From 71e8d2d8e2c751d72f0810540b7e9ad0962ea617 Mon Sep 17 00:00:00 2001 From: jinzr <60612200+JinZr@users.noreply.github.com> Date: Sun, 15 Oct 2023 16:04:00 +0800 Subject: [PATCH 1/5] init commit --- egs/ami/ASR/local/prepare_ami_gss.sh | 4 +- egs/ami/ASR/zipformer/.gitignore | 1 + egs/ami/ASR/zipformer/asr_datamodule.py | 1 + egs/ami/ASR/zipformer/beam_search.py | 1 + egs/ami/ASR/zipformer/ctc_decode.py | 847 +++++++++++ egs/ami/ASR/zipformer/decode.py | 1052 +++++++++++++ egs/ami/ASR/zipformer/decode_stream.py | 148 ++ egs/ami/ASR/zipformer/decoder.py | 1 + egs/ami/ASR/zipformer/encoder_interface.py | 1 + .../ASR/zipformer/export-onnx-streaming.py | 775 ++++++++++ egs/ami/ASR/zipformer/export-onnx.py | 620 ++++++++ egs/ami/ASR/zipformer/export.py | 526 +++++++ .../ASR/zipformer/generate_averaged_model.py | 1 + egs/ami/ASR/zipformer/jit_pretrained.py | 280 ++++ egs/ami/ASR/zipformer/jit_pretrained_ctc.py | 436 ++++++ .../ASR/zipformer/jit_pretrained_streaming.py | 273 ++++ egs/ami/ASR/zipformer/joiner.py | 1 + egs/ami/ASR/zipformer/model.py | 1 + egs/ami/ASR/zipformer/onnx_check.py | 240 +++ egs/ami/ASR/zipformer/onnx_decode.py | 325 ++++ .../zipformer/onnx_pretrained-streaming.py | 546 +++++++ egs/ami/ASR/zipformer/onnx_pretrained.py | 421 ++++++ egs/ami/ASR/zipformer/optim.py | 1 + egs/ami/ASR/zipformer/pretrained.py | 381 +++++ egs/ami/ASR/zipformer/pretrained_ctc.py | 455 ++++++ egs/ami/ASR/zipformer/profile.py | 1 + egs/ami/ASR/zipformer/scaling.py | 1 + egs/ami/ASR/zipformer/scaling_converter.py | 1 + .../ASR/zipformer/streaming_beam_search.py | 1 + egs/ami/ASR/zipformer/streaming_decode.py | 853 +++++++++++ egs/ami/ASR/zipformer/subsampling.py | 1 + egs/ami/ASR/zipformer/test_scaling.py | 1 + egs/ami/ASR/zipformer/test_subsampling.py | 1 + egs/ami/ASR/zipformer/train.py | 1344 +++++++++++++++++ egs/ami/ASR/zipformer/zipformer.py | 1 + 35 files changed, 9541 insertions(+), 2 deletions(-) create mode 100644 egs/ami/ASR/zipformer/.gitignore create mode 120000 egs/ami/ASR/zipformer/asr_datamodule.py create mode 120000 egs/ami/ASR/zipformer/beam_search.py create mode 100755 egs/ami/ASR/zipformer/ctc_decode.py create mode 100755 egs/ami/ASR/zipformer/decode.py create mode 100644 egs/ami/ASR/zipformer/decode_stream.py create mode 120000 egs/ami/ASR/zipformer/decoder.py create mode 120000 egs/ami/ASR/zipformer/encoder_interface.py create mode 100755 egs/ami/ASR/zipformer/export-onnx-streaming.py create mode 100755 egs/ami/ASR/zipformer/export-onnx.py create mode 100755 egs/ami/ASR/zipformer/export.py create mode 120000 egs/ami/ASR/zipformer/generate_averaged_model.py create mode 100755 egs/ami/ASR/zipformer/jit_pretrained.py create mode 100755 egs/ami/ASR/zipformer/jit_pretrained_ctc.py create mode 100755 egs/ami/ASR/zipformer/jit_pretrained_streaming.py create mode 120000 egs/ami/ASR/zipformer/joiner.py create mode 120000 egs/ami/ASR/zipformer/model.py create mode 100755 egs/ami/ASR/zipformer/onnx_check.py create mode 100755 egs/ami/ASR/zipformer/onnx_decode.py create mode 100755 egs/ami/ASR/zipformer/onnx_pretrained-streaming.py create mode 100755 egs/ami/ASR/zipformer/onnx_pretrained.py create mode 120000 egs/ami/ASR/zipformer/optim.py create mode 100755 egs/ami/ASR/zipformer/pretrained.py create mode 100755 egs/ami/ASR/zipformer/pretrained_ctc.py create mode 120000 egs/ami/ASR/zipformer/profile.py create mode 120000 egs/ami/ASR/zipformer/scaling.py create mode 120000 egs/ami/ASR/zipformer/scaling_converter.py create mode 120000 egs/ami/ASR/zipformer/streaming_beam_search.py create mode 100755 egs/ami/ASR/zipformer/streaming_decode.py create mode 120000 egs/ami/ASR/zipformer/subsampling.py create mode 120000 egs/ami/ASR/zipformer/test_scaling.py create mode 120000 egs/ami/ASR/zipformer/test_subsampling.py create mode 100755 egs/ami/ASR/zipformer/train.py create mode 120000 egs/ami/ASR/zipformer/zipformer.py diff --git a/egs/ami/ASR/local/prepare_ami_gss.sh b/egs/ami/ASR/local/prepare_ami_gss.sh index d5422458b..414c22b12 100755 --- a/egs/ami/ASR/local/prepare_ami_gss.sh +++ b/egs/ami/ASR/local/prepare_ami_gss.sh @@ -58,7 +58,7 @@ if [ $stage -le 4 ]; then # for train, we use smaller context and larger batches to speed-up processing for JOB in $(seq $nj); do gss enhance cuts $EXP_DIR/cuts_train.jsonl.gz \ - $EXP_DIR/cuts_per_segment_train_split$nj/cuts_per_segment_train.JOB.jsonl.gz $EXP_DIR/enhanced \ + $EXP_DIR/cuts_per_segment_train_split$nj/cuts_per_segment_train.$JOB.jsonl.gz $EXP_DIR/enhanced \ --bss-iterations 10 \ --context-duration 5.0 \ --use-garbage-class \ @@ -77,7 +77,7 @@ if [ $stage -le 5 ]; then for part in dev test; do for JOB in $(seq $nj); do gss enhance cuts $EXP_DIR/cuts_${part}.jsonl.gz \ - $EXP_DIR/cuts_per_segment_${part}_split$nj/cuts_per_segment_${part}.JOB.jsonl.gz \ + $EXP_DIR/cuts_per_segment_${part}_split$nj/cuts_per_segment_${part}.