if + you want to train a streaming model, this is expected to be True. + When setting True, it will use a masking strategy to make the attention + see only limited left and right context. + short_chunk_threshold (float): a threshold to determinize the chunk size + to be used in masking training, if the randomly generated chunk size + is greater than ``max_len * short_chunk_threshold`` (max_len is the + max sequence length of current batch) then it will use + full context in training (i.e. with chunk size equals to max_len). + This will be used only when dynamic_chunk_training is True. + short_chunk_size (int): see docs above, if the randomly generated chunk + size equals to or less than ``max_len * short_chunk_threshold``, the + chunk size will be sampled uniformly from 1 to short_chunk_size. + This also will be used only when dynamic_chunk_training is True. + num_left_chunks (int): the left context (in chunks) attention can see, the + chunk size is decided by short_chunk_threshold and short_chunk_size. + A minus value means seeing full left context. + This also will be used only when dynamic_chunk_training is True. + causal (bool): Whether to use causal convolution in conformer encoder + layer. This MUST be True when using dynamic_chunk_training. + """ + + def __init__( + self, + num_features: int, + subsampling_factor: int = 4, + d_model: int = 256, + nhead: int = 4, + dim_feedforward: int = 2048, + num_encoder_layers: int = 12, + dropout: float = 0.1, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 31, + dynamic_chunk_training: bool = False, + short_chunk_threshold: float = 0.75, + short_chunk_size: int = 25, + num_left_chunks: int = -1, + causal: bool = False, + ) -> None: + super(Conformer, self).__init__() + + self.num_features = num_features + self.subsampling_factor = subsampling_factor + if subsampling_factor != 4: + raise NotImplementedError("Support only 'subsampling_factor=4'.") + + # self.encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, T//subsampling_factor, d_model). + # That is, it does two things simultaneously: + # (1) subsampling: T -> T//subsampling_factor + # (2) embedding: num_features -> d_model + self.encoder_embed = Conv2dSubsampling(num_features, d_model) + + self.encoder_layers = num_encoder_layers + self.d_model = d_model + self.cnn_module_kernel = cnn_module_kernel + self.causal = causal + self.dynamic_chunk_training = dynamic_chunk_training + self.short_chunk_threshold = short_chunk_threshold + self.short_chunk_size = short_chunk_size + self.num_left_chunks = num_left_chunks + + self.encoder_pos = RelPositionalEncoding(d_model, dropout) + + encoder_layer = ConformerEncoderLayer( + d_model, + nhead, + dim_feedforward, + dropout, + layer_dropout, + cnn_module_kernel, + causal, + ) + self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers) + self._init_state: List[torch.Tensor] = [torch.empty(0)] + + def forward( + self, x: torch.Tensor, x_lens: torch.Tensor, warmup: float = 1.0 + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Args: + x: + The input tensor. Its shape is (batch_size, seq_len, feature_dim). + x_lens: + A tensor of shape (batch_size,) containing the number of frames in + `x` before padding. + warmup: + A floating point value that gradually increases from 0 throughout + training; when it is >= 1.0 we are "fully warmed up". It is used + to turn modules on sequentially. + Returns: + Return a tuple containing 2 tensors: + - embeddings: its shape is (batch_size, output_seq_len, d_model) + - lengths, a tensor of shape (batch_size,) containing the number + of frames in `embeddings` before padding. + """ + x = self.encoder_embed(x) + x, pos_emb = self.encoder_pos(x) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + # Caution: We assume the subsampling factor is 4! + + # lengths = ((x_lens - 1) // 2 - 1) // 2 # issue an warning + # + # Note: rounding_mode in torch.div() is available only in torch >= 1.8.0 + lengths = (((x_lens - 1) >> 1) - 1) >> 1 + + assert x.size(0) == lengths.max().item() + + src_key_padding_mask = make_pad_mask(lengths) + + if self.dynamic_chunk_training: + assert ( + self.causal + ), "Causal convolution is required for streaming conformer." + max_len = x.size(0) + chunk_size = torch.randint(1, max_len, (1,)).item() + if chunk_size > (max_len * self.short_chunk_threshold): + chunk_size = max_len + else: + chunk_size = chunk_size % self.short_chunk_size + 1 + + mask = ~subsequent_chunk_mask( + size=x.size(0), + chunk_size=chunk_size, + num_left_chunks=self.num_left_chunks, + device=x.device, + ) + x = self.encoder( + x, + pos_emb, + mask=mask, + src_key_padding_mask=src_key_padding_mask, + warmup=warmup, + ) # (T, N, C) + else: + x = self.encoder( + x, + pos_emb, + mask=None, + src_key_padding_mask=src_key_padding_mask, + warmup=warmup, + ) # (T, N, C) + + x = x.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + return x, lengths + + @torch.jit.export + def get_init_state( + self, left_context: int, device: torch.device + ) -> List[torch.Tensor]: + """Return the initial cache state of the model. + + Args: + left_context: The left context size (in frames after subsampling). + + Returns: + Return the initial state of the model, it is a list containing two + tensors, the first one is the cache for attentions which has a shape + of (num_encoder_layers, left_context, encoder_dim), the second one + is the cache of conv_modules which has a shape of + (num_encoder_layers, cnn_module_kernel - 1, encoder_dim). + + NOTE: the returned tensors are on the given device. + """ + if ( + len(self._init_state) == 2 + and self._init_state[0].size(1) == left_context + ): + # Note: It is OK to share the init state as it is + # not going to be modified by the model + return self._init_state + + init_states: List[torch.Tensor] = [ + torch.zeros( + ( + self.encoder_layers, + left_context, + self.d_model, + ), + device=device, + ), + torch.zeros( + ( + self.encoder_layers, + self.cnn_module_kernel - 1, + self.d_model, + ), + device=device, + ), + ] + + self._init_state = init_states + + return init_states + + @torch.jit.export + def streaming_forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + states: Optional[List[Tensor]] = None, + processed_lens: Optional[Tensor] = None, + left_context: int = 64, + right_context: int = 4, + chunk_size: int = 16, + simulate_streaming: bool = False, + warmup: float = 1.0, + ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]: + """ + Args: + x: + The input tensor. Its shape is (batch_size, seq_len, feature_dim). + x_lens: + A tensor of shape (batch_size,) containing the number of frames in + `x` before padding. + states: + The decode states for previous frames which contains the cached data. + It has two elements, the first element is the attn_cache which has + a shape of (encoder_layers, left_context, batch, attention_dim), + the second element is the conv_cache which has a shape of + (encoder_layers, cnn_module_kernel-1, batch, conv_dim). + Note: states will be modified in this function. + processed_lens: + How many frames (after subsampling) have been processed for each sequence. + left_context: + How many previous frames the attention can see in current chunk. + Note: It's not that each individual frame has `left_context` frames + of left context, some have more. + right_context: + How many future frames the attention can see in current chunk. + Note: It's not that each individual frame has `right_context` frames + of right context, some have more. + chunk_size: + The chunk size for decoding, this will be used to simulate streaming + decoding using masking. + simulate_streaming: + If setting True, it will use a masking strategy to simulate streaming + fashion (i.e. every chunk data only see limited left context and + right context). The whole sequence is supposed to be send at a time + When using simulate_streaming. + warmup: + A floating point value that gradually increases from 0 throughout + training; when it is >= 1.0 we are "fully warmed up". It is used + to turn modules on sequentially. + Returns: + Return a tuple containing 2 tensors: + - logits, its shape is (batch_size, output_seq_len, output_dim) + - logit_lens, a tensor of shape (batch_size,) containing the number + of frames in `logits` before padding. + - decode_states, the updated states including the information + of current chunk. + """ + + # x: [N, T, C] + # Caution: We assume the subsampling factor is 4! + + # lengths = ((x_lens - 1) // 2 - 1) // 2 # issue an warning + # + # Note: rounding_mode in torch.div() is available only in torch >= 1.8.0 + lengths = (((x_lens - 1) >> 1) - 1) >> 1 + + if not simulate_streaming: + assert states is not None + assert processed_lens is not None + assert ( + len(states) == 2 + and states[0].shape + == (self.encoder_layers, left_context, x.size(0), self.d_model) + and states[1].shape + == ( + self.encoder_layers, + self.cnn_module_kernel - 1, + x.size(0), + self.d_model, + ) + ), f"""The length of states MUST be equal to 2, and the shape of + first element should be {(self.encoder_layers, left_context, x.size(0), self.d_model)}, + given {states[0].shape}. the shape of second element should be + {(self.encoder_layers, self.cnn_module_kernel - 1, x.size(0), self.d_model)}, + given {states[1].shape}.""" + + lengths -= 2 # we will cut off 1 frame on each side of encoder_embed output + + src_key_padding_mask = make_pad_mask(lengths) + + processed_mask = torch.arange(left_context, device=x.device).expand( + x.size(0), left_context + ) + processed_lens = processed_lens.view(x.size(0), 1) + processed_mask = (processed_lens <= processed_mask).flip(1) + + src_key_padding_mask = torch.cat( + [processed_mask, src_key_padding_mask], dim=1 + ) + + embed = self.encoder_embed(x) + + # cut off 1 frame on each size of embed as they see the padding + # value which causes a training and decoding mismatch. + embed = embed[:, 1:-1, :] + + embed, pos_enc = self.encoder_pos(embed, left_context) + embed = embed.permute(1, 0, 2) # (B, T, F) -> (T, B, F) + + x, states = self.encoder.chunk_forward( + embed, + pos_enc, + src_key_padding_mask=src_key_padding_mask, + warmup=warmup, + states=states, + left_context=left_context, + right_context=right_context, + ) # (T, B, F) + if right_context > 0: + x = x[0:-right_context, ...] + lengths -= right_context + else: + assert states is None + states = [] # just to make torch.script.jit happy + # this branch simulates streaming decoding using mask as we are + # using in training time. + src_key_padding_mask = make_pad_mask(lengths) + x = self.encoder_embed(x) + x, pos_emb = self.encoder_pos(x) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + assert x.size(0) == lengths.max().item() + + num_left_chunks = -1 + if left_context >= 0: + assert left_context % chunk_size == 0 + num_left_chunks = left_context // chunk_size + + mask = ~subsequent_chunk_mask( + size=x.size(0), + chunk_size=chunk_size, + num_left_chunks=num_left_chunks, + device=x.device, + ) + x = self.encoder( + x, + pos_emb, + mask=mask, + src_key_padding_mask=src_key_padding_mask, + warmup=warmup, + ) # (T, N, C) + + x = x.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + return x, lengths, states + + +class ConformerEncoderLayer(nn.Module): + """ + ConformerEncoderLayer is made up of self-attn, feedforward and convolution networks. + See: "Conformer: Convolution-augmented Transformer for Speech Recognition" + + Args: + d_model: the number of expected features in the input (required). + nhead: the number of heads in the multiheadattention models (required). + dim_feedforward: the dimension of the feedforward network model (default=2048). + dropout: the dropout value (default=0.1). + cnn_module_kernel (int): Kernel size of convolution module. + causal (bool): Whether to use causal convolution in conformer encoder + layer. This MUST be True when using dynamic_chunk_training and streaming decoding. + + Examples:: + >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8) + >>> src = torch.rand(10, 32, 512) + >>> pos_emb = torch.rand(32, 19, 512) + >>> out = encoder_layer(src, pos_emb) + """ + + def __init__( + self, + d_model: int, + nhead: int, + dim_feedforward: int = 2048, + dropout: float = 0.1, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 31, + causal: bool = False, + ) -> None: + super(ConformerEncoderLayer, self).__init__() + + self.layer_dropout = layer_dropout + + self.d_model = d_model + + self.self_attn = RelPositionMultiheadAttention( + d_model, nhead, dropout=0.0 + ) + + self.feed_forward = nn.Sequential( + ScaledLinear(d_model, dim_feedforward), + ActivationBalancer(channel_dim=-1), + DoubleSwish(), + nn.Dropout(dropout), + ScaledLinear(dim_feedforward, d_model, initial_scale=0.25), + ) + + self.feed_forward_macaron = nn.Sequential( + ScaledLinear(d_model, dim_feedforward), + ActivationBalancer(channel_dim=-1), + DoubleSwish(), + nn.Dropout(dropout), + ScaledLinear(dim_feedforward, d_model, initial_scale=0.25), + ) + + self.conv_module = ConvolutionModule( + d_model, cnn_module_kernel, causal=causal + ) + + self.norm_final = BasicNorm(d_model) + + # try to ensure the output is close to zero-mean (or at least, zero-median). + self.balancer = ActivationBalancer( + channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0 + ) + + self.dropout = nn.Dropout(dropout) + + def forward( + self, + src: Tensor, + pos_emb: Tensor, + src_mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + warmup: float = 1.0, + ) -> Tensor: + """ + Pass the input through the encoder layer. + + Args: + src: the sequence to the encoder layer (required). + pos_emb: Positional embedding tensor (required). + src_mask: the mask for the src sequence (optional). + src_key_padding_mask: the mask for the src keys per batch (optional). + warmup: controls selective bypass of of layers; if < 1.0, we will + bypass layers more frequently. + Shape: + src: (S, N, E). + pos_emb: (N, 2*S-1, E) + src_mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, N is the batch size, E is the feature number + """ + src_orig = src + + warmup_scale = min(0.1 + warmup, 1.0) + # alpha = 1.0 means fully use this encoder layer, 0.0 would mean + # completely bypass it. + if self.training: + alpha = ( + warmup_scale + if torch.rand(()).item() <= (1.0 - self.layer_dropout) + else 0.1 + ) + else: + alpha = 1.0 + + # macaron style feed forward module + src = src + self.dropout(self.feed_forward_macaron(src)) + + # multi-headed self-attention module + src_att = self.self_attn( + src, + src, + src, + pos_emb=pos_emb, + attn_mask=src_mask, + key_padding_mask=src_key_padding_mask, + )[0] + + src = src + self.dropout(src_att) + + # convolution module + conv, _ = self.conv_module(src) + src = src + self.dropout(conv) + + # feed forward module + src = src + self.dropout(self.feed_forward(src)) + + src = self.norm_final(self.balancer(src)) + + if alpha != 1.0: + src = alpha * src + (1 - alpha) * src_orig + + return src + + @torch.jit.export + def chunk_forward( + self, + src: Tensor, + pos_emb: Tensor, + states: List[Tensor], + src_mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + warmup: float = 1.0, + left_context: int = 0, + right_context: int = 0, + ) -> Tuple[Tensor, List[Tensor]]: + """ + Pass the input through the encoder layer. + + Args: + src: the sequence to the encoder layer (required). + pos_emb: Positional embedding tensor (required). + states: + The decode states for previous frames which contains the cached data. + It has two elements, the first element is the attn_cache which has + a shape of (left_context, batch, attention_dim), + the second element is the conv_cache which has a shape of + (cnn_module_kernel-1, batch, conv_dim). + Note: states will be modified in this function. + src_mask: the mask for the src sequence (optional). + src_key_padding_mask: the mask for the src keys per batch (optional). + warmup: controls selective bypass of of layers; if < 1.0, we will + bypass layers more frequently. + left_context: + How many previous frames the attention can see in current chunk. + Note: It's not that each individual frame has `left_context` frames + of left context, some have more. + right_context: + How many future frames the attention can see in current chunk. + Note: It's not that each individual frame has `right_context` frames + of right context, some have more. + + Shape: + src: (S, N, E). + pos_emb: (N, 2*(S+left_context)-1, E). + src_mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, N is the batch size, E is the feature number + """ + + assert not self.training + assert len(states) == 2 + assert states[0].shape == (left_context, src.size(1), src.size(2)) + + # macaron style feed forward module + src = src + self.dropout(self.feed_forward_macaron(src)) + + # We put the attention cache this level (i.e. before linear transformation) + # to save memory consumption, when decoding in streaming fashion, the + # batch size would be thousands (for 32GB machine), if we cache key & val + # separately, it needs extra several GB memory. + # TODO(WeiKang): Move cache to self_attn level (i.e. cache key & val + # separately) if needed. + key = torch.cat([states[0], src], dim=0) + val = key + if right_context > 0: + states[0] = key[ + -(left_context + right_context) : -right_context, ... # noqa + ] + else: + states[0] = key[-left_context:, ...] + + # multi-headed self-attention module + src_att = self.self_attn( + src, + key, + val, + pos_emb=pos_emb, + attn_mask=src_mask, + key_padding_mask=src_key_padding_mask, + left_context=left_context, + )[0] + + src = src + self.dropout(src_att) + + # convolution module + conv, conv_cache = self.conv_module(src, states[1], right_context) + states[1] = conv_cache + + src = src + self.dropout(conv) + + # feed forward module + src = src + self.dropout(self.feed_forward(src)) + + src = self.norm_final(self.balancer(src)) + + return src, states + + +class ConformerEncoder(nn.Module): + r"""ConformerEncoder is a stack of N encoder layers + + Args: + encoder_layer: an instance of the ConformerEncoderLayer() class (required). + num_layers: the number of sub-encoder-layers in the encoder (required). + + Examples:: + >>> encoder_layer = ConformerEncoderLayer(d_model=512, nhead=8) + >>> conformer_encoder = ConformerEncoder(encoder_layer, num_layers=6) + >>> src = torch.rand(10, 32, 512) + >>> pos_emb = torch.rand(32, 19, 512) + >>> out = conformer_encoder(src, pos_emb) + """ + + def __init__(self, encoder_layer: nn.Module, num_layers: int) -> None: + super().__init__() + self.layers = nn.ModuleList( + [copy.deepcopy(encoder_layer) for i in range(num_layers)] + ) + self.num_layers = num_layers + + def forward( + self, + src: Tensor, + pos_emb: Tensor, + mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + warmup: float = 1.0, + ) -> Tensor: + r"""Pass the input through the encoder layers in turn. + + Args: + src: the sequence to the encoder (required). + pos_emb: Positional embedding tensor (required). + mask: the mask for the src sequence (optional). + src_key_padding_mask: the mask for the src keys per batch (optional). + warmup: controls selective bypass of of layers; if < 1.0, we will + bypass layers more frequently. + + Shape: + src: (S, N, E). + pos_emb: (N, 2*S-1, E) + mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number + + """ + output = src + + for layer_index, mod in enumerate(self.layers): + output = mod( + output, + pos_emb, + src_mask=mask, + src_key_padding_mask=src_key_padding_mask, + warmup=warmup, + ) + + return output + + @torch.jit.export + def chunk_forward( + self, + src: Tensor, + pos_emb: Tensor, + states: List[Tensor], + mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + warmup: float = 1.0, + left_context: int = 0, + right_context: int = 0, + ) -> Tuple[Tensor, List[Tensor]]: + r"""Pass the input through the encoder layers in turn. + + Args: + src: the sequence to the encoder (required). + pos_emb: Positional embedding tensor (required). + states: + The decode states for previous frames which contains the cached data. + It has two elements, the first element is the attn_cache which has + a shape of (encoder_layers, left_context, batch, attention_dim), + the second element is the conv_cache which has a shape of + (encoder_layers, cnn_module_kernel-1, batch, conv_dim). + Note: states will be modified in this function. + mask: the mask for the src sequence (optional). + src_key_padding_mask: the mask for the src keys per batch (optional). + warmup: controls selective bypass of of layers; if < 1.0, we will + bypass layers more frequently. + left_context: + How many previous frames the attention can see in current chunk. + Note: It's not that each individual frame has `left_context` frames + of left context, some have more. + right_context: + How many future frames the attention can see in current chunk. + Note: It's not that each individual frame has `right_context` frames + of right context, some have more. + Shape: + src: (S, N, E). + pos_emb: (N, 2*(S+left_context)-1, E). + mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number + + """ + assert not self.training + assert len(states) == 2 + assert states[0].shape == ( + self.num_layers, + left_context, + src.size(1), + src.size(2), + ) + assert states[1].size(0) == self.num_layers + + output = src + + for layer_index, mod in enumerate(self.layers): + cache = [states[0][layer_index], states[1][layer_index]] + output, cache = mod.chunk_forward( + output, + pos_emb, + states=cache, + src_mask=mask, + src_key_padding_mask=src_key_padding_mask, + warmup=warmup, + left_context=left_context, + right_context=right_context, + ) + states[0][layer_index] = cache[0] + states[1][layer_index] = cache[1] + + return output, states + + +class RelPositionalEncoding(torch.nn.Module): + """Relative positional encoding module. + + See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py + + Args: + d_model: Embedding dimension. + dropout_rate: Dropout rate. + max_len: Maximum input length. + + """ + + def __init__( + self, d_model: int, dropout_rate: float, max_len: int = 5000 + ) -> None: + """Construct an PositionalEncoding object.""" + super(RelPositionalEncoding, self).__init__() + self.d_model = d_model + self.dropout = torch.nn.Dropout(p=dropout_rate) + self.pe = None + self.extend_pe(torch.tensor(0.0).expand(1, max_len)) + + def extend_pe(self, x: Tensor, left_context: int = 0) -> None: + """Reset the positional encodings.""" + x_size_1 = x.size(1) + left_context + if self.pe is not None: + # self.pe contains both positive and negative parts + # the length of self.pe is 2 * input_len - 1 + if self.pe.size(1) >= x_size_1 * 2 - 1: + # Note: TorchScript doesn't implement operator== for torch.Device + if self.pe.dtype != x.dtype or str(self.pe.device) != str( + x.device + ): + self.pe = self.pe.to(dtype=x.dtype, device=x.device) + return + # Suppose `i` means to the position of query vector and `j` means the + # position of key vector. We use position relative positions when keys + # are to the left (i>j) and negative relative positions otherwise (i Tuple[Tensor, Tensor]: + """Add positional encoding. + + Args: + x (torch.Tensor): Input tensor (batch, time, `*`). + left_context (int): left context (in frames) used during streaming decoding. + this is used only in real streaming decoding, in other circumstances, + it MUST be 0. + + Returns: + torch.Tensor: Encoded tensor (batch, time, `*`). + torch.Tensor: Encoded tensor (batch, 2*time-1, `*`). + + """ + self.extend_pe(x, left_context) + x_size_1 = x.size(1) + left_context + pos_emb = self.pe[ + :, + self.pe.size(1) // 2 + - x_size_1 + + 1 : self.pe.size(1) // 2 # noqa E203 + + x.size(1), + ] + return self.dropout(x), self.dropout(pos_emb) + + +class RelPositionMultiheadAttention(nn.Module): + r"""Multi-Head Attention layer with relative position encoding + + See reference: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" + + Args: + embed_dim: total dimension of the model. + num_heads: parallel attention heads. + dropout: a Dropout layer on attn_output_weights. Default: 0.0. + + Examples:: + + >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads) + >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb) + """ + + def __init__( + self, + embed_dim: int, + num_heads: int, + dropout: float = 0.0, + ) -> None: + super(RelPositionMultiheadAttention, self).__init__() + self.embed_dim = embed_dim + self.num_heads = num_heads + self.dropout = dropout + self.head_dim = embed_dim // num_heads + assert ( + self.head_dim * num_heads == self.embed_dim + ), "embed_dim must be divisible by num_heads" + + self.in_proj = ScaledLinear(embed_dim, 3 * embed_dim, bias=True) + self.out_proj = ScaledLinear( + embed_dim, embed_dim, bias=True, initial_scale=0.25 + ) + + # linear transformation for positional encoding. + self.linear_pos = ScaledLinear(embed_dim, embed_dim, bias=False) + # these two learnable bias are used in matrix c and matrix d + # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3 + self.pos_bias_u = nn.Parameter(torch.Tensor(num_heads, self.head_dim)) + self.pos_bias_v = nn.Parameter(torch.Tensor(num_heads, self.head_dim)) + self.pos_bias_u_scale = nn.Parameter(torch.zeros(()).detach()) + self.pos_bias_v_scale = nn.Parameter(torch.zeros(()).detach()) + self._reset_parameters() + + def _pos_bias_u(self): + return self.pos_bias_u * self.pos_bias_u_scale.exp() + + def _pos_bias_v(self): + return self.pos_bias_v * self.pos_bias_v_scale.exp() + + def _reset_parameters(self) -> None: + nn.init.normal_(self.pos_bias_u, std=0.01) + nn.init.normal_(self.pos_bias_v, std=0.01) + + def forward( + self, + query: Tensor, + key: Tensor, + value: Tensor, + pos_emb: Tensor, + key_padding_mask: Optional[Tensor] = None, + need_weights: bool = True, + attn_mask: Optional[Tensor] = None, + left_context: int = 0, + ) -> Tuple[Tensor, Optional[Tensor]]: + r""" + Args: + query, key, value: map a query and a set of key-value pairs to an output. + pos_emb: Positional embedding tensor + key_padding_mask: if provided, specified padding elements in the key will + be ignored by the attention. When given a binary mask and a value is True, + the corresponding value on the attention layer will be ignored. When given + a byte mask and a value is non-zero, the corresponding value on the attention + layer will be ignored + need_weights: output attn_output_weights. + attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all + the batches while a 3D mask allows to specify a different mask for the entries of each batch. + left_context (int): left context (in frames) used during streaming decoding. + this is used only in real streaming decoding, in other circumstances, + it MUST be 0. + + Shape: + - Inputs: + - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is + the embedding dimension. + - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is + the embedding dimension. + - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length. + If a ByteTensor is provided, the non-zero positions will be ignored while the position + with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the + value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged. + - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length. + 3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length, + S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked + positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend + while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True`` + is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor + is provided, it will be added to the attention weight. + + - Outputs: + - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, + E is the embedding dimension. + - attn_output_weights: :math:`(N, L, S)` where N is the batch size, + L is the target sequence length, S is the source sequence length. + """ + return self.multi_head_attention_forward( + query, + key, + value, + pos_emb, + self.embed_dim, + self.num_heads, + self.in_proj.get_weight(), + self.in_proj.get_bias(), + self.dropout, + self.out_proj.get_weight(), + self.out_proj.get_bias(), + training=self.training, + key_padding_mask=key_padding_mask, + need_weights=need_weights, + attn_mask=attn_mask, + left_context=left_context, + ) + + def rel_shift(self, x: Tensor, left_context: int = 0) -> Tensor: + """Compute relative positional encoding. + + Args: + x: Input tensor (batch, head, time1, 2*time1-1). + time1 means the length of query vector. + left_context (int): left context (in frames) used during streaming decoding. + this is used only in real streaming decoding, in other circumstances, + it MUST be 0. + + Returns: + Tensor: tensor of shape (batch, head, time1, time2) + (note: time2 has the same value as time1, but it is for + the key, while time1 is for the query). + """ + (batch_size, num_heads, time1, n) = x.shape + + time2 = time1 + left_context + assert ( + n == left_context + 2 * time1 - 1 + ), f"{n} == {left_context} + 2 * {time1} - 1" + + # Note: TorchScript requires explicit arg for stride() + batch_stride = x.stride(0) + head_stride = x.stride(1) + time1_stride = x.stride(2) + n_stride = x.stride(3) + return x.as_strided( + (batch_size, num_heads, time1, time2), + (batch_stride, head_stride, time1_stride - n_stride, n_stride), + storage_offset=n_stride * (time1 - 1), + ) + + def multi_head_attention_forward( + self, + query: Tensor, + key: Tensor, + value: Tensor, + pos_emb: Tensor, + embed_dim_to_check: int, + num_heads: int, + in_proj_weight: Tensor, + in_proj_bias: Tensor, + dropout_p: float, + out_proj_weight: Tensor, + out_proj_bias: Tensor, + training: bool = True, + key_padding_mask: Optional[Tensor] = None, + need_weights: bool = True, + attn_mask: Optional[Tensor] = None, + left_context: int = 0, + ) -> Tuple[Tensor, Optional[Tensor]]: + r""" + Args: + query, key, value: map a query and a set of key-value pairs to an output. + pos_emb: Positional embedding tensor + embed_dim_to_check: total dimension of the model. + num_heads: parallel attention heads. + in_proj_weight, in_proj_bias: input projection weight and bias. + dropout_p: probability of an element to be zeroed. + out_proj_weight, out_proj_bias: the output projection weight and bias. + training: apply dropout if is ``True``. + key_padding_mask: if provided, specified padding elements in the key will + be ignored by the attention. This is an binary mask. When the value is True, + the corresponding value on the attention layer will be filled with -inf. + need_weights: output attn_output_weights. + attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all + the batches while a 3D mask allows to specify a different mask for the entries of each batch. + left_context (int): left context (in frames) used during streaming decoding. + this is used only in real streaming decoding, in other circumstances, + it MUST be 0. + + Shape: + Inputs: + - query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is + the embedding dimension. + - value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is + the embedding dimension. + - pos_emb: :math:`(N, 2*L-1, E)` or :math:`(1, 2*L-1, E)` where L is the target sequence + length, N is the batch size, E is the embedding dimension. + - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length. + If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions + will be unchanged. If a BoolTensor is provided, the positions with the + value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged. + - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length. + 3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length, + S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked + positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend + while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True`` + are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor + is provided, it will be added to the attention weight. + + Outputs: + - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, + E is the embedding dimension. + - attn_output_weights: :math:`(N, L, S)` where N is the batch size, + L is the target sequence length, S is the source sequence length. + """ + + tgt_len, bsz, embed_dim = query.size() + assert embed_dim == embed_dim_to_check + assert key.size(0) == value.size(0) and key.size(1) == value.size(1) + + head_dim = embed_dim // num_heads + assert ( + head_dim * num_heads == embed_dim + ), "embed_dim must be divisible by num_heads" + + scaling = float(head_dim) ** -0.5 + + if torch.equal(query, key) and torch.equal(key, value): + # self-attention + q, k, v = nn.functional.linear( + query, in_proj_weight, in_proj_bias + ).chunk(3, dim=-1) + + elif torch.equal(key, value): + # encoder-decoder attention + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = 0 + _end = embed_dim + _w = in_proj_weight[_start:_end, :] + if _b is not None: + _b = _b[_start:_end] + q = nn.functional.linear(query, _w, _b) + + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = embed_dim + _end = None + _w = in_proj_weight[_start:, :] + if _b is not None: + _b = _b[_start:] + k, v = nn.functional.linear(key, _w, _b).chunk(2, dim=-1) + + else: + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = 0 + _end = embed_dim + _w = in_proj_weight[_start:_end, :] + if _b is not None: + _b = _b[_start:_end] + q = nn.functional.linear(query, _w, _b) + + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = embed_dim + _end = embed_dim * 2 + _w = in_proj_weight[_start:_end, :] + if _b is not None: + _b = _b[_start:_end] + k = nn.functional.linear(key, _w, _b) + + # This is inline in_proj function with in_proj_weight and in_proj_bias + _b = in_proj_bias + _start = embed_dim * 2 + _end = None + _w = in_proj_weight[_start:, :] + if _b is not None: + _b = _b[_start:] + v = nn.functional.linear(value, _w, _b) + + if attn_mask is not None: + assert ( + attn_mask.dtype == torch.float32 + or attn_mask.dtype == torch.float64 + or attn_mask.dtype == torch.float16 + or attn_mask.dtype == torch.uint8 + or attn_mask.dtype == torch.bool + ), "Only float, byte, and bool types are supported for attn_mask, not {}".format( + attn_mask.dtype + ) + if attn_mask.dtype == torch.uint8: + warnings.warn( + "Byte tensor for attn_mask is deprecated. Use bool tensor instead." + ) + attn_mask = attn_mask.to(torch.bool) + + if attn_mask.dim() == 2: + attn_mask = attn_mask.unsqueeze(0) + if list(attn_mask.size()) != [1, query.size(0), key.size(0)]: + raise RuntimeError( + "The size of the 2D attn_mask is not correct." + ) + elif attn_mask.dim() == 3: + if list(attn_mask.size()) != [ + bsz * num_heads, + query.size(0), + key.size(0), + ]: + raise RuntimeError( + "The size of the 3D attn_mask is not correct." + ) + else: + raise RuntimeError( + "attn_mask's dimension {} is not supported".format( + attn_mask.dim() + ) + ) + # attn_mask's dim is 3 now. + + # convert ByteTensor key_padding_mask to bool + if ( + key_padding_mask is not None + and key_padding_mask.dtype == torch.uint8 + ): + warnings.warn( + "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead." + ) + key_padding_mask = key_padding_mask.to(torch.bool) + + q = (q * scaling).contiguous().view(tgt_len, bsz, num_heads, head_dim) + k = k.contiguous().view(-1, bsz, num_heads, head_dim) + v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1) + + src_len = k.size(0) + + if key_padding_mask is not None: + assert key_padding_mask.size(0) == bsz, "{} == {}".format( + key_padding_mask.size(0), bsz + ) + assert key_padding_mask.size(1) == src_len, "{} == {}".format( + key_padding_mask.size(1), src_len + ) + + q = q.transpose(0, 1) # (batch, time1, head, d_k) + + pos_emb_bsz = pos_emb.size(0) + assert pos_emb_bsz in (1, bsz) # actually it is 1 + p = self.linear_pos(pos_emb).view(pos_emb_bsz, -1, num_heads, head_dim) + # (batch, 2*time1, head, d_k) --> (batch, head, d_k, 2*time -1) + p = p.permute(0, 2, 3, 1) + + q_with_bias_u = (q + self._pos_bias_u()).transpose( + 1, 2 + ) # (batch, head, time1, d_k) + + q_with_bias_v = (q + self._pos_bias_v()).transpose( + 1, 2 + ) # (batch, head, time1, d_k) + + # compute attention score + # first compute matrix a and matrix c + # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3 + k = k.permute(1, 2, 3, 0) # (batch, head, d_k, time2) + matrix_ac = torch.matmul( + q_with_bias_u, k + ) # (batch, head, time1, time2) + + # compute matrix b and matrix d + matrix_bd = torch.matmul( + q_with_bias_v, p + ) # (batch, head, time1, 2*time1-1) + matrix_bd = self.rel_shift(matrix_bd, left_context) + + attn_output_weights = ( + matrix_ac + matrix_bd + ) # (batch, head, time1, time2) + + attn_output_weights = attn_output_weights.view( + bsz * num_heads, tgt_len, -1 + ) + + assert list(attn_output_weights.size()) == [ + bsz * num_heads, + tgt_len, + src_len, + ] + + if attn_mask is not None: + if attn_mask.dtype == torch.bool: + attn_output_weights.masked_fill_(attn_mask, float("-inf")) + else: + attn_output_weights += attn_mask + + if key_padding_mask is not None: + attn_output_weights = attn_output_weights.view( + bsz, num_heads, tgt_len, src_len + ) + attn_output_weights = attn_output_weights.masked_fill( + key_padding_mask.unsqueeze(1).unsqueeze(2), + float("-inf"), + ) + attn_output_weights = attn_output_weights.view( + bsz * num_heads, tgt_len, src_len + ) + + attn_output_weights = nn.functional.softmax(attn_output_weights, dim=-1) + + # If we are using dynamic_chunk_training and setting a limited + # num_left_chunks, the attention may only see the padding values which + # will also be masked out by `key_padding_mask`, at this circumstances, + # the whole column of `attn_output_weights` will be `-inf` + # (i.e. be `nan` after softmax), so, we fill `0.0` at the masking + # positions to avoid invalid loss value below. + if ( + attn_mask is not None + and attn_mask.dtype == torch.bool + and key_padding_mask is not None + ): + if attn_mask.size(0) != 1: + attn_mask = attn_mask.view(bsz, num_heads, tgt_len, src_len) + combined_mask = attn_mask | key_padding_mask.unsqueeze( + 1 + ).unsqueeze(2) + else: + # attn_mask.shape == (1, tgt_len, src_len) + combined_mask = attn_mask.unsqueeze( + 0 + ) | key_padding_mask.unsqueeze(1).unsqueeze(2) + + attn_output_weights = attn_output_weights.view( + bsz, num_heads, tgt_len, src_len + ) + attn_output_weights = attn_output_weights.masked_fill( + combined_mask, 0.0 + ) + attn_output_weights = attn_output_weights.view( + bsz * num_heads, tgt_len, src_len + ) + + attn_output_weights = nn.functional.dropout( + attn_output_weights, p=dropout_p, training=training + ) + + attn_output = torch.bmm(attn_output_weights, v) + assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim] + attn_output = ( + attn_output.transpose(0, 1) + .contiguous() + .view(tgt_len, bsz, embed_dim) + ) + attn_output = nn.functional.linear( + attn_output, out_proj_weight, out_proj_bias + ) + + if need_weights: + # average attention weights over heads + attn_output_weights = attn_output_weights.view( + bsz, num_heads, tgt_len, src_len + ) + return attn_output, attn_output_weights.sum(dim=1) / num_heads + else: + return attn_output, None + + +class ConvolutionModule(nn.Module): + """ConvolutionModule in Conformer model. + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/conformer/convolution.py + + Args: + channels (int): The number of channels of conv layers. + kernel_size (int): Kernerl size of conv layers. + bias (bool): Whether to use bias in conv layers (default=True). + causal (bool): Whether to use causal convolution. + """ + + def __init__( + self, + channels: int, + kernel_size: int, + bias: bool = True, + causal: bool = False, + ) -> None: + """Construct an ConvolutionModule object.""" + super(ConvolutionModule, self).__init__() + # kernerl_size should be a odd number for 'SAME' padding + assert (kernel_size - 1) % 2 == 0 + self.causal = causal + + self.pointwise_conv1 = ScaledConv1d( + channels, + 2 * channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + ) + + # after pointwise_conv1 we put x through a gated linear unit (nn.functional.glu). + # For most layers the normal rms value of channels of x seems to be in the range 1 to 4, + # but sometimes, for some reason, for layer 0 the rms ends up being very large, + # between 50 and 100 for different channels. This will cause very peaky and + # sparse derivatives for the sigmoid gating function, which will tend to make + # the loss function not learn effectively. (for most layers the average absolute values + # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion, + # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different + # layers, which likely breaks down as 0.5 for the "linear" half and + # 0.2 to 0.3 for the part that goes into the sigmoid. The idea is that if we + # constrain the rms values to a reasonable range via a constraint of max_abs=10.0, + # it will be in a better position to start learning something, i.e. to latch onto + # the correct range. + self.deriv_balancer1 = ActivationBalancer( + channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0 + ) + + self.lorder = kernel_size - 1 + padding = (kernel_size - 1) // 2 + if self.causal: + padding = 0 + + self.depthwise_conv = ScaledConv1d( + channels, + channels, + kernel_size, + stride=1, + padding=padding, + groups=channels, + bias=bias, + ) + + self.deriv_balancer2 = ActivationBalancer( + channel_dim=1, min_positive=0.05, max_positive=1.0 + ) + + self.activation = DoubleSwish() + + self.pointwise_conv2 = ScaledConv1d( + channels, + channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + initial_scale=0.25, + ) + + def forward( + self, + x: Tensor, + cache: Optional[Tensor] = None, + right_context: int = 0, + ) -> Tuple[Tensor, Tensor]: + """Compute convolution module. + + Args: + x: Input tensor (#time, batch, channels). + cache: The cache of depthwise_conv, only used in real streaming + decoding. + right_context: + How many future frames the attention can see in current chunk. + Note: It's not that each individual frame has `right_context` frames + of right context, some have more. + + Returns: + If cache is None return the output tensor (#time, batch, channels). + If cache is not None, return a tuple of Tensor, the first one is + the output tensor (#time, batch, channels), the second one is the + new cache for next chunk (#kernel_size - 1, batch, channels). + + """ + # exchange the temporal dimension and the feature dimension + x = x.permute(1, 2, 0) # (#batch, channels, time). + + # GLU mechanism + x = self.pointwise_conv1(x) # (batch, 2*channels, time) + + x = self.deriv_balancer1(x) + x = nn.functional.glu(x, dim=1) # (batch, channels, time) + + # 1D Depthwise Conv + if self.causal and self.lorder > 0: + if cache is None: + # Make depthwise_conv causal by + # manualy padding self.lorder zeros to the left + x = nn.functional.pad(x, (self.lorder, 0), "constant", 0.0) + else: + assert ( + not self.training + ), "Cache should be None in training time" + assert cache.size(0) == self.lorder + x = torch.cat([cache.permute(1, 2, 0), x], dim=2) + if right_context > 0: + cache = x.permute(2, 0, 1)[ + -(self.lorder + right_context) : ( # noqa + -right_context + ), + ..., + ] + else: + cache = x.permute(2, 0, 1)[-self.lorder :, ...] # noqa + x = self.depthwise_conv(x) + + x = self.deriv_balancer2(x) + x = self.activation(x) + + x = self.pointwise_conv2(x) # (batch, channel, time) + + # torch.jit.script requires return types be the same as annotated above + if cache is None: + cache = torch.empty(0) + + return x.permute(2, 0, 1), cache + + +class Conv2dSubsampling(nn.Module): + """Convolutional 2D subsampling (to 1/4 length). + + Convert an input of shape (N, T, idim) to an output + with shape (N, T', odim), where + T' = ((T-1)//2 - 1)//2, which approximates T' == T//4 + + It is based on + https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py # noqa + """ + + def __init__( + self, + in_channels: int, + out_channels: int, + layer1_channels: int = 8, + layer2_channels: int = 32, + layer3_channels: int = 128, + ) -> None: + """ + Args: + in_channels: + Number of channels in. The input shape is (N, T, in_channels). + Caution: It requires: T >=7, in_channels >=7 + out_channels + Output dim. The output shape is (N, ((T-1)//2 - 1)//2, out_channels) + layer1_channels: + Number of channels in layer1 + layer1_channels: + Number of channels in layer2 + """ + assert in_channels >= 7 + super().