$JOB.jsonl.gz \ $EXP_DIR/enhanced \ --bss-iterations 10 \ --context-duration 15.0 \ diff --git a/egs/ami/ASR/zipformer/.gitignore b/egs/ami/ASR/zipformer/.gitignore new file mode 100644 index 000000000..e47ac1582 --- /dev/null +++ b/egs/ami/ASR/zipformer/.gitignore @@ -0,0 +1 @@ +swoosh.pdf diff --git a/egs/ami/ASR/zipformer/asr_datamodule.py b/egs/ami/ASR/zipformer/asr_datamodule.py new file mode 120000 index 000000000..c274de28a --- /dev/null +++ b/egs/ami/ASR/zipformer/asr_datamodule.py @@ -0,0 +1 @@ +../pruned_transducer_stateless7/asr_datamodule.py \ No newline at end of file diff --git a/egs/ami/ASR/zipformer/beam_search.py b/egs/ami/ASR/zipformer/beam_search.py new file mode 120000 index 000000000..e24eca39f --- /dev/null +++ b/egs/ami/ASR/zipformer/beam_search.py @@ -0,0 +1 @@ +../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py \ No newline at end of file diff --git a/egs/ami/ASR/zipformer/ctc_decode.py b/egs/ami/ASR/zipformer/ctc_decode.py new file mode 100755 index 000000000..4db50b981 --- /dev/null +++ b/egs/ami/ASR/zipformer/ctc_decode.py @@ -0,0 +1,847 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Liyong Guo, +# Quandong Wang, +# Zengwei Yao) +# +# 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) ctc-decoding +./zipformer/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --use-ctc 1 \ + --max-duration 600 \ + --decoding-method ctc-decoding + +(2) 1best +./zipformer/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --use-ctc 1 \ + --max-duration 600 \ + --hlg-scale 0.6 \ + --decoding-method 1best + +(3) nbest +./zipformer/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --use-ctc 1 \ + --max-duration 600 \ + --hlg-scale 0.6 \ + --decoding-method nbest + +(4) nbest-rescoring +./zipformer/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --use-ctc 1 \ + --max-duration 600 \ + --hlg-scale 0.6 \ + --nbest-scale 1.0 \ + --lm-dir data/lm \ + --decoding-method nbest-rescoring + +(5) whole-lattice-rescoring +./zipformer/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --use-ctc 1 \ + --max-duration 600 \ + --hlg-scale 0.6 \ + --nbest-scale 1.0 \ + --lm-dir data/lm \ + --decoding-method whole-lattice-rescoring +""" + + +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 LibriSpeechAsrDataModule +from train import add_model_arguments, get_params, get_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.decode import ( + get_lattice, + nbest_decoding, + nbest_oracle, + one_best_decoding, + rescore_with_n_best_list, + rescore_with_whole_lattice, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + get_texts, + 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=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + 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( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/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( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="ctc-decoding", + help="""Decoding method. + Supported values are: + - (1) ctc-decoding. Use CTC decoding. It uses a sentence piece + model, i.e., lang_dir/bpe.model, to convert word pieces to words. + It needs neither a lexicon nor an n-gram LM. + - (2) 1best. Extract the best path from the decoding lattice as the + decoding result. + - (3) nbest. Extract n paths from the decoding lattice; the path + with the highest score is the decoding result. + - (4) nbest-rescoring. Extract n paths from the decoding lattice, + rescore them with an n-gram LM (e.g., a 4-gram LM), the path with + the highest score is the decoding result. + - (5) whole-lattice-rescoring. Rescore the decoding lattice with an + n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice + is the decoding result. + you have trained an RNN LM using ./rnn_lm/train.py + - (6) nbest-oracle. Its WER is the lower bound of any n-best + rescoring method can achieve. Useful for debugging n-best + rescoring method. + """, + ) + + parser.add_argument( + "--num-paths", + type=int, + default=100, + help="""Number of paths for n-best based decoding method. + Used only when "method" is one of the following values: + nbest, nbest-rescoring, and nbest-oracle + """, + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=1.0, + help="""The scale to be applied to `lattice.scores`. + It's needed if you use any kinds of n-best based rescoring. + Used only when "method" is one of the following values: + nbest, nbest-rescoring, and nbest-oracle + A smaller value results in more unique paths. + """, + ) + + parser.add_argument( + "--hlg-scale", + type=float, + default=0.6, + help="""The scale to be applied to `hlg.scores`. + """, + ) + + parser.add_argument( + "--lm-dir", + type=str, + default="data/lm", + help="""The n-gram LM dir. + It should contain either G_4_gram.pt or G_4_gram.fst.txt + """, + ) + + add_model_arguments(parser) + + return parser + + +def get_decoding_params() -> AttributeDict: + """Parameters for decoding.""" + params = AttributeDict( + { + "frame_shift_ms": 10, + "search_beam": 20, + "output_beam": 8, + "min_active_states": 30, + "max_active_states": 10000, + "use_double_scores": True, + } + ) + return params + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + HLG: Optional[k2.Fsa], + H: Optional[k2.Fsa], + bpe_model: Optional[spm.SentencePieceProcessor], + batch: dict, + word_table: k2.SymbolTable, + G: 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 no rescoring is used, the key is the string `no_rescore`. + If LM rescoring is used, the key is the string `lm_scale_xxx`, + where `xxx` is the value of `lm_scale`. An example key is + `lm_scale_0.7` + - 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`. + + - params.decoding_method is "1best", it uses 1best decoding without LM rescoring. + - params.decoding_method is "nbest", it uses nbest decoding without LM rescoring. + - params.decoding_method is "nbest-rescoring", it uses nbest LM rescoring. + - params.decoding_method is "whole-lattice-rescoring", it uses whole lattice LM + rescoring. + + model: + The neural model. + HLG: + The decoding graph. Used only when params.decoding_method is NOT ctc-decoding. + H: + The ctc topo. Used only when params.decoding_method is ctc-decoding. + bpe_model: + The BPE model. Used only when params.decoding_method is ctc-decoding. + 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. + G: + An LM. It is not None when params.decoding_method is "nbest-rescoring" + or "whole-lattice-rescoring". In general, the G in HLG + is a 3-gram LM, while this G is a 4-gram LM. + Returns: + Return the decoding result. See above description for the format of + the returned dict. Note: If it decodes to nothing, then return None. + """ + if HLG is not None: + device = HLG.device + else: + device = H.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) + + if params.causal: + # this seems to cause insertions at the end of the utterance if used with zipformer. + pad_len = 30 + feature_lens += pad_len + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, pad_len), + value=LOG_EPS, + ) + + encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens) + ctc_output = model.ctc_output(encoder_out) # (N, T, C) + + supervision_segments = torch.stack( + ( + supervisions["sequence_idx"], + torch.div( + supervisions["start_frame"], + params.subsampling_factor, + rounding_mode="floor", + ), + torch.div( + supervisions["num_frames"], + params.subsampling_factor, + rounding_mode="floor", + ), + ), + 1, + ).to(torch.int32) + + if H is None: + assert HLG is not None + decoding_graph = HLG + else: + assert HLG is None + assert bpe_model is not None + decoding_graph = H + + lattice = get_lattice( + nnet_output=ctc_output, + decoding_graph=decoding_graph, + supervision_segments=supervision_segments, + search_beam=params.search_beam, + output_beam=params.output_beam, + min_active_states=params.min_active_states, + max_active_states=params.max_active_states, + subsampling_factor=params.subsampling_factor, + ) + + if params.decoding_method == "ctc-decoding": + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + # Note: `best_path.aux_labels` contains token IDs, not word IDs + # since we are using H, not HLG here. + # + # token_ids is a lit-of-list of IDs + token_ids = get_texts(best_path) + + # hyps is a list of str, e.g., ['xxx yyy zzz', ...] + hyps = bpe_model.decode(token_ids) + + # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ] + hyps = [s.split() for s in hyps] + key = "ctc-decoding" + return {key: hyps} + + if params.decoding_method == "nbest-oracle": + # Note: You can also pass rescored lattices to it. + # We choose the HLG decoded lattice for speed reasons + # as HLG decoding is faster and the oracle WER + # is only slightly worse than that of rescored lattices. + best_path = nbest_oracle( + lattice=lattice, + num_paths=params.num_paths, + ref_texts=supervisions["text"], + word_table=word_table, + nbest_scale=params.nbest_scale, + oov="", + ) + hyps = get_texts(best_path) + hyps = [[word_table[i] for i in ids] for ids in hyps] + key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}" # noqa + return {key: hyps} + + if params.decoding_method in ["1best", "nbest"]: + if params.decoding_method == "1best": + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + key = "no_rescore" + else: + best_path = nbest_decoding( + lattice=lattice, + num_paths=params.num_paths, + use_double_scores=params.use_double_scores, + nbest_scale=params.nbest_scale, + ) + key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}" # noqa + + hyps = get_texts(best_path) + hyps = [[word_table[i] for i in ids] for ids in hyps] + return {key: hyps} + + assert params.decoding_method in [ + "nbest-rescoring", + "whole-lattice-rescoring", + ] + + lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7] + lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3] + lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0] + + if params.decoding_method == "nbest-rescoring": + best_path_dict = rescore_with_n_best_list( + lattice=lattice, + G=G, + num_paths=params.num_paths, + lm_scale_list=lm_scale_list, + nbest_scale=params.nbest_scale, + ) + elif params.decoding_method == "whole-lattice-rescoring": + best_path_dict = rescore_with_whole_lattice( + lattice=lattice, + G_with_epsilon_loops=G, + lm_scale_list=lm_scale_list, + ) + else: + assert False, f"Unsupported decoding method: {params.decoding_method}" + + ans = dict() + if best_path_dict is not None: + for lm_scale_str, best_path in best_path_dict.items(): + hyps = get_texts(best_path) + hyps = [[word_table[i] for i in ids] for ids in hyps] + ans[lm_scale_str] = hyps + else: + ans = None + return ans + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + HLG: Optional[k2.Fsa], + H: Optional[k2.Fsa], + bpe_model: Optional[spm.SentencePieceProcessor], + word_table: k2.SymbolTable, + G: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[str, 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. + HLG: + The decoding graph. Used only when params.decoding_method is NOT ctc-decoding. + H: + The ctc topo. Used only when params.decoding_method is ctc-decoding. + bpe_model: + The BPE model. Used only when params.decoding_method is ctc-decoding. + word_table: + It is the word symbol table. + G: + An LM. It is not None when params.decoding_method is "nbest-rescoring" + or "whole-lattice-rescoring". In general, the G in HLG + is a 3-gram LM, while this G is a 4-gram LM. + Returns: + Return a dict, whose key may be "no-rescore" if no LM rescoring + is used, or it may be "lm_scale_0.7" if LM rescoring 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 = "?" + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + HLG=HLG, + H=H, + bpe_model=bpe_model, + batch=batch, + word_table=word_table, + G=G, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % 100 == 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[str, 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}-{params.suffix}.txt" + 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}-{params.suffix}.txt" + with open(errs_filename, "w") as f: + wer = write_error_stats(f, f"{test_set_name}-{key}", results) + 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}-{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() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + args.lang_dir = Path(args.lang_dir) + args.lm_dir = Path(args.lm_dir) + + params = get_params() + # add decoding params + params.update(get_decoding_params()) + params.update(vars(args)) + + assert params.decoding_method in ( + "ctc-decoding", + "1best", + "nbest", + "nbest-rescoring", + "whole-lattice-rescoring", + "nbest-oracle", + ) + 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.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + 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}") + logging.info(params) + + lexicon = Lexicon(params.lang_dir) + max_token_id = max(lexicon.tokens) + num_classes = max_token_id + 1 # +1 for the blank + + params.vocab_size = num_classes + # and are defined in local/train_bpe_model.py + params.blank_id = 0 + + if params.decoding_method == "ctc-decoding": + HLG = None + H = k2.ctc_topo( + max_token=max_token_id, + modified=False, + device=device, + ) + bpe_model = spm.SentencePieceProcessor() + bpe_model.load(str(params.lang_dir / "bpe.model")) + else: + H = None + bpe_model = None + HLG = k2.Fsa.from_dict( + torch.load(f"{params.lang_dir}/HLG.pt", map_location=device) + ) + assert HLG.requires_grad is False + + HLG.scores *= params.hlg_scale + if not hasattr(HLG, "lm_scores"): + HLG.lm_scores = HLG.scores.clone() + + if params.decoding_method in ( + "nbest-rescoring", + "whole-lattice-rescoring", + ): + if not (params.lm_dir / "G_4_gram.pt").is_file(): + logging.info("Loading G_4_gram.fst.txt") + logging.warning("It may take 8 minutes.") + with open(params.lm_dir / "G_4_gram.fst.txt") as f: + first_word_disambig_id = lexicon.word_table["#0"] + + G = k2.Fsa.from_openfst(f.read(), acceptor=False) + # G.aux_labels is not needed in later computations, so + # remove it here. + del G.aux_labels + # CAUTION: The following line is crucial. + # Arcs entering the back-off state have label equal to #0. + # We have to change it to 0 here. + G.labels[G.labels >= first_word_disambig_id] = 0 + # See https://github.com/k2-fsa/k2/issues/874 + # for why we need to set G.properties to None + G.