__init__() + + self.conv = nn.Sequential( + ScaledConv2d( + in_channels=1, + out_channels=layer1_channels, + kernel_size=3, + padding=1, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ScaledConv2d( + in_channels=layer1_channels, + out_channels=layer2_channels, + kernel_size=3, + stride=2, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ScaledConv2d( + in_channels=layer2_channels, + out_channels=layer3_channels, + kernel_size=3, + stride=2, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ) + self.out = ScaledLinear( + layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels + ) + # set learn_eps=False because out_norm is preceded by `out`, and `out` + # itself has learned scale, so the extra degree of freedom is not + # needed. + self.out_norm = BasicNorm(out_channels, learn_eps=False) + # constrain median of output to be close to zero. + self.out_balancer = ActivationBalancer( + channel_dim=-1, min_positive=0.45, max_positive=0.55 + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """Subsample x. + + Args: + x: + Its shape is (N, T, idim). + + Returns: + Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim) + """ + # On entry, x is (N, T, idim) + x = x.unsqueeze(1) # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W) + x = self.conv(x) + # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2) + b, c, t, f = x.size() + x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f)) + # Now x is of shape (N, ((T-1)//2 - 1))//2, odim) + x = self.out_norm(x) + x = self.out_balancer(x) + return x + + +if __name__ == "__main__": + feature_dim = 50 + c = Conformer(num_features=feature_dim, d_model=128, nhead=4) + batch_size = 5 + seq_len = 20 + # Just make sure the forward pass runs. + f = c( + torch.randn(batch_size, seq_len, feature_dim), + torch.full((batch_size,), seq_len, dtype=torch.int64), + warmup=0.5, + ) diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decode.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decode.py new file mode 100755 index 000000000..0ebe20372 --- /dev/null +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decode.py @@ -0,0 +1,815 @@ +#!/usr/bin/env python3 +# +# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./pruned_transducer_stateless2/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless2/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./pruned_transducer_stateless2/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless2/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./pruned_transducer_stateless2/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless2/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search (one best) +./pruned_transducer_stateless2/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless2/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(5) fast beam search (nbest) +./pruned_transducer_stateless2/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless2/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 \ + --num-paths 200 \ + --nbest-scale 0.5 + +(6) fast beam search (nbest oracle WER) +./pruned_transducer_stateless2/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless2/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_oracle \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 \ + --num-paths 200 \ + --nbest-scale 0.5 + +(7) fast beam search (with LG) +./pruned_transducer_stateless2/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless2/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_LG \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(8) decode in streaming mode (take greedy search as an example) +./pruned_transducer_stateless2/decode.py \ + --epoch 28 \ + --avg 15 \ + --simulate-streaming 1 \ + --causal-convolution 1 \ + --decode-chunk-size 16 \ + --left-context 64 \ + --exp-dir ./pruned_transducer_stateless2/exp \ + --max-duration 600 \ + --decoding-method greedy_search + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 +""" + + +import argparse +import logging +import math +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import FisherSwbdSpeechAsrDataModule + + +from beam_search import ( + beam_search, + fast_beam_search_nbest, + fast_beam_search_nbest_LG, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=28, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 0. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless2/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_500", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + - fast_beam_search_nbest + - fast_beam_search_nbest_oracle + - fast_beam_search_nbest_LG + If you use fast_beam_search_nbest_LG, you have to specify + `--lang-dir`, which should contain `LG.pt`. + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search, + fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=0.01, + help=""" + Used only when --decoding_method is fast_beam_search_nbest_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--simulate-streaming", + type=str2bool, + default=False, + help="""Whether to simulate streaming in decoding, this is a good way to + test a streaming model. + """, + ) + + parser.add_argument( + "--decode-chunk-size", + type=int, + default=16, + help="The chunk size for decoding (in frames after subsampling)", + ) + + parser.add_argument( + "--left-context", + type=int, + default=64, + help="left context can be seen during decoding (in frames after subsampling)", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + add_model_arguments(parser) + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = model.device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + feature_lens += params.left_context + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, params.left_context), + value=LOG_EPS, + ) + + if params.simulate_streaming: + encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward( + x=feature, + x_lens=feature_lens, + chunk_size=params.decode_chunk_size, + left_context=params.left_context, + simulate_streaming=True, + ) + else: + encoder_out, encoder_out_lens = model.encoder( + x=feature, x_lens=feature_lens + ) + + hyps = [] + + if params.decoding_method == "fast_beam_search": + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_LG": + hyp_tokens = fast_beam_search_nbest_LG( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + ) + for hyp in hyp_tokens: + hyps.append([word_table[i] for i in hyp]) + elif params.decoding_method == "fast_beam_search_nbest": + hyp_tokens = fast_beam_search_nbest( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_oracle": + hyp_tokens = fast_beam_search_nbest_oracle( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + ref_texts=sp.encode(supervisions["text"]), + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif ( + params.decoding_method == "greedy_search" + and params.max_sym_per_frame == 1 + ): + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif params.decoding_method == "fast_beam_search": + return { + ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}" + ): hyps + } + elif "fast_beam_search" in params.decoding_method: + key = f"beam_{params.beam}_" + key += f"max_contexts_{params.max_contexts}_" + key += f"max_states_{params.max_states}" + if "nbest" in params.decoding_method: + key += f"_num_paths_{params.num_paths}_" + key += f"nbest_scale_{params.nbest_scale}" + if "LG" in params.decoding_method: + key += f"_ngram_lm_scale_{params.ngram_lm_scale}" + + return {key: hyps} + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for hyp_words, ref_text in zip(hyps, texts): + ref_words = ref_text.split() + this_batch.append((ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info( + f"batch {batch_str}, cuts processed until now is {num_cuts}" + ) + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[List[int], List[int]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir + / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + FisherSwbdSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "fast_beam_search_nbest", + "fast_beam_search_nbest_LG", + "fast_beam_search_nbest_oracle", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.simulate_streaming: + params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}" + params.suffix += f"-left-context-{params.left_context}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + if "nbest" in params.decoding_method: + params.suffix += f"-nbest-scale-{params.nbest_scale}" + params.suffix += f"-num-paths-{params.num_paths}" + if "LG" in params.decoding_method: + params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" + elif "beam_search" in params.decoding_method: + params.suffix += ( + f"-{params.decoding_method}-beam-size-{params.beam_size}" + ) + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + if params.simulate_streaming: + assert ( + params.causal_convolution + ), "Decoding in streaming requires causal convolution" + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if start >= 0: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + + model.to(device) + model.eval() + model.device = device + + if "fast_beam_search" in params.decoding_method: + if params.decoding_method == "fast_beam_search_nbest_LG": + lexicon = Lexicon(params.lang_dir) + word_table = lexicon.word_table + lg_filename = params.lang_dir / "LG.pt" + logging.info(f"Loading {lg_filename}") + decoding_graph = k2.Fsa.from_dict( + torch.load(lg_filename, map_location=device) + ) + decoding_graph.scores *= params.ngram_lm_scale + else: + word_table = None + decoding_graph = k2.trivial_graph( + params.vocab_size - 1, device=device + ) + else: + decoding_graph = None + word_table = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + fisherswbd = FisherSwbdSpeechAsrDataModule(args) + + test_eval2000_cuts = fisherswbd.test_eval2000_cuts() + test_swbd_cuts = fisherswbd.test_swbd_cuts() + test_callhome_cuts = fisherswbd.test_callhome_cuts() + + test_eval2000_dl = fisherswbd.test_dataloaders(test_eval2000_cuts) + test_swbd_dl = fisherswbd.test_dataloaders(test_swbd_cuts) + test_callhome_dl = fisherswbd.test_dataloaders(test_callhome_cuts) + + test_sets = ["eval2000", "swbd", "callhome"] + test_dl = [test_eval2000_dl, test_swbd_dl, test_callhome_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decode_stream.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decode_stream.py new file mode 100644 index 000000000..ba5e80555 --- /dev/null +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decode_stream.py @@ -0,0 +1,126 @@ +# Copyright 2022 Xiaomi Corp. (authors: Wei Kang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from typing import List, Optional, Tuple + +import k2 +import torch + +from icefall.utils import AttributeDict + + +class DecodeStream(object): + def __init__( + self, + params: AttributeDict, + initial_states: List[torch.Tensor], + decoding_graph: Optional[k2.Fsa] = None, + device: torch.device = torch.device("cpu"), + ) -> None: + """ + Args: + initial_states: + Initial decode states of the model, e.g. the return value of + `get_init_state` in conformer.py + decoding_graph: + Decoding graph used for decoding, may be a TrivialGraph or a HLG. + Used only when decoding_method is fast_beam_search. + device: + The device to run this stream. + """ + if decoding_graph is not None: + assert device == decoding_graph.device + + self.params = params + self.LOG_EPS = math.log(1e-10) + + self.states = initial_states + + # It contains a 2-D tensors representing the feature frames. + self.features: torch.Tensor = None + + self.num_frames: int = 0 + # how many frames have been processed. (before subsampling). + # we only modify this value in `func:get_feature_frames`. + self.num_processed_frames: int = 0 + + self._done: bool = False + + # The transcript of current utterance. + self.ground_truth: str = "" + + # The decoding result (partial or final) of current utterance. + self.hyp: List = [] + + # how many frames have been processed, after subsampling (i.e. a + # cumulative sum of the second return value of + # encoder.streaming_forward + self.done_frames: int = 0 + + self.pad_length = ( + params.right_context + 2 + ) * params.subsampling_factor + 3 + + if params.decoding_method == "greedy_search": + self.hyp = [params.blank_id] * params.context_size + elif params.decoding_method == "fast_beam_search": + # The rnnt_decoding_stream for fast_beam_search. + self.rnnt_decoding_stream: k2.RnntDecodingStream = ( + k2.RnntDecodingStream(decoding_graph) + ) + else: + assert ( + False + ), f"Decoding method :{params.decoding_method} do not support." + + @property + def done(self) -> bool: + """Return True if all the features are processed.""" + return self._done + + def set_features( + self, + features: torch.Tensor, + ) -> None: + """Set features tensor of current utterance.""" + assert features.dim() == 2, features.dim() + self.features = torch.nn.functional.pad( + features, + (0, 0, 0, self.pad_length), + mode="constant", + value=self.LOG_EPS, + ) + self.num_frames = self.features.size(0) + + def get_feature_frames(self, chunk_size: int) -> Tuple[torch.Tensor, int]: + """Consume chunk_size frames of features""" + chunk_length = chunk_size + self.pad_length + + ret_length = min( + self.num_frames - self.num_processed_frames, chunk_length + ) + + ret_features = self.features[ + self.num_processed_frames : self.num_processed_frames # noqa + + ret_length + ] + + self.num_processed_frames += chunk_size + if self.num_processed_frames >= self.num_frames: + self._done = True + + return ret_features, ret_length diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decoder.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decoder.py new file mode 100644 index 000000000..1ddfce034 --- /dev/null +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decoder.py @@ -0,0 +1,106 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +import torch.nn as nn +import torch.nn.functional as F +from scaling import ScaledConv1d, ScaledEmbedding + + +class Decoder(nn.Module): + """This class modifies the stateless decoder from the following paper: + + RNN-transducer with stateless prediction network + https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419 + + It removes the recurrent connection from the decoder, i.e., the prediction + network. Different from the above paper, it adds an extra Conv1d + right after the embedding layer. + + TODO: Implement https://arxiv.org/pdf/2109.07513.pdf + """ + + def __init__( + self, + vocab_size: int, + decoder_dim: int, + blank_id: int, + context_size: int, + ): + """ + Args: + vocab_size: + Number of tokens of the modeling unit including blank. + decoder_dim: + Dimension of the input embedding, and of the decoder output. + blank_id: + The ID of the blank symbol. + context_size: + Number of previous words to use to predict the next word. + 1 means bigram; 2 means trigram. n means (n+1)-gram. + """ + super().__init__() + + self.embedding = ScaledEmbedding( + num_embeddings=vocab_size, + embedding_dim=decoder_dim, + padding_idx=blank_id, + ) + self.blank_id = blank_id + + assert context_size >= 1, context_size + self.context_size = context_size + self.vocab_size = vocab_size + if context_size > 1: + self.conv = ScaledConv1d( + in_channels=decoder_dim, + out_channels=decoder_dim, + kernel_size=context_size, + padding=0, + groups=decoder_dim, + bias=False, + ) + else: + # It is to support torch script + self.conv = nn.Identity() + + def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor: + """ + Args: + y: + A 2-D tensor of shape (N, U). + need_pad: + True to left pad the input. Should be True during training. + False to not pad the input. Should be False during inference. + Returns: + Return a tensor of shape (N, U, decoder_dim). + """ + y = y.to(torch.int64) + embedding_out = self.embedding(y) + if self.context_size > 1: + embedding_out = embedding_out.permute(0, 2, 1) + if need_pad is True: + embedding_out = F.pad( + embedding_out, pad=(self.context_size - 1, 0) + ) + else: + # During inference time, there is no need to do extra padding + # as we only need one output + assert embedding_out.size(-1) == self.context_size + embedding_out = self.conv(embedding_out) + embedding_out = embedding_out.permute(0, 2, 1) + embedding_out = F.relu(embedding_out) + return embedding_out diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/encoder_interface.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/encoder_interface.py new file mode 100644 index 000000000..257facce4 --- /dev/null +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/encoder_interface.py @@ -0,0 +1,43 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Tuple + +import torch +import torch.nn as nn + + +class EncoderInterface(nn.Module): + def forward( + self, x: torch.Tensor, x_lens: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Args: + x: + A tensor of shape (batch_size, input_seq_len, num_features) + containing the input features. + x_lens: + A tensor of shape (batch_size,) containing the number of frames + in `x` before padding. + Returns: + Return a tuple containing two tensors: + - encoder_out, a tensor of (batch_size, out_seq_len, output_dim) + containing unnormalized probabilities, i.e., the output of a + linear layer. + - encoder_out_lens, a tensor of shape (batch_size,) containing + the number of frames in `encoder_out` before padding. + """ + raise NotImplementedError("Please implement it in a subclass") diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/export.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/export.py new file mode 100755 index 000000000..f1a8ea589 --- /dev/null +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/export.py @@ -0,0 +1,230 @@ +#!/usr/bin/env python3 +# +# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This script converts several saved checkpoints +# to a single one using model averaging. +""" +Usage: +./pruned_transducer_stateless2/export.py \ + --exp-dir ./pruned_transducer_stateless2/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 20 \ + --avg 10 + +It will generate a file exp_dir/pretrained.pt + +To use the generated file with `pruned_transducer_stateless2/decode.py`, +you can do: + + cd /path/to/exp_dir + ln -s pretrained.pt epoch-9999.pt + + cd /path/to/egs/librispeech/ASR + ./pruned_transducer_stateless2/decode.py \ + --exp-dir ./pruned_transducer_stateless2/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 100 \ + --bpe-model data/lang_bpe_500/bpe.model +""" + +import argparse +import logging +from pathlib import Path + +import sentencepiece as spm +import torch +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import str2bool + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=28, + help="""It specifies the checkpoint to use for averaging. + Note: Epoch counts from 0. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless2/exp", + help="""It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--jit", + type=str2bool, + default=False, + help="""True to save a model after applying torch.jit.script. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + + parser.add_argument( + "--streaming-model", + type=str2bool, + default=False, + help="""Whether to export a streaming model, if the models in exp-dir + are streaming model, this should be True. + """, + ) + + add_model_arguments(parser) + return parser + + +def main(): + args = get_parser().parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + if params.streaming_model: + assert params.causal_convolution + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + model.to(device) + + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if start >= 0: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + + model.eval() + + model.to("cpu") + model.eval() + + if params.jit: + # We won't use the forward() method of the model in C++, so just ignore + # it here. + # Otherwise, one of its arguments is a ragged tensor and is not + # torch scriptabe. + model.__class__.forward = torch.jit.ignore(model.__class__.forward) + logging.info("Using torch.jit.script") + model = torch.jit.script(model) + filename = params.exp_dir / "cpu_jit.pt" + model.save(str(filename)) + logging.info(f"Saved to {filename}") + else: + logging.info("Not using torch.jit.script") + # Save it using a format so that it can be loaded + # by :func:`load_checkpoint` + filename = params.exp_dir / "pretrained.pt" + torch.save({"model": model.state_dict()}, str(filename)) + logging.info(f"Saved to {filename}") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/joiner.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/joiner.py new file mode 100644 index 000000000..b916addf0 --- /dev/null +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/joiner.py @@ -0,0 +1,69 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +import torch.nn as nn +from scaling import ScaledLinear + + +class Joiner(nn.Module): + def __init__( + self, + encoder_dim: int, + decoder_dim: int, + joiner_dim: int, + vocab_size: int, + ): + super().__init__() + + self.encoder_proj = ScaledLinear(encoder_dim, joiner_dim) + self.decoder_proj = ScaledLinear(decoder_dim, joiner_dim) + self.output_linear = ScaledLinear(joiner_dim, vocab_size) + + def forward( + self, + encoder_out: torch.Tensor, + decoder_out: torch.Tensor, + project_input: bool = True, + ) -> torch.Tensor: + """ + Args: + encoder_out: + Output from the encoder. Its shape is (N, T, s_range, C). + decoder_out: + Output from the decoder. Its shape is (N, T, s_range, C). + project_input: + If true, apply input projections encoder_proj and decoder_proj. + If this is false, it is the user's responsibility to do this + manually. + Returns: + Return a tensor of shape (N, T, s_range, C). + """ + + assert encoder_out.ndim == decoder_out.ndim + assert encoder_out.ndim in (2, 4) + assert encoder_out.shape == decoder_out.shape + + if project_input: + logit = self.encoder_proj(encoder_out) + self.decoder_proj( + decoder_out + ) + else: + logit = encoder_out + decoder_out + + logit = self.output_linear(torch.tanh(logit)) + + return logit diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/model.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/model.py new file mode 100644 index 000000000..2434fd41d --- /dev/null +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/model.py @@ -0,0 +1,194 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, Wei Kang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import k2 +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface +from scaling import ScaledLinear + +from icefall.utils import add_sos + + +class Transducer(nn.Module): + """It implements https://arxiv.org/pdf/1211.3711.pdf + "Sequence Transduction with Recurrent Neural Networks" + """ + + def __init__( + self, + encoder: EncoderInterface, + decoder: nn.Module, + joiner: nn.Module, + encoder_dim: int, + decoder_dim: int, + joiner_dim: int, + vocab_size: int, + ): + """ + Args: + encoder: + It is the transcription network in the paper. Its accepts + two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,). + It returns two tensors: `logits` of shape (N, T, encoder_dm) and + `logit_lens` of shape (N,). + decoder: + It is the prediction network in the paper. Its input shape + is (N, U) and its output shape is (N, U, decoder_dim). + It should contain one attribute: `blank_id`. + joiner: + It has two inputs with shapes: (N, T, encoder_dim) and + (N, U, decoder_dim). + Its output shape is (N, T, U, vocab_size). Note that its output + contains unnormalized probs, i.e., not processed by log-softmax. + """ + super().__init__() + assert isinstance(encoder, EncoderInterface), type(encoder) + assert hasattr(decoder, "blank_id") + + self.encoder = encoder + self.decoder = decoder + self.joiner = joiner + + self.simple_am_proj = ScaledLinear( + encoder_dim, vocab_size, initial_speed=0.5 + ) + self.simple_lm_proj = ScaledLinear(decoder_dim, vocab_size) + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + y: k2.RaggedTensor, + prune_range: int = 5, + am_scale: float = 0.0, + lm_scale: float = 0.0, + warmup: float = 1.0, + ) -> torch.Tensor: + """ + Args: + x: + A 3-D tensor of shape (N, T, C). + x_lens: + A 1-D tensor of shape (N,). It contains the number of frames in `x` + before padding. + y: + A ragged tensor with 2 axes [utt][label]. It contains labels of each + utterance. + prune_range: + The prune range for rnnt loss, it means how many symbols(context) + we are considering for each frame to compute the loss. + am_scale: + The scale to smooth the loss with am (output of encoder network) + part + lm_scale: + The scale to smooth the loss with lm (output of predictor network) + part + warmup: + A value warmup >= 0 that determines which modules are active, values + warmup > 1 "are fully warmed up" and all modules will be active. + Returns: + Return the transducer loss. + + Note: + Regarding am_scale & lm_scale, it will make the loss-function one of + the form: + lm_scale * lm_probs + am_scale * am_probs + + (1-lm_scale-am_scale) * combined_probs + """ + assert x.ndim == 3, x.shape + assert x_lens.ndim == 1, x_lens.shape + assert y.num_axes == 2, y.num_axes + + assert x.size(0) == x_lens.size(0) == y.dim0 + + encoder_out, x_lens = self.encoder(x, x_lens, warmup=warmup) + assert torch.all(x_lens > 0) + + # Now for the decoder, i.e., the prediction network + row_splits = y.shape.row_splits(1) + y_lens = row_splits[1:] - row_splits[:-1] + + blank_id = self.decoder.blank_id + sos_y = add_sos(y, sos_id=blank_id) + + # sos_y_padded: [B, S + 1], start with SOS. + sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id) + + # decoder_out: [B, S + 1, decoder_dim] + decoder_out = self.decoder(sos_y_padded) + + # Note: y does not start with SOS + # y_padded : [B, S] + y_padded = y.pad(mode="constant", padding_value=0) + + y_padded = y_padded.to(torch.int64) + boundary = torch.zeros( + (x.size(0), 4), dtype=torch.int64, device=x.device + ) + boundary[:, 2] = y_lens + boundary[:, 3] = x_lens + + lm = self.simple_lm_proj(decoder_out) + am = self.simple_am_proj(encoder_out) + + with torch.cuda.amp.autocast(enabled=False): + simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed( + lm=lm.float(), + am=am.float(), + symbols=y_padded, + termination_symbol=blank_id, + lm_only_scale=lm_scale, + am_only_scale=am_scale, + boundary=boundary, + reduction="sum", + return_grad=True, + ) + + # ranges : [B, T, prune_range] + ranges = k2.get_rnnt_prune_ranges( + px_grad=px_grad, + py_grad=py_grad, + boundary=boundary, + s_range=prune_range, + ) + + # am_pruned : [B, T, prune_range, encoder_dim] + # lm_pruned : [B, T, prune_range, decoder_dim] + am_pruned, lm_pruned = k2.do_rnnt_pruning( + am=self.joiner.encoder_proj(encoder_out), + lm=self.joiner.decoder_proj(decoder_out), + ranges=ranges, + ) + + # logits : [B, T, prune_range, vocab_size] + + # project_input=False since we applied the decoder's input projections + # prior to do_rnnt_pruning (this is an optimization for speed). + logits = self.joiner(am_pruned, lm_pruned, project_input=False) + + with torch.cuda.amp.autocast(enabled=False): + pruned_loss = k2.rnnt_loss_pruned( + logits=logits.float(), + symbols=y_padded, + ranges=ranges, + termination_symbol=blank_id, + boundary=boundary, + reduction="sum", + ) + + return (simple_loss, pruned_loss) diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/optim.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/optim.py new file mode 100644 index 000000000..432bf8220 --- /dev/null +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/optim.py @@ -0,0 +1,331 @@ +# Copyright 2022 Xiaomi Corp. (authors: Daniel Povey) +# +# See ../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +from typing import List, Optional, Union + +import torch +from torch.optim import Optimizer + + +class Eve(Optimizer): + r""" + Implements Eve algorithm. This is a modified version of AdamW with a special + way of setting the weight-decay / shrinkage-factor, which is designed to make the + rms of the parameters approach a particular target_rms (default: 0.1). This is + for use with networks with 'scaled' versions of modules (see scaling.py), which + will be close to invariant to the absolute scale on the parameter matrix. + + The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_. + The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_. + Eve is unpublished so far. + + Arguments: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups + lr (float, optional): learning rate (default: 1e-3) + betas (Tuple[float, float], optional): coefficients used for computing + running averages of gradient and its square (default: (0.9, 0.999)) + eps (float, optional): term added to the denominator to improve + numerical stability (default: 1e-8) + weight_decay (float, optional): weight decay coefficient (default: 3e-4; + this value means that the weight would decay significantly after + about 3k minibatches. Is not multiplied by learning rate, but + is conditional on RMS-value of parameter being > target_rms. + target_rms (float, optional): target root-mean-square value of + parameters, if they fall below this we will stop applying weight decay. + + + .. _Adam\: A Method for Stochastic Optimization: + https://arxiv.org/abs/1412.6980 + .. _Decoupled Weight Decay Regularization: + https://arxiv.org/abs/1711.05101 + .. _On the Convergence of Adam and Beyond: + https://openreview.net/forum?id=ryQu7f-RZ + """ + + def __init__( + self, + params, + lr=1e-3, + betas=(0.9, 0.98), + eps=1e-8, + weight_decay=1e-3, + target_rms=0.1, + ): + + if not 0.0 <= lr: + raise ValueError("Invalid learning rate: {}".format(lr)) + if not 0.0 <= eps: + raise ValueError("Invalid epsilon value: {}".format(eps)) + if not 0.0 <= betas[0] < 1.0: + raise ValueError( + "Invalid beta parameter at index 0: {}".format(betas[0]) + ) + if not 0.0 <= betas[1] < 1.0: + raise ValueError( + "Invalid beta parameter at index 1: {}".format(betas[1]) + ) + if not 0 <= weight_decay <= 0.1: + raise ValueError( + "Invalid weight_decay value: {}".format(weight_decay) + ) + if not 0 < target_rms <= 10.0: + raise ValueError("Invalid target_rms value: {}".format(target_rms)) + defaults = dict( + lr=lr, + betas=betas, + eps=eps, + weight_decay=weight_decay, + target_rms=target_rms, + ) + super(Eve, self).__init__(params, defaults) + + def __setstate__(self, state): + super(Eve, self).__setstate__(state) + + @torch.no_grad() + def step(self, closure=None): + """Performs a single optimization step. + + Arguments: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + for group in self.param_groups: + for p in group["params"]: + if p.grad is None: + continue + + # Perform optimization step + grad = p.grad + if grad.is_sparse: + raise RuntimeError( + "AdamW does not support sparse gradients" + ) + + state = self.state[p] + + # State initialization + if len(state) == 0: + state["step"] = 0 + # Exponential moving average of gradient values + state["exp_avg"] = torch.zeros_like( + p, memory_format=torch.preserve_format + ) + # Exponential moving average of squared gradient values + state["exp_avg_sq"] = torch.zeros_like( + p, memory_format=torch.preserve_format + ) + + exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"] + + beta1, beta2 = group["betas"] + + state["step"] += 1 + bias_correction1 = 1 - beta1 ** state["step"] + bias_correction2 = 1 - beta2 ** state["step"] + + # Decay the first and second moment running average coefficient + exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1) + exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2) + denom = (exp_avg_sq.sqrt() * (bias_correction2 ** -0.5)).add_( + group["eps"] + ) + + step_size = group["lr"] / bias_correction1 + target_rms = group["target_rms"] + weight_decay = group["weight_decay"] + + if p.numel() > 1: + # avoid applying this weight-decay on "scaling factors" + # (which are scalar). + is_above_target_rms = p.norm() > ( + target_rms * (p.numel() ** 0.5) + ) + p.mul_(1 - (weight_decay * is_above_target_rms)) + p.addcdiv_(exp_avg, denom, value=-step_size) + + return loss + + +class LRScheduler(object): + """ + Base-class for learning rate schedulers where the learning-rate depends on both the + batch and the epoch. + """ + + def __init__(self, optimizer: Optimizer, verbose: bool = False): + # Attach optimizer + if not isinstance(optimizer, Optimizer): + raise TypeError( + "{} is not an Optimizer".format(type(optimizer).__name__) + ) + self.optimizer = optimizer + self.verbose = verbose + + for group in optimizer.param_groups: + group.setdefault("initial_lr", group["lr"]) + + self.base_lrs = [ + group["initial_lr"] for group in optimizer.param_groups + ] + + self.epoch = 0 + self.batch = 0 + + def state_dict(self): + """Returns the state of the scheduler as a :class:`dict`. + + It contains an entry for every variable in self.__dict__ which + is not the optimizer. + """ + return { + "base_lrs": self.base_lrs, + "epoch": self.epoch, + "batch": self.batch, + } + + def load_state_dict(self, state_dict): + """Loads the schedulers state. + + Args: + state_dict (dict): scheduler state. Should be an object returned + from a call to :meth:`state_dict`. + """ + self.__dict__.update(state_dict) + + def get_last_lr(self) -> List[float]: + """Return last computed learning rate by current scheduler. Will be a list of float.""" + return self._last_lr + + def get_lr(self): + # Compute list of learning rates from self.epoch and self.batch and + # self.base_lrs; this must be overloaded by the user. + # e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ] + raise NotImplementedError + + def step_batch(self, batch: Optional[int] = None) -> None: + # Step the batch index, or just set it. If `batch` is specified, it + # must be the batch index from the start of training, i.e. summed over + # all epochs. + # You can call this in any order; if you don't provide 'batch', it should + # of course be called once per batch. + if batch is not None: + self.batch = batch + else: + self.batch = self.batch + 1 + self._set_lrs() + + def step_epoch(self, epoch: Optional[int] = None): + # Step the epoch index, or just set it. If you provide the 'epoch' arg, + # you should call this at the start of the epoch; if you don't provide the 'epoch' + # arg, you should call it at the end of the epoch. + if epoch is not None: + self.epoch = epoch + else: + self.epoch = self.epoch + 1 + self._set_lrs() + + def _set_lrs(self): + values = self.get_lr() + assert len(values) == len(self.optimizer.param_groups) + + for i, data in enumerate(zip(self.optimizer.param_groups, values)): + param_group, lr = data + param_group["lr"] = lr + self.print_lr(self.verbose, i, lr) + self._last_lr = [group["lr"] for group in self.optimizer.param_groups] + + def print_lr(self, is_verbose, group, lr): + """Display the current learning rate.""" + if is_verbose: + print( + f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate" + f" of group {group} to {lr:.4e}." + ) + + +class Eden(LRScheduler): + """ + Eden scheduler. + lr = initial_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 * + (((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25)) + + E.g. suggest initial-lr = 0.003 (passed to optimizer). + + Args: + optimizer: the optimizer to change the learning rates on + lr_batches: the number of batches after which we start significantly + decreasing the learning rate, suggest 5000. + lr_epochs: the number of epochs after which we start significantly + decreasing the learning rate, suggest 6 if you plan to do e.g. + 20 to 40 epochs, but may need smaller number if dataset is huge + and you will do few epochs. + """ + + def __init__( + self, + optimizer: Optimizer, + lr_batches: Union[int, float], + lr_epochs: Union[int, float], + verbose: bool = False, + ): + super(Eden, self).__init__(optimizer, verbose) + self.lr_batches = lr_batches + self.lr_epochs = lr_epochs + + def get_lr(self): + factor = ( + (self.batch ** 2 + self.lr_batches ** 2) / self.lr_batches ** 2 + ) ** -0.25 * ( + ((self.epoch ** 2 + self.lr_epochs ** 2) / self.lr_epochs ** 2) + ** -0.25 + ) + return [x * factor for x in self.base_lrs] + + +def _test_eden(): + m = torch.nn.Linear(100, 100) + optim = Eve(m.parameters(), lr=0.003) + + scheduler = Eden(optim, lr_batches=30, lr_epochs=2, verbose=True) + + for epoch in range(10): + scheduler.step_epoch(epoch) # sets epoch to `epoch` + + for step in range(20): + x = torch.randn(200, 100).detach() + x.requires_grad = True + y = m(x) + dy = torch.randn(200, 100).detach() + f = (y * dy).sum() + f.backward() + + optim.step() + scheduler.step_batch() + optim.zero_grad() + print("last lr = ", scheduler.get_last_lr()) + print("state dict = ", scheduler.state_dict()) + + +if __name__ == "__main__": + _test_eden() diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/pretrained.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/pretrained.py new file mode 100755 index 000000000..f52cb22ab --- /dev/null +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/pretrained.py @@ -0,0 +1,390 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +(1) greedy search +./pruned_transducer_stateless2/pretrained.py \ + --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method greedy_search \ + /path/to/foo.wav \ + /path/to/bar.wav + +(2) beam search +./pruned_transducer_stateless2/pretrained.py \ + --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav + +(3) modified beam search +./pruned_transducer_stateless2/pretrained.py \ + --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method modified_beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav + +(4) fast beam search +./pruned_transducer_stateless2/pretrained.py \ + --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method fast_beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav + +You can also use `./pruned_transducer_stateless2/exp/epoch-xx.pt`. + +Note: ./pruned_transducer_stateless2/exp/pretrained.pt is generated by +./pruned_transducer_stateless2/export.py +""" + + +import argparse +import logging +import math +from typing import List + +import k2 +import kaldifeat +import sentencepiece as spm +import torch +import torchaudio +from beam_search import ( + beam_search, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from torch.nn.utils.rnn import pad_sequence +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.utils import str2bool + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--checkpoint", + type=str, + required=True, + help="Path to the checkpoint. " + "The checkpoint is assumed to be saved by " + "icefall.checkpoint.save_checkpoint().", + ) + + parser.add_argument( + "--bpe-model", + type=str, + help="""Path to bpe.model.""", + ) + + parser.add_argument( + "--method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "sound_files", + type=str, + nargs="+", + help="The input sound file(s) to transcribe. " + "Supported formats are those supported by torchaudio.load(). " + "For example, wav and flac are supported. " + "The sample rate has to be 16kHz.", + ) + + parser.add_argument( + "--sample-rate", + type=int, + default=16000, + help="The sample rate of the input sound file", + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=4, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --method is fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --method is fast_beam_search""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. Used only when + --method is greedy_search. + """, + ) + + parser.add_argument( + "--simulate-streaming", + type=str2bool, + default=False, + help="""Whether to simulate streaming in decoding, this is a good way to + test a streaming model. + """, + ) + + parser.add_argument( + "--decode-chunk-size", + type=int, + default=16, + help="The chunk size for decoding (in frames after subsampling)", + ) + parser.add_argument( + "--left-context", + type=int, + default=64, + help="left context can be seen during decoding (in frames after subsampling)", + ) + + add_model_arguments(parser) + + return parser + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float +) -> List[torch.Tensor]: + """Read a list of sound files into a list 1-D float32 torch tensors. + Args: + filenames: + A list of sound filenames. + expected_sample_rate: + The expected sample rate of the sound files. + Returns: + Return a list of 1-D float32 torch tensors. + """ + ans = [] + for f in filenames: + wave, sample_rate = torchaudio.load(f) + assert sample_rate == expected_sample_rate, ( + f"expected sample rate: {expected_sample_rate}. " + f"Given: {sample_rate}" + ) + # We use only the first channel + ans.append(wave[0]) + return ans + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + + params = get_params() + + params.update(vars(args)) + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + if params.simulate_streaming: + assert ( + params.causal_convolution + ), "Decoding in streaming requires causal convolution" + + logging.info(f"{params}") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + logging.info("Creating model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + checkpoint = torch.load(args.checkpoint, map_location="cpu") + model.load_state_dict(checkpoint["model"], strict=False) + model.to(device) + model.eval() + model.device = device + + logging.info("Constructing Fbank computer") + opts = kaldifeat.FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = params.sample_rate + opts.mel_opts.num_bins = params.feature_dim + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {params.sound_files}") + waves = read_sound_files( + filenames=params.sound_files, expected_sample_rate=params.sample_rate + ) + waves = [w.to(device) for w in waves] + + logging.info("Decoding started") + features = fbank(waves) + feature_lengths = [f.size(0) for f in features] + + features = pad_sequence( + features, batch_first=True, padding_value=math.log(1e-10) + ) + + feature_lengths = torch.tensor(feature_lengths, device=device) + + if params.simulate_streaming: + encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward( + x=features, + x_lens=feature_lengths, + chunk_size=params.decode_chunk_size, + left_context=params.left_context, + simulate_streaming=True, + ) + else: + encoder_out, encoder_out_lens = model.encoder( + x=features, x_lens=feature_lengths + ) + + num_waves = encoder_out.size(0) + hyps = [] + msg = f"Using {params.method}" + if params.method == "beam_search": + msg += f" with beam size {params.beam_size}" + logging.info(msg) + + if params.method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + for i in range(num_waves): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError(f"Unsupported method: {params.method}") + + hyps.append(sp.decode(hyp).split()) + + s = "\n" + for filename, hyp in zip(params.sound_files, hyps): + words = " ".join(hyp) + s += f"{filename}:\n{words}\n\n" + logging.info(s) + + logging.info("Decoding Done") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/scaling.