__dict__["_properties"] = None + G = k2.Fsa.from_fsas([G]).to(device) + G = k2.arc_sort(G) + # Save a dummy value so that it can be loaded in C++. + # See https://github.com/pytorch/pytorch/issues/67902 + # for why we need to do this. + G.dummy = 1 + + torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt") + else: + logging.info("Loading pre-compiled G_4_gram.pt") + d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device) + G = k2.Fsa.from_dict(d) + + if params.decoding_method == "whole-lattice-rescoring": + # Add epsilon self-loops to G as we will compose + # it with the whole lattice later + G = k2.add_epsilon_self_loops(G) + G = k2.arc_sort(G) + G = G.to(device) + + # G.lm_scores is used to replace HLG.lm_scores during + # LM rescoring. + G.lm_scores = G.scores.clone() + else: + G = None + + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + 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 i >= 1: + 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)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_other_dl = librispeech.test_dataloaders(test_other_cuts) + + test_sets = ["test-clean", "test-other"] + test_dl = [test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + HLG=HLG, + H=H, + bpe_model=bpe_model, + word_table=lexicon.word_table, + G=G, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/ami/ASR/zipformer/decode.py b/egs/ami/ASR/zipformer/decode.py new file mode 100755 index 000000000..3531d657f --- /dev/null +++ b/egs/ami/ASR/zipformer/decode.py @@ -0,0 +1,1052 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# 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 +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search (one best) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(5) fast beam search (nbest) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/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) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/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) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_LG \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 +""" + + +import argparse +import logging +import math +import os +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 LibriSpeechAsrDataModule +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, + modified_beam_search_lm_rescore, + modified_beam_search_lm_rescore_LODR, + modified_beam_search_lm_shallow_fusion, + modified_beam_search_LODR, +) +from train import add_model_arguments, get_model, get_params + +from icefall import ContextGraph, LmScorer, NgramLm +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + make_pad_mask, + 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=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + 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( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/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 + - modified_beam_search_LODR + - 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( + "--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""", + ) + + parser.add_argument( + "--use-shallow-fusion", + type=str2bool, + default=False, + help="""Use neural network LM for shallow fusion. + If you want to use LODR, you will also need to set this to true + """, + ) + + parser.add_argument( + "--lm-type", + type=str, + default="rnn", + help="Type of NN lm", + choices=["rnn", "transformer"], + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.3, + help="""The scale of the neural network LM + Used only when `--use-shallow-fusion` is set to True. + """, + ) + + parser.add_argument( + "--tokens-ngram", + type=int, + default=2, + help="""The order of the ngram lm. + """, + ) + + parser.add_argument( + "--backoff-id", + type=int, + default=500, + help="ID of the backoff symbol in the ngram LM", + ) + + parser.add_argument( + "--context-score", + type=float, + default=2, + help=""" + The bonus score of each token for the context biasing words/phrases. + Used only when --decoding-method is modified_beam_search and + modified_beam_search_LODR. + """, + ) + + parser.add_argument( + "--context-file", + type=str, + default="", + help=""" + The path of the context biasing lists, one word/phrase each line + Used only when --decoding-method is modified_beam_search and + modified_beam_search_LODR. + """, + ) + 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, + context_graph: Optional[ContextGraph] = None, + LM: Optional[LmScorer] = None, + ngram_lm=None, + ngram_lm_scale: float = 0.0, +) -> 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. + LM: + A neural network language model. + ngram_lm: + A ngram language model + ngram_lm_scale: + The scale for the ngram language model. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).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) + + if params.causal: + # this seems to cause insertions at the end of the utterance if used with zipformer. + pad_len = 30 + feature_lens += pad_len + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, pad_len), + value=LOG_EPS, + ) + + encoder_out, encoder_out_lens = model.forward_encoder(feature, 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, + context_graph=context_graph, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_lm_shallow_fusion": + hyp_tokens = modified_beam_search_lm_shallow_fusion( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_LODR": + hyp_tokens = modified_beam_search_LODR( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LODR_lm=ngram_lm, + LODR_lm_scale=ngram_lm_scale, + LM=LM, + context_graph=context_graph, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_lm_rescore": + lm_scale_list = [0.01 * i for i in range(10, 50)] + ans_dict = modified_beam_search_lm_rescore( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + lm_scale_list=lm_scale_list, + ) + elif params.decoding_method == "modified_beam_search_lm_rescore_LODR": + lm_scale_list = [0.02 * i for i in range(2, 30)] + ans_dict = modified_beam_search_lm_rescore_LODR( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + LODR_lm=ngram_lm, + sp=sp, + lm_scale_list=lm_scale_list, + ) + 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 "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} + elif "modified_beam_search" in params.decoding_method: + prefix = f"beam_size_{params.beam_size}" + if params.decoding_method in ( + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + ): + ans = dict() + assert ans_dict is not None + for key, hyps in ans_dict.items(): + hyps = [sp.decode(hyp).split() for hyp in hyps] + ans[f"{prefix}_{key}"] = hyps + return ans + else: + if params.has_contexts: + prefix += f"-context-score-{params.context_score}" + return {prefix: 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, + context_graph: Optional[ContextGraph] = None, + LM: Optional[LmScorer] = None, + ngram_lm=None, + ngram_lm_scale: float = 0.0, +) -> Dict[str, List[Tuple[str, 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"] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + context_graph=context_graph, + word_table=word_table, + batch=batch, + LM=LM, + ngram_lm=ngram_lm, + ngram_lm_scale=ngram_lm_scale, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = ref_text.split() + this_batch.append((cut_id, 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[str, 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" + ) + 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() + LibriSpeechAsrDataModule.add_arguments(parser) + LmScorer.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", + "modified_beam_search_LODR", + "modified_beam_search_lm_shallow_fusion", + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if os.path.exists(params.context_file): + params.has_contexts = True + else: + params.has_contexts = False + + 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.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + 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}" + if params.decoding_method in ( + "modified_beam_search", + "modified_beam_search_LODR", + ): + if params.has_contexts: + params.suffix += f"-context-score-{params.context_score}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_shallow_fusion: + params.suffix += f"-{params.lm_type}-lm-scale-{params.lm_scale}" + + if "LODR" in params.decoding_method: + params.suffix += ( + f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}" + ) + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + 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() + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + 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 i >= 1: + 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)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + # only load the neural network LM if required + if params.use_shallow_fusion or params.decoding_method in ( + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + "modified_beam_search_lm_shallow_fusion", + "modified_beam_search_LODR", + ): + LM = LmScorer( + lm_type=params.lm_type, + params=params, + device=device, + lm_scale=params.lm_scale, + ) + LM.to(device) + LM.eval() + else: + LM = None + + # only load N-gram LM when needed + if params.decoding_method == "modified_beam_search_lm_rescore_LODR": + try: + import kenlm + except ImportError: + print("Please install kenlm first. You can use") + print(" pip install https://github.com/kpu/kenlm/archive/master.zip") + print("to install it") + import sys + + sys.exit(-1) + ngram_file_name = str(params.lang_dir / f"{params.tokens_ngram}gram.arpa") + logging.info(f"lm filename: {ngram_file_name}") + ngram_lm = kenlm.Model(ngram_file_name) + ngram_lm_scale = None # use a list to search + + elif params.decoding_method == "modified_beam_search_LODR": + lm_filename = f"{params.tokens_ngram}gram.fst.txt" + logging.info(f"Loading token level lm: {lm_filename}") + ngram_lm = NgramLm( + str(params.lang_dir / lm_filename), + backoff_id=params.backoff_id, + is_binary=False, + ) + logging.info(f"num states: {ngram_lm.lm.num_states}") + ngram_lm_scale = params.ngram_lm_scale + else: + ngram_lm = None + ngram_lm_scale = None + + 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 + + if "modified_beam_search" in params.decoding_method: + if os.path.exists(params.context_file): + contexts = [] + for line in open(params.context_file).readlines(): + contexts.append(line.strip()) + context_graph = ContextGraph(params.context_score) + context_graph.build(sp.encode(contexts)) + else: + context_graph = None + else: + context_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_other_dl = librispeech.test_dataloaders(test_other_cuts) + + test_sets = ["test-clean", "test-other"] + test_dl = [test_clean_dl, test_other_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, + context_graph=context_graph, + LM=LM, + ngram_lm=ngram_lm, + ngram_lm_scale=ngram_lm_scale, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/ami/ASR/zipformer/decode_stream.py b/egs/ami/ASR/zipformer/decode_stream.py new file mode 100644 index 000000000..d6918bf32 --- /dev/null +++ b/egs/ami/ASR/zipformer/decode_stream.py @@ -0,0 +1,148 @@ +# Copyright 2022 Xiaomi Corp. (authors: Wei Kang, +# Zengwei Yao) +# +# 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 beam_search import Hypothesis, HypothesisList + +from icefall.utils import AttributeDict + + +class DecodeStream(object): + def __init__( + self, + params: AttributeDict, + cut_id: str, + 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 params.decoding_method == "fast_beam_search": + assert decoding_graph is not None + assert device == decoding_graph.device + + self.params = params + self.cut_id = cut_id + 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, at encoder output + self.done_frames: int = 0 + + # The encoder_embed subsample features (T - 7) // 2 + # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling + self.pad_length = 7 + 2 * 3 + + if params.decoding_method == "greedy_search": + self.hyp = [-1] * (params.context_size - 1) + [params.blank_id] + elif params.decoding_method == "modified_beam_search": + self.hyps = HypothesisList() + self.hyps.add( + Hypothesis( + ys=[-1] * (params.context_size - 1) + [params.blank_id], + log_prob=torch.zeros(1, dtype=torch.float32, device=device), + ) + ) + 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: + raise ValueError(f"Unsupported decoding method: {params.decoding_method}") + + @property + def done(self) -> bool: + """Return True if all the features are processed.""" + return self._done + + @property + def id(self) -> str: + return self.cut_id + + def set_features( + self, + features: torch.Tensor, + tail_pad_len: int = 0, + ) -> 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 + tail_pad_len), + 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 + ret_length # noqa + ] + + self.num_processed_frames += chunk_size + if self.num_processed_frames >= self.num_frames: + self._done = True + + return ret_features, ret_length + + def decoding_result(self) -> List[int]: + """Obtain current decoding result.""" + if self.params.decoding_method == "greedy_search": + return self.hyp[self.params.context_size :] # noqa + elif self.params.decoding_method == "modified_beam_search": + best_hyp = self.hyps.get_most_probable(length_norm=True) + return best_hyp.ys[self.params.context_size :] # noqa + else: + assert self.params.decoding_method == "fast_beam_search" + return self.hyp diff --git a/egs/ami/ASR/zipformer/decoder.py b/egs/ami/ASR/zipformer/decoder.py new file mode 120000 index 000000000..5a8018680 --- /dev/null +++ b/egs/ami/ASR/zipformer/decoder.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/decoder.py \ No newline at end of file diff --git a/egs/ami/ASR/zipformer/encoder_interface.py b/egs/ami/ASR/zipformer/encoder_interface.py new file mode 120000 index 000000000..653c5b09a --- /dev/null +++ b/egs/ami/ASR/zipformer/encoder_interface.py @@ -0,0 +1 @@ +../../../librispeech/ASR/transducer_stateless/encoder_interface.py \ No newline at end of file diff --git a/egs/ami/ASR/zipformer/export-onnx-streaming.py b/egs/ami/ASR/zipformer/export-onnx-streaming.py new file mode 100755 index 000000000..e2c7d7d95 --- /dev/null +++ b/egs/ami/ASR/zipformer/export-onnx-streaming.py @@ -0,0 +1,775 @@ +#!