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/scaling.py new file mode 100644 index 000000000..c190be626 --- /dev/null +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/scaling.py @@ -0,0 +1,733 @@ +# Copyright 2022 Xiaomi Corp. (authors: Daniel Povey) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import collections +from itertools import repeat +from typing import Optional, Tuple + +import torch +import torch.nn as nn +from torch import Tensor + + +def _ntuple(n): + def parse(x): + if isinstance(x, collections.Iterable): + return x + return tuple(repeat(x, n)) + + return parse + + +_single = _ntuple(1) +_pair = _ntuple(2) + + +class ActivationBalancerFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, + x: Tensor, + channel_dim: int, + min_positive: float, # e.g. 0.05 + max_positive: float, # e.g. 0.95 + max_factor: float, # e.g. 0.01 + min_abs: float, # e.g. 0.2 + max_abs: float, # e.g. 100.0 + ) -> Tensor: + if x.requires_grad: + if channel_dim < 0: + channel_dim += x.ndim + + # sum_dims = [d for d in range(x.ndim) if d != channel_dim] + # The above line is not torch scriptable for torch 1.6.0 + # torch.jit.frontend.NotSupportedError: comprehension ifs not supported yet: # noqa + sum_dims = [] + for d in range(x.ndim): + if d != channel_dim: + sum_dims.append(d) + + xgt0 = x > 0 + proportion_positive = torch.mean( + xgt0.to(x.dtype), dim=sum_dims, keepdim=True + ) + factor1 = ( + (min_positive - proportion_positive).relu() + * (max_factor / min_positive) + if min_positive != 0.0 + else 0.0 + ) + factor2 = ( + (proportion_positive - max_positive).relu() + * (max_factor / (max_positive - 1.0)) + if max_positive != 1.0 + else 0.0 + ) + factor = factor1 + factor2 + if isinstance(factor, float): + factor = torch.zeros_like(proportion_positive) + + mean_abs = torch.mean(x.abs(), dim=sum_dims, keepdim=True) + below_threshold = mean_abs < min_abs + above_threshold = mean_abs > max_abs + + ctx.save_for_backward( + factor, xgt0, below_threshold, above_threshold + ) + ctx.max_factor = max_factor + ctx.sum_dims = sum_dims + return x + + @staticmethod + def backward( + ctx, x_grad: Tensor + ) -> Tuple[Tensor, None, None, None, None, None, None]: + factor, xgt0, below_threshold, above_threshold = ctx.saved_tensors + dtype = x_grad.dtype + scale_factor = ( + (below_threshold.to(dtype) - above_threshold.to(dtype)) + * (xgt0.to(dtype) - 0.5) + * (ctx.max_factor * 2.0) + ) + + neg_delta_grad = x_grad.abs() * (factor + scale_factor) + return x_grad - neg_delta_grad, None, None, None, None, None, None + + +class BasicNorm(torch.nn.Module): + """ + This is intended to be a simpler, and hopefully cheaper, replacement for + LayerNorm. The observation this is based on, is that Transformer-type + networks, especially with pre-norm, sometimes seem to set one of the + feature dimensions to a large constant value (e.g. 50), which "defeats" + the LayerNorm because the output magnitude is then not strongly dependent + on the other (useful) features. Presumably the weight and bias of the + LayerNorm are required to allow it to do this. + + So the idea is to introduce this large constant value as an explicit + parameter, that takes the role of the "eps" in LayerNorm, so the network + doesn't have to do this trick. We make the "eps" learnable. + + Args: + num_channels: the number of channels, e.g. 512. + channel_dim: the axis/dimension corresponding to the channel, + interprted as an offset from the input's ndim if negative. + shis is NOT the num_channels; it should typically be one of + {-2, -1, 0, 1, 2, 3}. + eps: the initial "epsilon" that we add as ballast in: + scale = ((input_vec**2).mean() + epsilon)**-0.5 + Note: our epsilon is actually large, but we keep the name + to indicate the connection with conventional LayerNorm. + learn_eps: if true, we learn epsilon; if false, we keep it + at the initial value. + """ + + def __init__( + self, + num_channels: int, + channel_dim: int = -1, # CAUTION: see documentation. + eps: float = 0.25, + learn_eps: bool = True, + ) -> None: + super(BasicNorm, self).__init__() + self.num_channels = num_channels + self.channel_dim = channel_dim + if learn_eps: + self.eps = nn.Parameter(torch.tensor(eps).log().detach()) + else: + self.register_buffer("eps", torch.tensor(eps).log().detach()) + + def forward(self, x: Tensor) -> Tensor: + assert x.shape[self.channel_dim] == self.num_channels + scales = ( + torch.mean(x ** 2, dim=self.channel_dim, keepdim=True) + + self.eps.exp() + ) ** -0.5 + return x * scales + + +class ScaledLinear(nn.Linear): + """ + A modified version of nn.Linear where the parameters are scaled before + use, via: + weight = self.weight * self.weight_scale.exp() + bias = self.bias * self.bias_scale.exp() + + Args: + Accepts the standard args and kwargs that nn.Linear accepts + e.g. in_features, out_features, bias=False. + + initial_scale: you can override this if you want to increase + or decrease the initial magnitude of the module's output + (affects the initialization of weight_scale and bias_scale). + Another option, if you want to do something like this, is + to re-initialize the parameters. + initial_speed: this affects how fast the parameter will + learn near the start of training; you can set it to a + value less than one if you suspect that a module + is contributing to instability near the start of training. + Nnote: regardless of the use of this option, it's best to + use schedulers like Noam that have a warm-up period. + Alternatively you can set it to more than 1 if you want it to + initially train faster. Must be greater than 0. + """ + + def __init__( + self, + *args, + initial_scale: float = 1.0, + initial_speed: float = 1.0, + **kwargs + ): + super(ScaledLinear, self).__init__(*args, **kwargs) + initial_scale = torch.tensor(initial_scale).log() + self.weight_scale = nn.Parameter(initial_scale.clone().detach()) + if self.bias is not None: + self.bias_scale = nn.Parameter(initial_scale.clone().detach()) + else: + self.register_parameter("bias_scale", None) + + self._reset_parameters( + initial_speed + ) # Overrides the reset_parameters in nn.Linear + + def _reset_parameters(self, initial_speed: float): + std = 0.1 / initial_speed + a = (3 ** 0.5) * std + nn.init.uniform_(self.weight, -a, a) + if self.bias is not None: + nn.init.constant_(self.bias, 0.0) + fan_in = self.weight.shape[1] * self.weight[0][0].numel() + scale = fan_in ** -0.5 # 1/sqrt(fan_in) + with torch.no_grad(): + self.weight_scale += torch.tensor(scale / std).log() + + def get_weight(self): + return self.weight * self.weight_scale.exp() + + def get_bias(self): + if self.bias is None or self.bias_scale is None: + return None + else: + return self.bias * self.bias_scale.exp() + + def forward(self, input: Tensor) -> Tensor: + return torch.nn.functional.linear( + input, self.get_weight(), self.get_bias() + ) + + +class ScaledConv1d(nn.Conv1d): + # See docs for ScaledLinear + def __init__( + self, + *args, + initial_scale: float = 1.0, + initial_speed: float = 1.0, + **kwargs + ): + super(ScaledConv1d, self).__init__(*args, **kwargs) + initial_scale = torch.tensor(initial_scale).log() + + self.bias_scale: Optional[nn.Parameter] # for torchscript + + self.weight_scale = nn.Parameter(initial_scale.clone().detach()) + if self.bias is not None: + self.bias_scale = nn.Parameter(initial_scale.clone().detach()) + else: + self.register_parameter("bias_scale", None) + self._reset_parameters( + initial_speed + ) # Overrides the reset_parameters in base class + + def _reset_parameters(self, initial_speed: float): + std = 0.1 / initial_speed + a = (3 ** 0.5) * std + nn.init.uniform_(self.weight, -a, a) + if self.bias is not None: + nn.init.constant_(self.bias, 0.0) + fan_in = self.weight.shape[1] * self.weight[0][0].numel() + scale = fan_in ** -0.5 # 1/sqrt(fan_in) + with torch.no_grad(): + self.weight_scale += torch.tensor(scale / std).log() + + def get_weight(self): + return self.weight * self.weight_scale.exp() + + def get_bias(self): + bias = self.bias + bias_scale = self.bias_scale + if bias is None or bias_scale is None: + return None + else: + return bias * bias_scale.exp() + + def forward(self, input: Tensor) -> Tensor: + F = torch.nn.functional + if self.padding_mode != "zeros": + return F.conv1d( + F.pad( + input, + self._reversed_padding_repeated_twice, + mode=self.padding_mode, + ), + self.get_weight(), + self.get_bias(), + self.stride, + (0,), + self.dilation, + self.groups, + ) + return F.conv1d( + input, + self.get_weight(), + self.get_bias(), + self.stride, + self.padding, + self.dilation, + self.groups, + ) + + +class ScaledConv2d(nn.Conv2d): + # See docs for ScaledLinear + def __init__( + self, + *args, + initial_scale: float = 1.0, + initial_speed: float = 1.0, + **kwargs + ): + super(ScaledConv2d, self).__init__(*args, **kwargs) + initial_scale = torch.tensor(initial_scale).log() + self.weight_scale = nn.Parameter(initial_scale.clone().detach()) + if self.bias is not None: + self.bias_scale = nn.Parameter(initial_scale.clone().detach()) + else: + self.register_parameter("bias_scale", None) + self._reset_parameters( + initial_speed + ) # Overrides the reset_parameters in base class + + def _reset_parameters(self, initial_speed: float): + std = 0.1 / initial_speed + a = (3 ** 0.5) * std + nn.init.uniform_(self.weight, -a, a) + if self.bias is not None: + nn.init.constant_(self.bias, 0.0) + fan_in = self.weight.shape[1] * self.weight[0][0].numel() + scale = fan_in ** -0.5 # 1/sqrt(fan_in) + with torch.no_grad(): + self.weight_scale += torch.tensor(scale / std).log() + + def get_weight(self): + return self.weight * self.weight_scale.exp() + + def get_bias(self): + # see https://github.com/pytorch/pytorch/issues/24135 + bias = self.bias + bias_scale = self.bias_scale + if bias is None or bias_scale is None: + return None + else: + return bias * bias_scale.exp() + + def _conv_forward(self, input, weight): + F = torch.nn.functional + if self.padding_mode != "zeros": + return F.conv2d( + F.pad( + input, + self._reversed_padding_repeated_twice, + mode=self.padding_mode, + ), + weight, + self.get_bias(), + self.stride, + (0, 0), + self.dilation, + self.groups, + ) + return F.conv2d( + input, + weight, + self.get_bias(), + self.stride, + self.padding, + self.dilation, + self.groups, + ) + + def forward(self, input: Tensor) -> Tensor: + return self._conv_forward(input, self.get_weight()) + + +class ActivationBalancer(torch.nn.Module): + """ + Modifies the backpropped derivatives of a function to try to encourage, for + each channel, that it is positive at least a proportion `threshold` of the + time. It does this by multiplying negative derivative values by up to + (1+max_factor), and positive derivative values by up to (1-max_factor), + interpolated from 1 at the threshold to those extremal values when none + of the inputs are positive. + + + Args: + channel_dim: the dimension/axis corresponding to the channel, e.g. + -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative. + min_positive: the minimum, per channel, of the proportion of the time + that (x > 0), below which we start to modify the derivatives. + max_positive: the maximum, per channel, of the proportion of the time + that (x > 0), above which we start to modify the derivatives. + max_factor: the maximum factor by which we modify the derivatives for + either the sign constraint or the magnitude constraint; + e.g. with max_factor=0.02, the the derivatives would be multiplied by + values in the range [0.98..1.02]. + min_abs: the minimum average-absolute-value per channel, which + we allow, before we start to modify the derivatives to prevent + this. + max_abs: the maximum average-absolute-value per channel, which + we allow, before we start to modify the derivatives to prevent + this. + """ + + def __init__( + self, + channel_dim: int, + min_positive: float = 0.05, + max_positive: float = 0.95, + max_factor: float = 0.01, + min_abs: float = 0.2, + max_abs: float = 100.0, + ): + super(ActivationBalancer, self).__init__() + self.channel_dim = channel_dim + self.min_positive = min_positive + self.max_positive = max_positive + self.max_factor = max_factor + self.min_abs = min_abs + self.max_abs = max_abs + + def forward(self, x: Tensor) -> Tensor: + if torch.jit.is_scripting(): + return x + else: + return ActivationBalancerFunction.apply( + x, + self.channel_dim, + self.min_positive, + self.max_positive, + self.max_factor, + self.min_abs, + self.max_abs, + ) + + +class DoubleSwishFunction(torch.autograd.Function): + """ + double_swish(x) = x * torch.sigmoid(x-1) + This is a definition, originally motivated by its close numerical + similarity to swish(swish(x)), where swish(x) = x * sigmoid(x). + + Memory-efficient derivative computation: + double_swish(x) = x * s, where s(x) = torch.sigmoid(x-1) + double_swish'(x) = d/dx double_swish(x) = x * s'(x) + x' * s(x) = x * s'(x) + s(x). + Now, s'(x) = s(x) * (1-s(x)). + double_swish'(x) = x * s'(x) + s(x). + = x * s(x) * (1-s(x)) + s(x). + = double_swish(x) * (1-s(x)) + s(x) + ... so we just need to remember s(x) but not x itself. + """ + + @staticmethod + def forward(ctx, x: Tensor) -> Tensor: + x = x.detach() + s = torch.sigmoid(x - 1.0) + y = x * s + ctx.save_for_backward(s, y) + return y + + @staticmethod + def backward(ctx, y_grad: Tensor) -> Tensor: + s, y = ctx.saved_tensors + return (y * (1 - s) + s) * y_grad + + +class DoubleSwish(torch.nn.Module): + def forward(self, x: Tensor) -> Tensor: + """Return double-swish activation function which is an approximation to Swish(Swish(x)), + that we approximate closely with x * sigmoid(x-1). + """ + if torch.jit.is_scripting(): + return x * torch.sigmoid(x - 1.0) + else: + return DoubleSwishFunction.apply(x) + + +class ScaledEmbedding(nn.Module): + r"""This is a modified version of nn.Embedding that introduces a learnable scale + on the parameters. Note: due to how we initialize it, it's best used with + schedulers like Noam that have a warmup period. + + It is a simple lookup table that stores embeddings of a fixed dictionary and size. + + This module is often used to store word embeddings and retrieve them using indices. + The input to the module is a list of indices, and the output is the corresponding + word embeddings. + + Args: + num_embeddings (int): size of the dictionary of embeddings + embedding_dim (int): the size of each embedding vector + padding_idx (int, optional): If given, pads the output with the embedding vector at :attr:`padding_idx` + (initialized to zeros) whenever it encounters the index. + max_norm (float, optional): If given, each embedding vector with norm larger than :attr:`max_norm` + is renormalized to have norm :attr:`max_norm`. + norm_type (float, optional): The p of the p-norm to compute for the :attr:`max_norm` option. Default ``2``. + scale_grad_by_freq (boolean, optional): If given, this will scale gradients by the inverse of frequency of + the words in the mini-batch. Default ``False``. + sparse (bool, optional): If ``True``, gradient w.r.t. :attr:`weight` matrix will be a sparse tensor. + See Notes for more details regarding sparse gradients. + + initial_speed (float, optional): This affects how fast the parameter will + learn near the start of training; you can set it to a value less than + one if you suspect that a module is contributing to instability near + the start of training. Nnote: regardless of the use of this option, + it's best to use schedulers like Noam that have a warm-up period. + Alternatively you can set it to more than 1 if you want it to + initially train faster. Must be greater than 0. + + + Attributes: + weight (Tensor): the learnable weights of the module of shape (num_embeddings, embedding_dim) + initialized from :math:`\mathcal{N}(0, 1)` + + Shape: + - Input: :math:`(*)`, LongTensor of arbitrary shape containing the indices to extract + - Output: :math:`(*, H)`, where `*` is the input shape and :math:`H=\text{embedding\_dim}` + + .. note:: + Keep in mind that only a limited number of optimizers support + sparse gradients: currently it's :class:`optim.SGD` (`CUDA` and `CPU`), + :class:`optim.SparseAdam` (`CUDA` and `CPU`) and :class:`optim.Adagrad` (`CPU`) + + .. note:: + With :attr:`padding_idx` set, the embedding vector at + :attr:`padding_idx` is initialized to all zeros. However, note that this + vector can be modified afterwards, e.g., using a customized + initialization method, and thus changing the vector used to pad the + output. The gradient for this vector from :class:`~torch.nn.Embedding` + is always zero. + + Examples:: + + >>> # an Embedding module containing 10 tensors of size 3 + >>> embedding = nn.Embedding(10, 3) + >>> # a batch of 2 samples of 4 indices each + >>> input = torch.LongTensor([[1,2,4,5],[4,3,2,9]]) + >>> embedding(input) + tensor([[[-0.0251, -1.6902, 0.7172], + [-0.6431, 0.0748, 0.6969], + [ 1.4970, 1.3448, -0.9685], + [-0.3677, -2.7265, -0.1685]], + + [[ 1.4970, 1.3448, -0.9685], + [ 0.4362, -0.4004, 0.9400], + [-0.6431, 0.0748, 0.6969], + [ 0.9124, -2.3616, 1.1151]]]) + + + >>> # example with padding_idx + >>> embedding = nn.Embedding(10, 3, padding_idx=0) + >>> input = torch.LongTensor([[0,2,0,5]]) + >>> embedding(input) + tensor([[[ 0.0000, 0.0000, 0.0000], + [ 0.1535, -2.0309, 0.9315], + [ 0.0000, 0.0000, 0.0000], + [-0.1655, 0.9897, 0.0635]]]) + + """ + __constants__ = [ + "num_embeddings", + "embedding_dim", + "padding_idx", + "scale_grad_by_freq", + "sparse", + ] + + num_embeddings: int + embedding_dim: int + padding_idx: int + scale_grad_by_freq: bool + weight: Tensor + sparse: bool + + def __init__( + self, + num_embeddings: int, + embedding_dim: int, + padding_idx: Optional[int] = None, + scale_grad_by_freq: bool = False, + sparse: bool = False, + initial_speed: float = 1.0, + ) -> None: + super(ScaledEmbedding, self).__init__() + self.num_embeddings = num_embeddings + self.embedding_dim = embedding_dim + if padding_idx is not None: + if padding_idx > 0: + assert ( + padding_idx < self.num_embeddings + ), "Padding_idx must be within num_embeddings" + elif padding_idx < 0: + assert ( + padding_idx >= -self.num_embeddings + ), "Padding_idx must be within num_embeddings" + padding_idx = self.num_embeddings + padding_idx + self.padding_idx = padding_idx + self.scale_grad_by_freq = scale_grad_by_freq + + self.scale = nn.Parameter(torch.zeros(())) # see reset_parameters() + self.sparse = sparse + + self.weight = nn.Parameter(torch.Tensor(num_embeddings, embedding_dim)) + self.reset_parameters(initial_speed) + + def reset_parameters(self, initial_speed: float = 1.0) -> None: + std = 0.1 / initial_speed + nn.init.normal_(self.weight, std=std) + nn.init.constant_(self.scale, torch.tensor(1.0 / std).log()) + + if self.padding_idx is not None: + with torch.no_grad(): + self.weight[self.padding_idx].fill_(0) + + def forward(self, input: Tensor) -> Tensor: + F = torch.nn.functional + scale = self.scale.exp() + if input.numel() < self.num_embeddings: + return ( + F.embedding( + input, + self.weight, + self.padding_idx, + None, + 2.0, # None, 2.0 relate to normalization + self.scale_grad_by_freq, + self.sparse, + ) + * scale + ) + else: + return F.embedding( + input, + self.weight * scale, + self.padding_idx, + None, + 2.0, # None, 2.0 relates to normalization + self.scale_grad_by_freq, + self.sparse, + ) + + def extra_repr(self) -> str: + s = "{num_embeddings}, {embedding_dim}, scale={scale}" + if self.padding_idx is not None: + s += ", padding_idx={padding_idx}" + if self.scale_grad_by_freq is not False: + s += ", scale_grad_by_freq={scale_grad_by_freq}" + if self.sparse is not False: + s += ", sparse=True" + return s.format(**self.__dict__) + + +def _test_activation_balancer_sign(): + probs = torch.arange(0, 1, 0.01) + N = 1000 + x = 1.0 * (torch.rand(probs.numel(), N) < probs.unsqueeze(-1)) + x = x.detach() + x.requires_grad = True + m = ActivationBalancer( + channel_dim=0, + min_positive=0.05, + max_positive=0.95, + max_factor=0.2, + min_abs=0.0, + ) + + y_grad = torch.sign(torch.randn(probs.numel(), N)) + + y = m(x) + y.backward(gradient=y_grad) + print("_test_activation_balancer_sign: x = ", x) + print("_test_activation_balancer_sign: y grad = ", y_grad) + print("_test_activation_balancer_sign: x grad = ", x.grad) + + +def _test_activation_balancer_magnitude(): + magnitudes = torch.arange(0, 1, 0.01) + N = 1000 + x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze( + -1 + ) + x = x.detach() + x.requires_grad = True + m = ActivationBalancer( + channel_dim=0, + min_positive=0.0, + max_positive=1.0, + max_factor=0.2, + min_abs=0.2, + max_abs=0.8, + ) + + y_grad = torch.sign(torch.randn(magnitudes.numel(), N)) + + y = m(x) + y.backward(gradient=y_grad) + print("_test_activation_balancer_magnitude: x = ", x) + print("_test_activation_balancer_magnitude: y grad = ", y_grad) + print("_test_activation_balancer_magnitude: x grad = ", x.grad) + + +def _test_basic_norm(): + num_channels = 128 + m = BasicNorm(num_channels=num_channels, channel_dim=1) + + x = torch.randn(500, num_channels) + + y = m(x) + + assert y.shape == x.shape + x_rms = (x ** 2).mean().sqrt() + y_rms = (y ** 2).mean().sqrt() + print("x rms = ", x_rms) + print("y rms = ", y_rms) + assert y_rms < x_rms + assert y_rms > 0.5 * x_rms + + +def _test_double_swish_deriv(): + x = torch.randn(10, 12, dtype=torch.double) * 0.5 + x.requires_grad = True + m = DoubleSwish() + torch.autograd.gradcheck(m, x) + + +if __name__ == "__main__": + _test_activation_balancer_sign() + _test_activation_balancer_magnitude() + _test_basic_norm() + _test_double_swish_deriv() diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/streaming_decode.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/streaming_decode.py new file mode 100755 index 000000000..52fe34e88 --- /dev/null +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/streaming_decode.py @@ -0,0 +1,693 @@ +#!/usr/bin/env python3 +# Copyright 2022 Xiaomi Corporation (Authors: Wei Kang, Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +Usage: +./pruned_transducer_stateless2/streaming_decode.py \ + --epoch 28 \ + --avg 15 \ + --left-context 32 \ + --decode-chunk-size 8 \ + --right-context 0 \ + --exp-dir ./pruned_transducer_stateless2/exp \ + --decoding_method greedy_search \ + --num-decode-streams 1000 +""" + +import argparse +import logging +import math +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import numpy as np +import sentencepiece as spm +import torch +import torch.nn as nn +#from asr_datamodule import LibriSpeechAsrDataModule +from asr_datamodule import FisherSwbdSpeechAsrDataModule +from decode_stream import DecodeStream +from kaldifeat import Fbank, FbankOptions +from lhotse import CutSet +from torch.nn.utils.rnn import pad_sequence +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + find_checkpoints, + load_checkpoint, +) +from icefall.decode import one_best_decoding +from icefall.utils import ( + AttributeDict, + get_texts, + setup_logger, + store_transcripts, + write_error_stats, +) + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=28, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 0. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless2/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Support only greedy_search and fast_beam_search now. + """, + ) + + parser.add_argument( + "--beam", + type=float, + default=4, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=32, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + + parser.add_argument( + "--decode-chunk-size", + type=int, + default=16, + help="The chunk size for decoding (in frames after subsampling)", + ) + + parser.add_argument( + "--left-context", + type=int, + default=64, + help="left context can be seen during decoding (in frames after subsampling)", + ) + + parser.add_argument( + "--right-context", + type=int, + default=0, + help="right context can be seen during decoding (in frames after subsampling)", + ) + + parser.add_argument( + "--num-decode-streams", + type=int, + default=2000, + help="The number of streams that can be decoded parallel.", + ) + + add_model_arguments(parser) + + return parser + + +def greedy_search( + model: nn.Module, + encoder_out: torch.Tensor, + streams: List[DecodeStream], +) -> List[List[int]]: + + assert len(streams) == encoder_out.size(0) + assert encoder_out.ndim == 3 + + blank_id = model.decoder.blank_id + context_size = model.decoder.context_size + device = model.device + T = encoder_out.size(1) + + decoder_input = torch.tensor( + [stream.hyp[-context_size:] for stream in streams], + device=device, + dtype=torch.int64, + ) + # decoder_out is of shape (N, decoder_out_dim) + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + # logging.info(f"decoder_out shape : {decoder_out.shape}") + + for t in range(T): + # current_encoder_out's shape: (batch_size, 1, encoder_out_dim) + current_encoder_out = encoder_out[:, t : t + 1, :] # noqa + + logits = model.joiner( + current_encoder_out.unsqueeze(2), + decoder_out.unsqueeze(1), + project_input=False, + ) + # logits'shape (batch_size, vocab_size) + logits = logits.squeeze(1).squeeze(1) + + assert logits.ndim == 2, logits.shape + y = logits.argmax(dim=1).tolist() + emitted = False + for i, v in enumerate(y): + if v != blank_id: + streams[i].hyp.append(v) + emitted = True + if emitted: + # update decoder output + decoder_input = torch.tensor( + [stream.hyp[-context_size:] for stream in streams], + device=device, + dtype=torch.int64, + ) + decoder_out = model.decoder( + decoder_input, + need_pad=False, + ) + decoder_out = model.joiner.decoder_proj(decoder_out) + + hyp_tokens = [] + for stream in streams: + hyp_tokens.append(stream.hyp) + return hyp_tokens + + +def fast_beam_search( + model: nn.Module, + encoder_out: torch.Tensor, + processed_lens: torch.Tensor, + decoding_streams: k2.RnntDecodingStreams, +) -> List[List[int]]: + + B, T, C = encoder_out.shape + for t in range(T): + # shape is a RaggedShape of shape (B, context) + # contexts is a Tensor of shape (shape.NumElements(), context_size) + shape, contexts = decoding_streams.get_contexts() + # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64 + contexts = contexts.to(torch.int64) + # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim) + decoder_out = model.decoder(contexts, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + # current_encoder_out is of shape + # (shape.NumElements(), 1, joiner_dim) + # fmt: off + current_encoder_out = torch.index_select( + encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64) + ) + # fmt: on + logits = model.joiner( + current_encoder_out.unsqueeze(2), + decoder_out.unsqueeze(1), + project_input=False, + ) + logits = logits.squeeze(1).squeeze(1) + log_probs = logits.log_softmax(dim=-1) + decoding_streams.advance(log_probs) + + decoding_streams.terminate_and_flush_to_streams() + + lattice = decoding_streams.format_output(processed_lens.tolist()) + best_path = one_best_decoding(lattice) + hyp_tokens = get_texts(best_path) + return hyp_tokens + + +def decode_one_chunk( + params: AttributeDict, + model: nn.Module, + decode_streams: List[DecodeStream], +) -> List[int]: + """Decode one chunk frames of features for each decode_streams and + return the indexes of finished streams in a List. + + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + decode_streams: + A List of DecodeStream, each belonging to a utterance. + Returns: + Return a List containing which DecodeStreams are finished. + """ + device = model.device + + features = [] + feature_lens = [] + states = [] + + rnnt_stream_list = [] + processed_lens = [] + + for stream in decode_streams: + feat, feat_len = stream.get_feature_frames( + params.decode_chunk_size * params.subsampling_factor + ) + features.append(feat) + feature_lens.append(feat_len) + states.append(stream.states) + processed_lens.append(stream.done_frames) + if params.decoding_method == "fast_beam_search": + rnnt_stream_list.append(stream.rnnt_decoding_stream) + + feature_lens = torch.tensor(feature_lens, device=device) + features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS) + + # if T is less than 7 there will be an error in time reduction layer, + # because we subsample features with ((x_len - 1) // 2 - 1) // 2 + # we plus 2 here because we will cut off one frame on each size of + # encoder_embed output as they see invalid paddings. so we need extra 2 + # frames. + tail_length = 7 + (2 + params.right_context) * params.subsampling_factor + if features.size(1) < tail_length: + feature_lens += tail_length - features.size(1) + features = torch.cat( + [ + features, + torch.tensor( + LOG_EPS, dtype=features.dtype, device=device + ).expand( + features.size(0), + tail_length - features.size(1), + features.size(2), + ), + ], + dim=1, + ) + + states = [ + torch.stack([x[0] for x in states], dim=2), + torch.stack([x[1] for x in states], dim=2), + ] + processed_lens = torch.tensor(processed_lens, device=device) + + encoder_out, encoder_out_lens, states = model.encoder.streaming_forward( + x=features, + x_lens=feature_lens, + states=states, + left_context=params.left_context, + right_context=params.right_context, + processed_lens=processed_lens, + ) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + if params.decoding_method == "greedy_search": + hyp_tokens = greedy_search(model, encoder_out, decode_streams) + elif params.decoding_method == "fast_beam_search": + config = k2.RnntDecodingConfig( + vocab_size=params.vocab_size, + decoder_history_len=params.context_size, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + decoding_streams = k2.RnntDecodingStreams(rnnt_stream_list, config) + processed_lens = processed_lens + encoder_out_lens + hyp_tokens = fast_beam_search( + model, encoder_out, processed_lens, decoding_streams + ) + else: + assert False + + states = [torch.unbind(states[0], dim=2), torch.unbind(states[1], dim=2)] + + finished_streams = [] + for i in range(len(decode_streams)): + decode_streams[i].states = [states[0][i], states[1][i]] + decode_streams[i].done_frames += encoder_out_lens[i] + if params.decoding_method == "fast_beam_search": + decode_streams[i].hyp = hyp_tokens[i] + if decode_streams[i].done: + finished_streams.append(i) + + return finished_streams + + +def decode_dataset( + cuts: CutSet, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[List[str], List[str]]]]: + """Decode dataset. + + Args: + cuts: + Lhotse Cutset containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + device = model.device + + opts = FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = 16000 + opts.mel_opts.num_bins = 80 + + log_interval = 50 + + decode_results = [] + # Contain decode streams currently running. + decode_streams = [] + initial_states = model.encoder.get_init_state( + params.left_context, device=device + ) + for num, cut in enumerate(cuts): + # each utterance has a DecodeStream. + decode_stream = DecodeStream( + params=params, + initial_states=initial_states, + decoding_graph=decoding_graph, + device=device, + ) + + audio: np.ndarray = cut.load_audio() + # audio.shape: (1, num_samples) + assert len(audio.shape) == 2 + assert audio.shape[0] == 1, "Should be single channel" + assert audio.dtype == np.float32, audio.dtype + + # The trained model is using normalized samples + assert audio.max() <= 1, "Should be normalized to [-1, 1])" + + samples = torch.from_numpy(audio).squeeze(0) + + fbank = Fbank(opts) + feature = fbank(samples.to(device)) + decode_stream.set_features(feature) + decode_stream.ground_truth = cut.supervisions[0].text + + decode_streams.append(decode_stream) + + while len(decode_streams) >= params.num_decode_streams: + finished_streams = decode_one_chunk( + params=params, model=model, decode_streams=decode_streams + ) + for i in sorted(finished_streams, reverse=True): + hyp = decode_streams[i].hyp + if params.decoding_method == "greedy_search": + hyp = hyp[params.context_size :] # noqa + decode_results.append( + ( + decode_streams[i].ground_truth.split(), + sp.decode(hyp).split(), + ) + ) + del decode_streams[i] + + if num % log_interval == 0: + logging.info(f"Cuts processed until now is {num}.") + + # decode final chunks of last sequences + while len(decode_streams): + finished_streams = decode_one_chunk( + params=params, model=model, decode_streams=decode_streams + ) + for i in sorted(finished_streams, reverse=True): + hyp = decode_streams[i].hyp + if params.decoding_method == "greedy_search": + hyp = hyp[params.context_size :] # noqa + decode_results.append( + ( + decode_streams[i].ground_truth.split(), + sp.decode(hyp).split(), + ) + ) + del decode_streams[i] + + key = "greedy_search" + if params.decoding_method == "fast_beam_search": + key = ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}" + ) + return {key: decode_results} + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + # sort results so we can easily compare the difference between two + # recognition results + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir + / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + FisherSwbdSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + params.res_dir = params.exp_dir / "streaming" / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + # for streaming + params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}" + params.suffix += f"-left-context-{params.left_context}" + params.suffix += f"-right-context-{params.right_context}" + + # for fast_beam_search + if params.decoding_method == "fast_beam_search": + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + # Decoding in streaming requires causal convolution + params.causal_convolution = True + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if start >= 0: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + + model.to(device) + model.eval() + model.device = device + + decoding_graph = None + if params.decoding_method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + fisherswbd = FisherSwbdSpeechAsrDataModule(args) + + test_eval2000_cuts = fisherswbd.test_eval2000_cuts() + test_swbd_cuts = fisherswbd.test_swbd_cuts () + test_callhome_cuts = fisherswbd.test_callhome_cuts() + + test_eval2000_dl = fisherswbd.test_dataloaders(test_eval2000_cuts) + test_swbd_dl = fisherswbd.test_dataloaders(test_swbd_cuts) + test_callhome_dl = fisherswbd.test_dataloaders(test_callhome_cuts) + + test_sets = ["eval2000", "swbd", "callhome"] + test_cuts = [test_eval2000_dl, test_swbd_dl, test_callhome_dl] + + for test_set, test_cut in zip(test_sets, test_cuts): + results_dict = decode_dataset( + cuts=test_cut, + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/test_model.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/test_model.py new file mode 100755 index 000000000..1858d6bf0 --- /dev/null +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/test_model.py @@ -0,0 +1,76 @@ +#!/usr/bin/env python3 +# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +To run this file, do: + + cd icefall/egs/librispeech/ASR + python ./pruned_transducer_stateless/test_model.py +""" + +import torch +from train import get_params, get_transducer_model + + +def test_model(): + params = get_params() + params.vocab_size = 500 + params.blank_id = 0 + params.context_size = 2 + params.unk_id = 2 + + params.dynamic_chunk_training = False + params.short_chunk_size = 25 + params.num_left_chunks = 4 + params.causal_convolution = False + + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + print(f"Number of model parameters: {num_param}") + model.__class__.forward = torch.jit.ignore(model.__class__.forward) + torch.jit.script(model) + + +def test_model_streaming(): + params = get_params() + params.vocab_size = 500 + params.blank_id = 0 + params.context_size = 2 + params.unk_id = 2 + + params.dynamic_chunk_training = True + params.short_chunk_size = 25 + params.num_left_chunks = 4 + params.causal_convolution = True + + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + print(f"Number of model parameters: {num_param}") + model.__class__.forward = torch.jit.ignore(model.__class__.forward) + torch.jit.script(model) + + +def main(): + test_model() + test_model_streaming() + + +if __name__ == "__main__": + main() diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/train.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/train.py new file mode 100755 index 000000000..354bd1d74 --- /dev/null +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/train.py @@ -0,0 +1,1085 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang +# Mingshuang Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./pruned_transducer_stateless2/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 0 \ + --exp-dir pruned_transducer_stateless2/exp \ + --max-duration 300 + +# For mix precision training: + +./pruned_transducer_stateless2/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 0 \ + --use-fp16 1 \ + --exp-dir pruned_transducer_stateless2/exp \ + --max-duration 550 + +# train a streaming model +./pruned_transducer_stateless2/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 0 \ + --exp-dir pruned_transducer_stateless/exp \ + --dynamic-chunk-training 1 \ + --causal-convolution 1 \ + --short-chunk-size 25 \ + --num-left-chunks 4 \ + --max-duration 300 +""" + + +import argparse +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import FisherSwbdSpeechAsrDataModule +from conformer import Conformer +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import Transducer +from optim import Eden, Eve +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import save_checkpoint_with_global_batch_idx +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +LRSchedulerType = Union[ + torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler +] + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--dynamic-chunk-training", + type=str2bool, + default=False, + help="""Whether to use dynamic_chunk_training, if you want a streaming + model, this requires to be True. + """, + ) + + parser.add_argument( + "--causal-convolution", + type=str2bool, + default=False, + help="""Whether to use causal convolution, this requires to be True when + using dynamic_chunk_training. + """, + ) + + parser.add_argument( + "--short-chunk-size", + type=int, + default=25, + help="""Chunk length of dynamic training, the chunk size would be either + max sequence length of current batch or uniformly sampled from (1, short_chunk_size). + """, + ) + + parser.add_argument( + "--num-left-chunks", + type=int, + default=4, + help="How many left context can be seen in chunks when calculating attention.", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=0, + help="""Resume training from from this epoch. + If it is positive, it will load checkpoint from + transducer_stateless2/exp/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless2/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--initial-lr", + type=float, + default=0.003, + help="The initial learning rate. This value should not need to " + "be changed.", + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=6, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" + "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=8000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=20, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warm_step for Noam optimizer. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for conformer + "feature_dim": 80, + "subsampling_factor": 4, + "encoder_dim": 512, + "nhead": 8, + "dim_feedforward": 2048, + "num_encoder_layers": 12, + # parameters for decoder + "decoder_dim": 512, + # parameters for joiner + "joiner_dim": 512, + # parameters for Noam + "model_warm_step": 3000, # arg given to model, not for lrate + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Conformer and Transformer + encoder = Conformer( + num_features=params.feature_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.encoder_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + dynamic_chunk_training=params.dynamic_chunk_training, + short_chunk_size=params.short_chunk_size, + num_left_chunks=params.num_left_chunks, + causal=params.causal_convolution, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=params.encoder_dim, + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=params.encoder_dim, + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is positive, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 0: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: nn.Module, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, + warmup: float = 1.0, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute RNN-T loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Conformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + warmup=warmup, + ) + # after the main warmup step, we keep pruned_loss_scale small + # for the same amount of time (model_warm_step), to avoid + # overwhelming the simple_loss and causing it to diverge, + # in case it had not fully learned the alignment yet. + pruned_loss_scale = ( + 0.0 + if warmup < 1.0 + else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0) + ) + loss = ( + params.simple_loss_scale * simple_loss + + pruned_loss_scale * pruned_loss + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = ( + (feature_lens // params.subsampling_factor).sum().item() + ) + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: nn.Module, + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + warmup=(params.batch_idx_train / params.model_warm_step), + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + scheduler.step_batch(params.batch_idx_train) + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 30: + return + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[0] + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}" + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary( + tb_writer, "train/tot_", params.batch_idx_train + ) + + if batch_idx > 0 and batch_idx % params.valid_interval == 0: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + if params.full_libri is False: + params.valid_interval = 1600 + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + if params.dynamic_chunk_training: + assert ( + params.causal_convolution + ), "dynamic_chunk_training requires causal convolution" + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + checkpoints = load_checkpoint_if_available(params=params, model=model) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank]) + model.device = device + + optimizer = Eve(model.parameters(), lr=params.initial_lr) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + diagnostic = diagnostics.attach_diagnostics(model) + + librispeech = FisherSwbdSpeechAsrDataModule(args) + + train_cuts = fisherswbd.train_fisher_cuts() + train_cuts += fisherswbd.train_swbd_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = fisherswbd.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = fisherswbd.dev_fisher_cuts() + valid_cuts += fisherswbd.dev_swbd_cuts() + valid_dl = fisherswbd.valid_dataloaders(valid_cuts) + + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + warmup=0.0 if params.start_epoch == 0 else 1.0, + ) + + scaler = GradScaler(enabled=params.use_fp16) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs): + scheduler.step_epoch(epoch) + fix_random_seed(params.seed + epoch) + train_dl.sampler.set_epoch(epoch) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: nn.Module, + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, + warmup: float, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 0 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + warmup=warmup, + ) + loss.backward() + optimizer.step() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + + +def main(): + parser = get_parser() + FisherSwbdSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/fisher_swbd/ASR/shared b/egs/fisher_swbd/ASR/shared new file mode 120000 index 000000000..4c5e91438 --- /dev/null +++ b/egs/fisher_swbd/ASR/shared @@ -0,0 +1 @@ +../../../icefall/shared/ \ No newline at end of file From e230654aba3a463bbcf0c8dc9bc22a16973343a0 Mon Sep 17 00:00:00 2001 From: s-mousmita Date: Sat, 23 Jul 2022 10:40:34 -0400 Subject: [PATCH 2/6] updating swbd prep part --- egs/fisher_swbd/ASR/prepare.sh | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/egs/fisher_swbd/ASR/prepare.sh b/egs/fisher_swbd/ASR/prepare.sh index 604eda754..69d8549eb 100755 --- a/egs/fisher_swbd/ASR/prepare.sh +++ b/egs/fisher_swbd/ASR/prepare.sh @@ -91,10 +91,10 @@ if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then log "Stage 2: Prepare SWBD manifests" mkdir -p data/manifests/swbd lhotse prepare switchboard --absolute-paths 1 --omit-silence $dl_dir/LDC97S62 data/manifests/swbd - gzip -d data/manifests/swbd/swbd_supervisions.jsonl > data/manifests/swbd/swbd_supervisions.jsonl.gz - python3 local/normalize_and_filter_supervisions.py data/manifests/swbd/swbd_supervisions.jsonl.gz data/manifests/supervisions_swbd.jsonl.gz + python3 local/normalize_and_filter_supervisions.py data/manifests/swbd/swbd_supervisions_all.jsonl.gz data/manifests/supervisions_swbd.jsonl.gz + cp data/manifests/swbd/swbd_recordings_all.jsonl.gz data/manifests/recordings_swbd.jsonl.gz gzip -d data/manifests/supervisions_swbd.jsonl.gz - cp data/manifests/swbd/swbd_recordings.jsonl data/manifests/recordings_swbd.jsonl + gzip -d data/manifests/recordings_swbd.jsonl.gz fi if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then From 3fb7393b06fbff73461b5b633bf8338708572514 Mon Sep 17 00:00:00 2001 From: s-mousmita Date: Sun, 24 Jul 2022 23:58:35 -0400 Subject: [PATCH 3/6] Update to prepare.sh --- egs/fisher_swbd/ASR/prepare.sh | 9 ++++----- 1 file changed, 4 insertions(+), 5 deletions(-) diff --git a/egs/fisher_swbd/ASR/prepare.sh b/egs/fisher_swbd/ASR/prepare.sh index 69d8549eb..7cb587f3d 100755 --- a/egs/fisher_swbd/ASR/prepare.sh +++ b/egs/fisher_swbd/ASR/prepare.sh @@ -100,12 +100,11 @@ fi if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then mkdir -p data/manifests/eval2000 lhotse prepare eval2000 --absolute-paths 1 $dl_dir data/manifests/eval2000 - gzip -c data/manifests/eval2000/supervisions_unnorm_eval2000.jsonl > data/manifests/eval2000/supervisions_unnorm_eval2000.jsonl.gz - python3 local/normalize_eval2000.py data/manifests/eval2000/supervisions_unnorm_eval2000.jsonl.gz data/manifests/eval2000/supervisions_eval2000.jsonl.gz - gzip -c data/manifests/eval2000/recordings_eval2000.jsonl >data/manifests/eval2000/recordings_eval2000.jsonl.gz - lhotse fix data/manifests/eval2000/recordings_eval2000.jsonl.gz data/manifests/eval2000/supervisions_eval2000.jsonl.gz data/manifests/ + python3 local/normalize_eval2000.py data/manifests/eval2000/eval2000_supervisions_unnorm.jsonl.gz data/manifests/eval2000/supervisions_eval2000.jsonl.gz + lhotse fix data/manifests/eval2000/eval2000_recordings_all.jsonl.gz data/manifests/eval2000/supervisions_eval2000.jsonl.gz data/manifests + mv data/manifests/eval2000_recordings_all.jsonl.gz data/manifests/recordings_eval2000.jsonl.gz gzip -d data/manifests/recordings_eval2000.jsonl.gz - gzip -d data/manifests/supervisions_eval2000.jsonl.gz + gzip -d data/manifests/supervisions_eval2000.jsonl.gz fi if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then From 34524acf44d1810147374964b08bfe036a7ae3ba Mon Sep 17 00:00:00 2001 From: s-mousmita Date: Mon, 25 Jul 2022 07:03:54 -0400 Subject: [PATCH 4/6] Removing some run time comment --- .../ASR/local/compute_fbank_fisher_swbd_eval2000.py | 9 +-------- 1 file changed, 1 insertion(+), 8 deletions(-) diff --git a/egs/fisher_swbd/ASR/local/compute_fbank_fisher_swbd_eval2000.py b/egs/fisher_swbd/ASR/local/compute_fbank_fisher_swbd_eval2000.py index f37cabc49..eb6ff6d83 100755 --- a/egs/fisher_swbd/ASR/local/compute_fbank_fisher_swbd_eval2000.py +++ b/egs/fisher_swbd/ASR/local/compute_fbank_fisher_swbd_eval2000.py @@ -47,18 +47,11 @@ def compute_fbank_fisher_swbd_eval2000(): num_jobs = min(25, os.cpu_count()) num_mel_bins = 80 sampling_rate = 8000 - dataset_parts = ( - "swbd", - ) - - """ dataset_parts = ( "eval2000", "fisher", "swbd", - ) - """ - + ) test_dataset=("eval2000",) manifests = read_manifests_if_cached( dataset_parts=dataset_parts, output_dir=src_dir, lazy=True, suffix="jsonl" From 67e3607863807b19037c9b371624f111e76f9c63 Mon Sep 17 00:00:00 2001 From: s-mousmita Date: Tue, 26 Jul 2022 10:30:34 -0400 Subject: [PATCH 5/6] Updating musan part to make compatible wiht recent lhotse nomenclature --- .../ASR/local/compute_fbank_musan.py | 23 ++++++++++++++----- egs/fisher_swbd/ASR/prepare.sh | 7 +----- 2 files changed, 18 insertions(+), 12 deletions(-) diff --git a/egs/fisher_swbd/ASR/local/compute_fbank_musan.py b/egs/fisher_swbd/ASR/local/compute_fbank_musan.py index f97002b82..acb477540 100755 --- a/egs/fisher_swbd/ASR/local/compute_fbank_musan.py +++ b/egs/fisher_swbd/ASR/local/compute_fbank_musan.py @@ -28,7 +28,7 @@ import os from pathlib import Path import torch -from lhotse import CutSet, Fbank, FbankConfig, combine +from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter, combine from lhotse.recipes.utils import read_manifests_if_cached from icefall.utils import get_executor @@ -52,12 +52,22 @@ def compute_fbank_musan(): "speech", "noise", ) + prefix = "musan" + suffix = "jsonl.gz" manifests = read_manifests_if_cached( - dataset_parts=dataset_parts, output_dir=src_dir, lazy=True, suffix="jsonl" + dataset_parts=dataset_parts, + output_dir=src_dir, + prefix=prefix, + suffix=suffix, ) assert manifests is not None - musan_cuts_path = output_dir / "cuts_musan.json.gz" + assert len(manifests) == len(dataset_parts), ( + len(manifests), + len(dataset_parts), + ) + + musan_cuts_path = output_dir / "musan_cuts.jsonl.gz" if musan_cuts_path.is_file(): logging.info(f"{musan_cuts_path} already exists - skipping") @@ -65,7 +75,7 @@ def compute_fbank_musan(): logging.info("Extracting features for Musan") - extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins, sampling_rate=sampling_rate)) + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins, sampling_rate=sampling_rate) with get_executor() as ex: # Initialize the executor only once. # create chunks of Musan with duration 5 - 10 seconds @@ -79,12 +89,13 @@ def compute_fbank_musan(): .filter(lambda c: c.duration > 5) .compute_and_store_features( extractor=extractor, - storage_path=f"{output_dir}/feats_musan", + storage_path=f"{output_dir}/musan_feats", num_jobs=num_jobs if ex is None else 80, executor=ex, + storage_type=LilcomChunkyWriter, ) ) - musan_cuts.to_json(musan_cuts_path) + musan_cuts.to_file(musan_cuts_path) if __name__ == "__main__": diff --git a/egs/fisher_swbd/ASR/prepare.sh b/egs/fisher_swbd/ASR/prepare.sh index 7cb587f3d..787e555fc 100755 --- a/egs/fisher_swbd/ASR/prepare.sh +++ b/egs/fisher_swbd/ASR/prepare.sh @@ -183,12 +183,7 @@ if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then # We assume that you have downloaded the musan corpus # to data/musan mkdir -p data/manifests/musan - lhotse prepare musan $dl_dir/musan data/manifests/musan - - for name in music noise speech ; do - jq -c '.[]' data/manifests/musan/recordings_${name}.json > data/manifests/recordings_${name}.jsonl - gzip -c data/manifests/recordings_${name}.jsonl > data/manifests/recordings_${name}.jsonl.gz - done + lhotse prepare musan $dl_dir/musan data/manifests/musan fi if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then From 83e2b30a224832d738235743ba88a70a7262e361 Mon Sep 17 00:00:00 2001 From: s-mousmita Date: Thu, 28 Jul 2022 07:58:20 -0400 Subject: [PATCH 6/6] adding black reformatted files --- .../compute_fbank_fisher_swbd_eval2000.py | 25 +-- .../ASR/local/compute_fbank_musan.py | 10 +- .../ASR/local/extract_json_cuts.py | 72 +++---- .../ASR/local/extract_json_supervision.py | 33 ++- .../ASR/local/extract_list_of_sph.py | 34 +-- .../ASR/local/normalize_eval2000.py | 199 +++++++++--------- egs/fisher_swbd/ASR/local/prepare_lang_bpe.py | 8 +- .../ASR/local/prepare_lang_g2pen.py | 15 +- egs/fisher_swbd/ASR/local/train_bpe_model.py | 10 +- .../asr_datamodule.py | 36 ++-- .../beam_search.py | 35 +-- .../pruned_transducer_stateless2/conformer.py | 34 +-- .../pruned_transducer_stateless2/decode.py | 10 +- .../decode_stream.py | 5 +- .../pretrained.py | 6 +- .../streaming_decode.py | 27 +-- .../ASR/pruned_transducer_stateless2/train.py | 30 +-- 17 files changed, 240 insertions(+), 349 deletions(-) diff --git a/egs/fisher_swbd/ASR/local/compute_fbank_fisher_swbd_eval2000.py b/egs/fisher_swbd/ASR/local/compute_fbank_fisher_swbd_eval2000.py index eb6ff6d83..d2bb3e3e2 100755 --- a/egs/fisher_swbd/ASR/local/compute_fbank_fisher_swbd_eval2000.py +++ b/egs/fisher_swbd/ASR/local/compute_fbank_fisher_swbd_eval2000.py @@ -1,5 +1,6 @@ #!/usr/bin/env python3 -# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) + + # # See ../../../../LICENSE for clarification regarding multiple authors # @@ -47,18 +48,19 @@ def compute_fbank_fisher_swbd_eval2000(): num_jobs = min(25, os.cpu_count()) num_mel_bins = 80 sampling_rate = 8000 - dataset_parts = ( - "eval2000", - "fisher", - "swbd", - ) - test_dataset=("eval2000",) + dataset_parts = ("eval2000", "fisher", "swbd") + test_dataset = ("eval2000",) manifests = read_manifests_if_cached( - dataset_parts=dataset_parts, output_dir=src_dir, lazy=True, suffix="jsonl" + dataset_parts=dataset_parts, + output_dir=src_dir, + lazy=True, + suffix="jsonl", ) assert manifests is not None - extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins, sampling_rate=sampling_rate)) + extractor = Fbank( + FbankConfig(num_mel_bins=num_mel_bins, sampling_rate=sampling_rate) + ) with get_executor() as ex: # Initialize the executor only once. for partition, m in manifests.items(): @@ -67,10 +69,9 @@ def compute_fbank_fisher_swbd_eval2000(): continue logging.info(f"Processing {partition}") cut_set = CutSet.from_manifests( - recordings=m["recordings"], - supervisions=m["supervisions"], + recordings=m["recordings"], supervisions=m["supervisions"] ) - #if "train" in partition: + # if "train" in partition: if partition not in test_dataset: logging.info(f"Adding speed perturbations to : {partition}") cut_set = ( diff --git a/egs/fisher_swbd/ASR/local/compute_fbank_musan.py b/egs/fisher_swbd/ASR/local/compute_fbank_musan.py index acb477540..5f217fd63 100755 --- a/egs/fisher_swbd/ASR/local/compute_fbank_musan.py +++ b/egs/fisher_swbd/ASR/local/compute_fbank_musan.py @@ -47,11 +47,7 @@ def compute_fbank_musan(): num_jobs = min(15, os.cpu_count()) num_mel_bins = 80 sampling_rate = 8000 - dataset_parts = ( - "music", - "speech", - "noise", - ) + dataset_parts = ("music", "speech", "noise") prefix = "musan" suffix = "jsonl.gz" manifests = read_manifests_if_cached( @@ -75,7 +71,9 @@ def compute_fbank_musan(): logging.info("Extracting features for Musan") - extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins, sampling_rate=sampling_rate) + extractor = Fbank( + FbankConfig(num_mel_bins=num_mel_bins, sampling_rate=sampling_rate) + ) with get_executor() as ex: # Initialize the executor only once. # create chunks of Musan with duration 5 - 10 seconds diff --git a/egs/fisher_swbd/ASR/local/extract_json_cuts.py b/egs/fisher_swbd/ASR/local/extract_json_cuts.py index 932f14714..df200c4ea 100644 --- a/egs/fisher_swbd/ASR/local/extract_json_cuts.py +++ b/egs/fisher_swbd/ASR/local/extract_json_cuts.py @@ -1,58 +1,56 @@ #!/usr/bin/env python3 - -# +# script to extract cutids corresponding to a list of source audio files. +# It takes three arguments: list of audio (.sph) , cut jsonl and out jsonl -import sys, json ; -import ntpath; +import sys, json +import ntpath -list_of_sph = sys.argv[1]; -jsonfile = sys.argv[2]; -out_partition_json = sys.argv[3]; +list_of_sph = sys.argv[1] +jsonfile = sys.argv[2] +out_partition_json = sys.argv[3] -list_of_sph=[line.rstrip('\n') for line in open(list_of_sph)] +list_of_sph = [line.rstrip("\n") for line in open(list_of_sph)] -sph_basename_list=[] +sph_basename_list = [] for f in list_of_sph: - bsname=ntpath.basename(f) - #print(bsname) + bsname = ntpath.basename(f) sph_basename_list.append(ntpath.basename(f)) -json_str=[line.rstrip('\n') for line in open(jsonfile)] +json_str = [line.rstrip("\n") for line in open(jsonfile)] num_json = len(json_str) -#cutid2sph=dict() -out_partition=open(out_partition_json,'w',encoding='utf-8') +out_partition = open(out_partition_json, "w", encoding="utf-8") for i in range(num_json): - if json_str[i] != '': - #print(json_str[i]) + if json_str[i] != "": + # print(json_str[i]) cur_json = json.loads(json_str[i]) - #print(cur_json) - cur_cutid= cur_json['id'] - cur_rec = cur_json['recording'] - cur_sources = cur_rec['sources'] - #print(cur_cutid) - #print(cur_rec) - #print(cur_sources) + # print(cur_json) + cur_cutid = cur_json["id"] + cur_rec = cur_json["recording"] + cur_sources = cur_rec["sources"] + # print(cur_cutid) + # print(cur_rec) + # print(cur_sources) for s in cur_sources: - cur_sph = s['source'] - cur_sph_basename=ntpath.basename(cur_sph) - #print(cur_sph) - #print(cur_sph_basename) - if cur_sph_basename in sph_basename_list : + cur_sph = s["source"] + cur_sph_basename = ntpath.basename(cur_sph) + # print(cur_sph) + # print(cur_sph_basename) + if cur_sph_basename in sph_basename_list: out_json_line = json_str[i] out_partition.write(out_json_line) out_partition.write("\n") - #for keys in cur_json: - #cur_cutid= cur_json['id'] - #cur_rec = cur_json['recording_id'] - #print(cur_cutid) - - + # for keys in cur_json: + # cur_cutid= cur_json['id'] + # cur_rec = cur_json['recording_id'] + # print(cur_cutid) + + """ for keys in cur_json: #print(keys) @@ -64,9 +62,3 @@ for i in range(num_json): out_partition.write(out_json_line) out_partition.write("\n") """ - - - - - - diff --git a/egs/fisher_swbd/ASR/local/extract_json_supervision.py b/egs/fisher_swbd/ASR/local/extract_json_supervision.py index 8dd714f23..0241e8012 100644 --- a/egs/fisher_swbd/ASR/local/extract_json_supervision.py +++ b/egs/fisher_swbd/ASR/local/extract_json_supervision.py @@ -1,40 +1,35 @@ #!/usr/bin/env python3 -# +# -import sys, json ; -import ntpath; +import sys, json +import ntpath -list_of_sph = sys.argv[1]; -jsonfile = sys.argv[2]; -out_partition_json = sys.argv[3]; +list_of_sph = sys.argv[1] +jsonfile = sys.argv[2] +out_partition_json = sys.argv[3] -list_of_sph=[line.rstrip('\n') for line in open(list_of_sph)] +list_of_sph = [line.rstrip("\n") for line in open(list_of_sph)] -sph_basename_list=[] +sph_basename_list = [] for f in list_of_sph: - bsname=ntpath.basename(f) - #print(bsname) + bsname = ntpath.basename(f) sph_basename_list.append(ntpath.basename(f)) -json_str=[line.rstrip('\n') for line in open(jsonfile)] +json_str = [line.rstrip("\n") for line in open(jsonfile)] num_json = len(json_str) -out_partition=open(out_partition_json,'w',encoding='utf-8') +out_partition = open(out_partition_json, "w", encoding="utf-8") for i in range(num_json): - if json_str[i] != '': - #print(json_str[i]) + if json_str[i] != "": cur_json = json.loads(json_str[i]) - #print(cur_json) - cur_rec = cur_json['recording_id'] - #print(cur_rec) + cur_rec = cur_json["recording_id"] cur_sph_basename = cur_rec + ".sph" - #print(cur_sph_basename) - if cur_sph_basename in sph_basename_list : + if cur_sph_basename in sph_basename_list: out_json_line = json_str[i] out_partition.write(out_json_line) out_partition.write("\n") diff --git a/egs/fisher_swbd/ASR/local/extract_list_of_sph.py b/egs/fisher_swbd/ASR/local/extract_list_of_sph.py index 708291820..d80cb8f6e 100644 --- a/egs/fisher_swbd/ASR/local/extract_list_of_sph.py +++ b/egs/fisher_swbd/ASR/local/extract_list_of_sph.py @@ -1,38 +1,20 @@ #!/usr/bin/env python3 - +# extract list of sph from a cut jsonl # python3 extract_list_of_sph.py dev_cuts_swbd.jsonl > data/fbank/dev_swbd_sph.list -import sys, json ; +import sys, json + inputfile = sys.argv[1] -json_str=[line.rstrip('\n') for line in open(inputfile)] +json_str = [line.rstrip("\n") for line in open(inputfile)] num_json = len(json_str) -#print(num_json) -#with open(inputfile, 'r',encoding='utf-8') as Jsonfile: -# print("Converting JSON encoded data into Python dictionary") -# json_dict = json.load(Jsonfile) -# for k,v in json_dict: -# print(k,v) - - - for i in range(num_json): - if json_str[i] != '': - #print(json_str[i]) + if json_str[i] != "": cur_json = json.loads(json_str[i]) - # print(cur_json) for keys in cur_json: - #print(keys) - cur_rec = cur_json['recording'] - cur_sources = cur_rec['sources'] - #print(cur_sources) + cur_rec = cur_json["recording"] + cur_sources = cur_rec["sources"] for s in cur_sources: - cur_sph = s['source'] + cur_sph = s["source"] print(cur_sph) - #cur_sph = cur_sources[2] - #print(cur_sph) - - - -#print(json.load(sys.stdin)['source']) diff --git a/egs/fisher_swbd/ASR/local/normalize_eval2000.py b/egs/fisher_swbd/ASR/local/normalize_eval2000.py index 244f20ea4..76efa176d 100644 --- a/egs/fisher_swbd/ASR/local/normalize_eval2000.py +++ b/egs/fisher_swbd/ASR/local/normalize_eval2000.py @@ -16,67 +16,72 @@ def get_args(): parser.add_argument("output_sups") return parser.parse_args() -def remove_punctutation_and_other_symbol(text:str) -> str: - text = text.replace("--"," ") - text = text.replace("//"," ") - text = text.replace("."," ") - text = text.replace("?"," ") - text = text.replace("~"," ") - text = text.replace(","," ") - text = text.replace(";"," ") - text = text.replace("("," ") - text = text.replace(")"," ") - text = text.replace("&"," ") - text = text.replace("%"," ") - text = text.replace("*"," ") - text = text.replace("{"," ") - text = text.replace("}"," ") + +def remove_punctutation_and_other_symbol(text: str) -> str: + text = text.replace("--", " ") + text = text.replace("//", " ") + text = text.replace(".", " ") + text = text.replace("?", " ") + text = text.replace("~", " ") + text = text.replace(",", " ") + text = text.replace(";", " ") + text = text.replace("(", " ") + text = text.replace(")", " ") + text = text.replace("&", " ") + text = text.replace("%", " ") + text = text.replace("*", " ") + text = text.replace("{", " ") + text = text.replace("}", " ") return text + def eval2000_clean_eform(text: str, eform_count) -> str: string_to_remove = [] - piece=text.split("\">") - for i in range(0,len(piece)): - s=piece[i]+"\">" - res = re.search(r'', s) + piece = text.split('">') + for i in range(0, len(piece)): + s = piece[i] + '">' + res = re.search(r"", s) if res is not None: - res_rm= res.group(1) + res_rm = res.group(1) string_to_remove.append(res_rm) for p in string_to_remove: eform_string = p text = text.replace(eform_string, " ") eform_1 = "" - text = text.replace(eform_2," ") - #print("TEXT final: ", text) + eform_2 = '">' + text = text.replace(eform_2, " ") + # print("TEXT final: ", text) return text - -def replace_silphone(text: str) -> str: + + +def replace_silphone(text: str) -> str: text = text.replace("[/BABY CRYING]", " ") - text = text.replace("[/CHILD]" , " ") - text = text.replace("[[DISTORTED]]" , " ") - text = text.replace("[/DISTORTION]" , " ") - text = text.replace("[[DRAWN OUT]]" , " ") - text = text.replace("[[DRAWN-OUT]]" , " ") - text = text.replace("[[FAINT]]" , " ") - text = text.replace("[SMACK]" , " ") - text = text.replace("[[MUMBLES]]" , " ") - text = text.replace("[[HIGH PITCHED SQUEAKY VOICE]]" , " ") - text = text.replace("[[IN THE LAUGH]]" , "[LAUGHTER]") - text = text.replace("[[LAST WORD SPOKEN WITH A LAUGH]]" , "[LAUGHTER]") - text = text.replace("[[PART OF FIRST SYLLABLE OF PREVIOUS WORD CUT OFF]]" , " ") - text = text.replace("[[PREVIOUS WORD SPOKEN WITH A LAUGH]]" , " ") - text = text.replace("[[PREVIOUS TWO WORDS SPOKEN WHILE LAUGHING]]" , " ") - text = text.replace("[[PROLONGED]]" , " ") - text = text.replace("[/RUNNING WATER]" , " ") - text = text.replace("[[SAYS LAUGHING]]" , "[LAUGHTER]") - text = text.replace("[[SINGING]]" , " ") - text = text.replace("[[SPOKEN WHILE LAUGHING]]" , "[LAUGHTER]") - text = text.replace("[/STATIC]" , " ") - text = text.replace("['THIRTIETH' DRAWN OUT]" , " ") - text = text.replace("[/VOICES]" , " ") - text = text.replace("[[WHISPERED]]" , " ") + text = text.replace("[/CHILD]", " ") + text = text.replace("[[DISTORTED]]", " ") + text = text.replace("[/DISTORTION]", " ") + text = text.replace("[[DRAWN OUT]]", " ") + text = text.replace("[[DRAWN-OUT]]", " ") + text = text.replace("[[FAINT]]", " ") + text = text.replace("[SMACK]", " ") + text = text.replace("[[MUMBLES]]", " ") + text = text.replace("[[HIGH PITCHED SQUEAKY VOICE]]", " ") + text = text.replace("[[IN THE LAUGH]]", "[LAUGHTER]") + text = text.replace("[[LAST WORD SPOKEN WITH A LAUGH]]", "[LAUGHTER]") + text = text.replace( + "[[PART OF FIRST SYLLABLE OF PREVIOUS WORD CUT OFF]]", " " + ) + text = text.replace("[[PREVIOUS WORD SPOKEN WITH A LAUGH]]", " ") + text = text.replace("[[PREVIOUS TWO WORDS SPOKEN WHILE LAUGHING]]", " ") + text = text.replace("[[PROLONGED]]", " ") + text = text.replace("[/RUNNING WATER]", " ") + text = text.replace("[[SAYS LAUGHING]]", "[LAUGHTER]") + text = text.replace("[[SINGING]]", " ") + text = text.replace("[[SPOKEN WHILE LAUGHING]]", "[LAUGHTER]") + text = text.replace("[/STATIC]", " ") + text = text.replace("['THIRTIETH' DRAWN OUT]", " ") + text = text.replace("[/VOICES]", " ") + text = text.replace("[[WHISPERED]]", " ") text = text.replace("[DISTORTION]", " ") text = text.replace("[DISTORTION, HIGH VOLUME ON WAVES]", " ") text = text.replace("[BACKGROUND LAUGHTER]", "[LAUGHTER]") @@ -95,24 +100,24 @@ def replace_silphone(text: str) -> str: text = text.replace("[BABY CRYING]", " ") text = text.replace("[METALLIC KNOCKING SOUND]", " ") text = text.replace("[METALLIC SOUND]", " ") - + text = text.replace("[PHONE JIGGLING]", " ") text = text.replace("[BACKGROUND SOUND]", " ") text = text.replace("[BACKGROUND VOICE]", " ") - text = text.replace("[BACKGROUND VOICES]", " ") + text = text.replace("[BACKGROUND VOICES]", " ") text = text.replace("[BACKGROUND NOISE]", " ") text = text.replace("[CAR HORNS IN BACKGROUND]", " ") text = text.replace("[CAR HORNS]", " ") - text = text.replace("[CARNATING]", " ") + text = text.replace("[CARNATING]", " ") text = text.replace("[CRYING CHILD]", " ") text = text.replace("[CHOPPING SOUND]", " ") text = text.replace("[BANGING]", " ") text = text.replace("[CLICKING NOISE]", " ") - text = text.replace("[CLATTERING]", " ") + text = text.replace("[CLATTERING]", " ") text = text.replace("[ECHO]", " ") - text = text.replace("[KNOCK]", " ") - text = text.replace("[NOISE-GOOD]", "[NOISE]") - text = text.replace("[RIGHT]", " ") + text = text.replace("[KNOCK]", " ") + text = text.replace("[NOISE-GOOD]", "[NOISE]") + text = text.replace("[RIGHT]", " ") text = text.replace("[SOUND]", " ") text = text.replace("[SQUEAK]", " ") text = text.replace("[STATIC]", " ") @@ -131,64 +136,65 @@ def replace_silphone(text: str) -> str: text = text.replace("Y[OU]I-", "YOU I") text = text.replace("-[A]ND", "AND") text = text.replace("JU[ST]", "JUST") - text = text.replace("{BREATH}" , " ") - text = text.replace("{BREATHY}" , " ") - text = text.replace("{CHANNEL NOISE}" , " ") - text = text.replace("{CLEAR THROAT}" , " ") + text = text.replace("{BREATH}", " ") + text = text.replace("{BREATHY}", " ") + text = text.replace("{CHANNEL NOISE}", " ") + text = text.replace("{CLEAR THROAT}", " ") - text = text.replace("{CLEARING THROAT}" , " ") - text = text.replace("{CLEARS THROAT}" , " ") - text = text.replace("{COUGH}" , " ") - text = text.replace("{DRAWN OUT}" , " ") - text = text.replace("{EXHALATION}" , " ") - text = text.replace("{EXHALE}" , " ") - text = text.replace("{GASP}" , " ") - text = text.replace("{HIGH SQUEAL}" , " ") - text = text.replace("{INHALE}" , " ") - text = text.replace("{LAUGH}" , "[LAUGHTER]") - text = text.replace("{LAUGH}" , "[LAUGHTER]") - text = text.replace("{LAUGH}" , "[LAUGHTER]") - text = text.replace("{LIPSMACK}" , " ") - text = text.replace("{LIPSMACK}" , " ") + text = text.replace("{CLEARING THROAT}", " ") + text = text.replace("{CLEARS THROAT}", " ") + text = text.replace("{COUGH}", " ") + text = text.replace("{DRAWN OUT}", " ") + text = text.replace("{EXHALATION}", " ") + text = text.replace("{EXHALE}", " ") + text = text.replace("{GASP}", " ") + text = text.replace("{HIGH SQUEAL}", " ") + text = text.replace("{INHALE}", " ") + text = text.replace("{LAUGH}", "[LAUGHTER]") + text = text.replace("{LAUGH}", "[LAUGHTER]") + text = text.replace("{LAUGH}", "[LAUGHTER]") + text = text.replace("{LIPSMACK}", " ") + text = text.replace("{LIPSMACK}", " ") - text = text.replace("{NOISE OF DISGUST}" , " ") - text = text.replace("{SIGH}" , " ") - text = text.replace("{SNIFF}" , " ") - text = text.replace("{SNORT}" , " ") - text = text.replace("{SHARP EXHALATION}" , " ") - text = text.replace("{BREATH LAUGH}" , " ") + text = text.replace("{NOISE OF DISGUST}", " ") + text = text.replace("{SIGH}", " ") + text = text.replace("{SNIFF}", " ") + text = text.replace("{SNORT}", " ") + text = text.replace("{SHARP EXHALATION}", " ") + text = text.replace("{BREATH LAUGH}", " ") return text -def remove_languagetag(text:str) -> str: - langtag = re.findall(r'<(.*?)>', text) + +def remove_languagetag(text: str) -> str: + langtag = re.findall(r"<(.*?)>", text) for t in langtag: text = text.replace(t, " ") - text = text.replace("<"," ") - text = text.replace(">"," ") + text = text.replace("<", " ") + text = text.replace(">", " ") return text - + + def eval2000_normalizer(text: str) -> str: - #print("TEXT original: ",text) - eform_count=text.count("contraction e_form") - #print("eform corunt:", eform_count) - if eform_count>0: - text = eval2000_clean_eform(text,eform_count) + # print("TEXT original: ",text) + eform_count = text.count("contraction e_form") + # print("eform corunt:", eform_count) + if eform_count > 0: + text = eval2000_clean_eform(text, eform_count) text = text.upper() text = remove_languagetag(text) text = replace_silphone(text) text = remove_punctutation_and_other_symbol(text) text = text.replace("IGNORE_TIME_SEGMENT_IN_SCORING", " ") text = text.replace("IGNORE_TIME_SEGMENT_SCORING", " ") - spaces = re.findall(r'\s+', text) + spaces = re.findall(r"\s+", text) for sp in spaces: - text = text.replace(sp," ") - text = text.strip() - #text = self.whitespace_regexp.sub(" ", text).strip() - #print(text) + text = text.replace(sp, " ") + text = text.strip() + # text = self.whitespace_regexp.sub(" ", text).strip() + # print(text) return text - def main(): args = get_args() sups = load_manifest_lazy_or_eager(args.input_sups) @@ -203,6 +209,7 @@ def main(): skip += 1 continue writer.write(sup) - + + if __name__ == "__main__": main() diff --git a/egs/fisher_swbd/ASR/local/prepare_lang_bpe.py b/egs/fisher_swbd/ASR/local/prepare_lang_bpe.py index b3f012e84..431ebd439 100755 --- a/egs/fisher_swbd/ASR/local/prepare_lang_bpe.py +++ b/egs/fisher_swbd/ASR/local/prepare_lang_bpe.py @@ -108,9 +108,7 @@ def lexicon_to_fst_no_sil( disambig_token = token2id["#0"] disambig_word = word2id["#0"] arcs = add_self_loops( - arcs, - disambig_token=disambig_token, - disambig_word=disambig_word, + arcs, disambig_token=disambig_token, disambig_word=disambig_word ) final_state = next_state @@ -223,9 +221,7 @@ def main(): write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig) L = lexicon_to_fst_no_sil( - lexicon, - token2id=token_sym_table, - word2id=word_sym_table, + lexicon, token2id=token_sym_table, word2id=word_sym_table ) L_disambig = lexicon_to_fst_no_sil( diff --git a/egs/fisher_swbd/ASR/local/prepare_lang_g2pen.py b/egs/fisher_swbd/ASR/local/prepare_lang_g2pen.py index 0549d7306..6a504d4c6 100755 --- a/egs/fisher_swbd/ASR/local/prepare_lang_g2pen.py +++ b/egs/fisher_swbd/ASR/local/prepare_lang_g2pen.py @@ -68,12 +68,7 @@ def get_args(): def get_g2p_sym2int(): # These symbols are removed from from g2p_en's vocabulary - excluded_symbols = [ - "", - "", - "", - "", - ] + excluded_symbols = ["", "", "", ""] symbols = [p for p in sorted(G2p().phonemes) if p not in excluded_symbols] # reserve 0 and 1 for blank and sos/eos/pad tokens @@ -345,9 +340,7 @@ def lexicon_to_fst( disambig_token = token2id["#0"] disambig_word = word2id["#0"] arcs = add_self_loops( - arcs, - disambig_token=disambig_token, - disambig_word=disambig_word, + arcs, disambig_token=disambig_token, disambig_word=disambig_word ) final_state = next_state @@ -396,9 +389,7 @@ def main(): print(vocab[:10]) if not lexicon_filename.is_file(): - lexicon = [ - ("!SIL", [sil_token]), - ] + lexicon = [("!SIL", [sil_token])] for symbol in special_symbols: lexicon.append((symbol, [symbol[1:-1]])) lexicon += [ diff --git a/egs/fisher_swbd/ASR/local/train_bpe_model.py b/egs/fisher_swbd/ASR/local/train_bpe_model.py index bc5812810..2581a90cb 100755 --- a/egs/fisher_swbd/ASR/local/train_bpe_model.py +++ b/egs/fisher_swbd/ASR/local/train_bpe_model.py @@ -43,16 +43,10 @@ def get_args(): """, ) - parser.add_argument( - "--transcript", - type=str, - help="Training transcript.", - ) + parser.add_argument("--transcript", type=str, help="Training transcript.") parser.add_argument( - "--vocab-size", - type=int, - help="Vocabulary size for BPE training", + "--vocab-size", type=int, help="Vocabulary size for BPE training" ) return parser.parse_args() diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/asr_datamodule.py index 706cdf34c..ef255cc47 100644 --- a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/asr_datamodule.py +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/asr_datamodule.py @@ -1,6 +1,6 @@ # Copyright 2021 Piotr Żelasko # Copyright 2022 Xiaomi Corporation (Author: Mingshuang Luo) -# +# # # See ../../../../LICENSE for clarification regarding multiple authors # @@ -356,13 +356,10 @@ class FisherSwbdSpeechAsrDataModule: ) else: validate = K2SpeechRecognitionDataset( - cut_transforms=transforms, - return_cuts=self.args.return_cuts, + cut_transforms=transforms, return_cuts=self.args.return_cuts ) valid_sampler = DynamicBucketingSampler( - cuts_valid, - max_duration=self.args.max_duration, - shuffle=False, + cuts_valid, max_duration=self.args.max_duration, shuffle=False ) logging.info("About to create dev dataloader") valid_dl = DataLoader( @@ -384,9 +381,7 @@ class FisherSwbdSpeechAsrDataModule: return_cuts=self.args.return_cuts, ) sampler = DynamicBucketingSampler( - cuts, - max_duration=self.args.max_duration, - shuffle=False, + cuts, max_duration=self.args.max_duration, shuffle=False ) logging.debug("About to create test dataloader") test_dl = DataLoader( @@ -396,41 +391,52 @@ class FisherSwbdSpeechAsrDataModule: num_workers=self.args.num_workers, ) return test_dl - + @lru_cache() def train_fisher_cuts(self) -> CutSet: logging.info("About to get fisher cuts") return load_manifest_lazy( self.args.manifest_dir / "train_cuts_fisher.jsonl.gz" ) + @lru_cache() def train_swbd_cuts(self) -> CutSet: logging.info("About to get train swbd cuts") return load_manifest_lazy( self.args.manifest_dir / "train_cuts_swbd.jsonl.gz" ) + @lru_cache() def dev_fisher_cuts(self) -> CutSet: logging.info("About to get dev fisher cuts") - return load_manifest_lazy(self.args.manifest_dir / "dev_cuts_fisher.jsonl.gz" + return load_manifest_lazy( + self.args.manifest_dir / "dev_cuts_fisher.jsonl.gz" ) + @lru_cache() def dev_swbd_cuts(self) -> CutSet: logging.info("About to get dev swbd cuts") - return load_manifest_lazy(self.args.manifest_dir / "dev_cuts_swbd.jsonl.gz" + return load_manifest_lazy( + self.args.manifest_dir / "dev_cuts_swbd.jsonl.gz" ) + @lru_cache() def test_eval2000_cuts(self) -> CutSet: logging.info("About to get test eval2000 cuts") - return load_manifest_lazy(self.args.manifest_dir / "cuts_eval2000.jsonl.gz" + return load_manifest_lazy( + self.args.manifest_dir / "cuts_eval2000.jsonl.gz" ) + @lru_cache() def test_swbd_cuts(self) -> CutSet: logging.info("About to get test eval2000 swbd cuts") - return load_manifest_lazy(self.args.manifest_dir / "cuts_eval2000_swbd.jsonl.gz" + return load_manifest_lazy( + self.args.manifest_dir / "cuts_eval2000_swbd.jsonl.gz" ) + @lru_cache() def test_callhome_cuts(self) -> CutSet: logging.info("About to get test eval2000 callhome cuts") - return load_manifest_lazy(self.args.manifest_dir / "cuts_eval2000_callhome.jsonl.gz" + return load_manifest_lazy( + self.args.manifest_dir / "cuts_eval2000_callhome.jsonl.gz" ) diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/beam_search.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/beam_search.py index ed6a6ea82..940da93ac 100644 --- a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/beam_search.py +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/beam_search.py @@ -550,9 +550,7 @@ def greedy_search( def greedy_search_batch( - model: Transducer, - encoder_out: torch.Tensor, - encoder_out_lens: torch.Tensor, + model: Transducer, encoder_out: torch.Tensor, encoder_out_lens: torch.Tensor ) -> List[List[int]]: """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1. Args: @@ -591,9 +589,7 @@ def greedy_search_batch( hyps = [[blank_id] * context_size for _ in range(N)] decoder_input = torch.tensor( - hyps, - device=device, - dtype=torch.int64, + hyps, device=device, dtype=torch.int64 ) # (N, context_size) decoder_out = model.decoder(decoder_input, need_pad=False) @@ -630,9 +626,7 @@ def greedy_search_batch( # update decoder output decoder_input = [h[-context_size:] for h in hyps[:batch_size]] decoder_input = torch.tensor( - decoder_input, - device=device, - dtype=torch.int64, + decoder_input, device=device, dtype=torch.int64 ) decoder_out = model.decoder(decoder_input, need_pad=False) decoder_out = model.joiner.decoder_proj(decoder_out) @@ -894,9 +888,7 @@ def modified_beam_search( ) # (num_hyps, 1, 1, encoder_out_dim) logits = model.joiner( - current_encoder_out, - decoder_out, - project_input=False, + current_encoder_out, decoder_out, project_input=False ) # (num_hyps, 1, 1, vocab_size) logits = logits.squeeze(1).squeeze(1) # (num_hyps, vocab_size) @@ -953,9 +945,7 @@ def modified_beam_search( def _deprecated_modified_beam_search( - model: Transducer, - encoder_out: torch.Tensor, - beam: int = 4, + model: Transducer, encoder_out: torch.Tensor, beam: int = 4 ) -> List[int]: """It limits the maximum number of symbols per frame to 1. @@ -1023,9 +1013,7 @@ def _deprecated_modified_beam_search( ) # (num_hyps, 1, 1, encoder_out_dim) logits = model.joiner( - current_encoder_out, - decoder_out, - project_input=False, + current_encoder_out, decoder_out, project_input=False ) # logits is of shape (num_hyps, 1, 1, vocab_size) logits = logits.squeeze(1).squeeze(1) @@ -1097,9 +1085,7 @@ def beam_search( device = next(model.parameters()).device decoder_input = torch.tensor( - [blank_id] * context_size, - device=device, - dtype=torch.int64, + [blank_id] * context_size, device=device, dtype=torch.int64 ).reshape(1, context_size) decoder_out = model.decoder(decoder_input, need_pad=False) @@ -1318,9 +1304,7 @@ def fast_beam_search_with_nbest_rescoring( num_unique_paths = len(word_ids_list) b_to_a_map = torch.zeros( - num_unique_paths, - dtype=torch.int32, - device=lattice.device, + num_unique_paths, dtype=torch.int32, device=lattice.device ) rescored_word_fsas = k2.intersect_device( @@ -1334,8 +1318,7 @@ def fast_beam_search_with_nbest_rescoring( rescored_word_fsas = k2.remove_epsilon_self_loops(rescored_word_fsas) rescored_word_fsas = k2.top_sort(k2.connect(rescored_word_fsas)) ngram_lm_scores = rescored_word_fsas.get_tot_scores( - use_double_scores=True, - log_semiring=False, + use_double_scores=True, log_semiring=False ) ans: Dict[str, List[List[int]]] = {} diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/conformer.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/conformer.py index fb8123838..f40a950c6 100644 --- a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/conformer.py +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/conformer.py @@ -223,19 +223,10 @@ class Conformer(EncoderInterface): init_states: List[torch.Tensor] = [ torch.zeros( - ( - self.encoder_layers, - left_context, - self.d_model, - ), - device=device, + (self.encoder_layers, left_context, self.d_model), device=device ), torch.zeros( - ( - self.encoder_layers, - self.cnn_module_kernel - 1, - self.d_model, - ), + (self.encoder_layers, self.cnn_module_kernel - 1, self.d_model), device=device, ), ] @@ -330,7 +321,9 @@ class Conformer(EncoderInterface): {(self.encoder_layers, self.cnn_module_kernel - 1, x.size(0), self.d_model)}, given {states[1].shape}.""" - lengths -= 2 # we will cut off 1 frame on each side of encoder_embed output + lengths -= ( + 2 + ) # we will cut off 1 frame on each side of encoder_embed output src_key_padding_mask = make_pad_mask(lengths) @@ -829,9 +822,7 @@ class RelPositionalEncoding(torch.nn.Module): self.pe = pe.to(device=x.device, dtype=x.dtype) def forward( - self, - x: torch.Tensor, - left_context: int = 0, + self, x: torch.Tensor, left_context: int = 0 ) -> Tuple[Tensor, Tensor]: """Add positional encoding. @@ -875,10 +866,7 @@ class RelPositionMultiheadAttention(nn.Module): """ def __init__( - self, - embed_dim: int, - num_heads: int, - dropout: float = 0.0, + self, embed_dim: int, num_heads: int, dropout: float = 0.0 ) -> None: super(RelPositionMultiheadAttention, self).__init__() self.embed_dim = embed_dim @@ -1272,8 +1260,7 @@ class RelPositionMultiheadAttention(nn.Module): bsz, num_heads, tgt_len, src_len ) attn_output_weights = attn_output_weights.masked_fill( - key_padding_mask.unsqueeze(1).unsqueeze(2), - float("-inf"), + key_padding_mask.unsqueeze(1).unsqueeze(2), float("-inf") ) attn_output_weights = attn_output_weights.view( bsz * num_heads, tgt_len, src_len @@ -1420,10 +1407,7 @@ class ConvolutionModule(nn.Module): ) def forward( - self, - x: Tensor, - cache: Optional[Tensor] = None, - right_context: int = 0, + self, x: Tensor, cache: Optional[Tensor] = None, right_context: int = 0 ) -> Tuple[Tensor, Tensor]: """Compute convolution module. diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decode.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decode.py index 0ebe20372..e5a4fc313 100755 --- a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decode.py +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decode.py @@ -384,9 +384,7 @@ def decode_one_batch( feature_lens += params.left_context feature = torch.nn.functional.pad( - feature, - pad=(0, 0, 0, params.left_context), - value=LOG_EPS, + feature, pad=(0, 0, 0, params.left_context), value=LOG_EPS ) if params.simulate_streaming: @@ -778,7 +776,7 @@ def main(): num_param = sum([p.numel() for p in model.parameters()]) logging.info(f"Number of model parameters: {num_param}") - + fisherswbd = FisherSwbdSpeechAsrDataModule(args) test_eval2000_cuts = fisherswbd.test_eval2000_cuts() @@ -803,9 +801,7 @@ def main(): ) save_results( - params=params, - test_set_name=test_set, - results_dict=results_dict, + params=params, test_set_name=test_set, results_dict=results_dict ) logging.info("Done!") diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decode_stream.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decode_stream.py index ba5e80555..2e90cdac4 100644 --- a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decode_stream.py +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/decode_stream.py @@ -92,10 +92,7 @@ class DecodeStream(object): """Return True if all the features are processed.""" return self._done - def set_features( - self, - features: torch.Tensor, - ) -> None: + def set_features(self, features: torch.Tensor) -> None: """Set features tensor of current utterance.""" assert features.dim() == 2, features.dim() self.features = torch.nn.functional.pad( diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/pretrained.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/pretrained.py index f52cb22ab..99cf5273d 100755 --- a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/pretrained.py +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/pretrained.py @@ -96,11 +96,7 @@ def get_parser(): "icefall.checkpoint.save_checkpoint().", ) - parser.add_argument( - "--bpe-model", - type=str, - help="""Path to bpe.model.""", - ) + parser.add_argument("--bpe-model", type=str, help="""Path to bpe.model.""") parser.add_argument( "--method", diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/streaming_decode.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/streaming_decode.py index 52fe34e88..76d50d28d 100755 --- a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/streaming_decode.py +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/streaming_decode.py @@ -39,7 +39,8 @@ import numpy as np import sentencepiece as spm import torch import torch.nn as nn -#from asr_datamodule import LibriSpeechAsrDataModule + +# from asr_datamodule import LibriSpeechAsrDataModule from asr_datamodule import FisherSwbdSpeechAsrDataModule from decode_stream import DecodeStream from kaldifeat import Fbank, FbankOptions @@ -187,9 +188,7 @@ def get_parser(): def greedy_search( - model: nn.Module, - encoder_out: torch.Tensor, - streams: List[DecodeStream], + model: nn.Module, encoder_out: torch.Tensor, streams: List[DecodeStream] ) -> List[List[int]]: assert len(streams) == encoder_out.size(0) @@ -236,10 +235,7 @@ def greedy_search( device=device, dtype=torch.int64, ) - decoder_out = model.decoder( - decoder_input, - need_pad=False, - ) + decoder_out = model.decoder(decoder_input, need_pad=False) decoder_out = model.joiner.decoder_proj(decoder_out) hyp_tokens = [] @@ -290,9 +286,7 @@ def fast_beam_search( def decode_one_chunk( - params: AttributeDict, - model: nn.Module, - decode_streams: List[DecodeStream], + params: AttributeDict, model: nn.Module, decode_streams: List[DecodeStream] ) -> List[int]: """Decode one chunk frames of features for each decode_streams and return the indexes of finished streams in a List. @@ -502,10 +496,7 @@ def decode_dataset( if params.decoding_method == "greedy_search": hyp = hyp[params.context_size :] # noqa decode_results.append( - ( - decode_streams[i].ground_truth.split(), - sp.decode(hyp).split(), - ) + (decode_streams[i].ground_truth.split(), sp.decode(hyp).split()) ) del decode_streams[i] @@ -661,7 +652,7 @@ def main(): fisherswbd = FisherSwbdSpeechAsrDataModule(args) test_eval2000_cuts = fisherswbd.test_eval2000_cuts() - test_swbd_cuts = fisherswbd.test_swbd_cuts () + test_swbd_cuts = fisherswbd.test_swbd_cuts() test_callhome_cuts = fisherswbd.test_callhome_cuts() test_eval2000_dl = fisherswbd.test_dataloaders(test_eval2000_cuts) @@ -681,9 +672,7 @@ def main(): ) save_results( - params=params, - test_set_name=test_set, - results_dict=results_dict, + params=params, test_set_name=test_set, results_dict=results_dict ) logging.info("Done!") diff --git a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/train.py b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/train.py index 354bd1d74..5b5af6720 100755 --- a/egs/fisher_swbd/ASR/pruned_transducer_stateless2/train.py +++ b/egs/fisher_swbd/ASR/pruned_transducer_stateless2/train.py @@ -155,10 +155,7 @@ def get_parser(): ) parser.add_argument( - "--num-epochs", - type=int, - default=30, - help="Number of epochs to train.", + "--num-epochs", type=int, default=30, help="Number of epochs to train." ) parser.add_argument( @@ -480,10 +477,7 @@ def load_checkpoint_if_available( assert filename.is_file(), f"{filename} does not exist!" saved_params = load_checkpoint( - filename, - model=model, - optimizer=optimizer, - scheduler=scheduler, + filename, model=model, optimizer=optimizer, scheduler=scheduler ) keys = [ @@ -646,11 +640,7 @@ def compute_validation_loss( for batch_idx, batch in enumerate(valid_dl): loss, loss_info = compute_loss( - params=params, - model=model, - sp=sp, - batch=batch, - is_training=False, + params=params, model=model, sp=sp, batch=batch, is_training=False ) assert loss.requires_grad is False tot_loss = tot_loss + loss_info @@ -767,9 +757,7 @@ def train_one_epoch( ) del params.cur_batch_idx remove_checkpoints( - out_dir=params.exp_dir, - topk=params.keep_last_k, - rank=rank, + out_dir=params.exp_dir, topk=params.keep_last_k, rank=rank ) if batch_idx % params.log_interval == 0: @@ -830,8 +818,6 @@ def run(rank, world_size, args): """ params = get_params() params.update(vars(args)) - if params.full_libri is False: - params.valid_interval = 1600 fix_random_seed(params.seed) if world_size > 1: @@ -897,11 +883,11 @@ def run(rank, world_size, args): if params.print_diagnostics: diagnostic = diagnostics.attach_diagnostics(model) - librispeech = FisherSwbdSpeechAsrDataModule(args) + fisherswbd = FisherSwbdSpeechAsrDataModule(args) train_cuts = fisherswbd.train_fisher_cuts() train_cuts += fisherswbd.train_swbd_cuts() - + def remove_short_and_long_utt(c: Cut): # Keep only utterances with duration between 1 second and 20 seconds # @@ -991,9 +977,7 @@ def run(rank, world_size, args): def display_and_save_batch( - batch: dict, - params: AttributeDict, - sp: spm.SentencePieceProcessor, + batch: dict, params: AttributeDict, sp: spm.SentencePieceProcessor ) -> None: """Display the batch statistics and save the batch into disk.