/usr/bin/env python3 +# +# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang, Wei Kang) +# Copyright 2023 Danqing Fu (danqing.fu@gmail.com) + +""" +This script exports a transducer model from PyTorch to ONNX. + +We use the pre-trained model from +https://huggingface.co/Zengwei/icefall-asr-librispeech-streaming-zipformer-2023-05-17 +as an example to show how to use this file. + +1. Download the pre-trained model + +cd egs/librispeech/ASR + +repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-streaming-zipformer-2023-05-17 +GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url +repo=$(basename $repo_url) + +pushd $repo +git lfs pull --include "exp/pretrained.pt" + +cd exp +ln -s pretrained.pt epoch-99.pt +popd + +2. Export the model to ONNX + +./zipformer/export-onnx-streaming.py \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ + --use-averaged-model 0 \ + --epoch 99 \ + --avg 1 \ + --exp-dir $repo/exp \ + --num-encoder-layers "2,2,3,4,3,2" \ + --downsampling-factor "1,2,4,8,4,2" \ + --feedforward-dim "512,768,1024,1536,1024,768" \ + --num-heads "4,4,4,8,4,4" \ + --encoder-dim "192,256,384,512,384,256" \ + --query-head-dim 32 \ + --value-head-dim 12 \ + --pos-head-dim 4 \ + --pos-dim 48 \ + --encoder-unmasked-dim "192,192,256,256,256,192" \ + --cnn-module-kernel "31,31,15,15,15,31" \ + --decoder-dim 512 \ + --joiner-dim 512 \ + --causal True \ + --chunk-size 16 \ + --left-context-frames 64 + +The --chunk-size in training is "16,32,64,-1", so we select one of them +(excluding -1) during streaming export. The same applies to `--left-context`, +whose value is "64,128,256,-1". + +It will generate the following 3 files inside $repo/exp: + + - encoder-epoch-99-avg-1-chunk-16-left-64.onnx + - decoder-epoch-99-avg-1-chunk-16-left-64.onnx + - joiner-epoch-99-avg-1-chunk-16-left-64.onnx + +See ./onnx_pretrained-streaming.py for how to use the exported ONNX models. +""" + +import argparse +import logging +from pathlib import Path +from typing import Dict, List, Tuple + +import k2 +import onnx +import torch +import torch.nn as nn +from decoder import Decoder +from onnxruntime.quantization import QuantType, quantize_dynamic +from scaling_converter import convert_scaled_to_non_scaled +from train import add_model_arguments, get_model, get_params +from zipformer import Zipformer2 + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import num_tokens, 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( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="""It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--tokens", + type=str, + default="data/lang_bpe_500/tokens.txt", + help="Path to the tokens.txt", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + add_model_arguments(parser) + + return parser + + +def add_meta_data(filename: str, meta_data: Dict[str, str]): + """Add meta data to an ONNX model. It is changed in-place. + + Args: + filename: + Filename of the ONNX model to be changed. + meta_data: + Key-value pairs. + """ + model = onnx.load(filename) + for key, value in meta_data.items(): + meta = model.metadata_props.add() + meta.key = key + meta.value = value + + onnx.save(model, filename) + + +class OnnxEncoder(nn.Module): + """A wrapper for Zipformer and the encoder_proj from the joiner""" + + def __init__( + self, encoder: Zipformer2, encoder_embed: nn.Module, encoder_proj: nn.Linear + ): + """ + Args: + encoder: + A Zipformer encoder. + encoder_proj: + The projection layer for encoder from the joiner. + """ + super().__init__() + self.encoder = encoder + self.encoder_embed = encoder_embed + self.encoder_proj = encoder_proj + self.chunk_size = encoder.chunk_size[0] + self.left_context_len = encoder.left_context_frames[0] + self.pad_length = 7 + 2 * 3 + + def forward( + self, + x: torch.Tensor, + states: List[torch.Tensor], + ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]: + N = x.size(0) + T = self.chunk_size * 2 + self.pad_length + x_lens = torch.tensor([T] * N, device=x.device) + left_context_len = self.left_context_len + + cached_embed_left_pad = states[-2] + x, x_lens, new_cached_embed_left_pad = self.encoder_embed.streaming_forward( + x=x, + x_lens=x_lens, + cached_left_pad=cached_embed_left_pad, + ) + assert x.size(1) == self.chunk_size, (x.size(1), self.chunk_size) + + src_key_padding_mask = torch.zeros(N, self.chunk_size, dtype=torch.bool) + + # processed_mask is used to mask out initial states + processed_mask = torch.arange(left_context_len, device=x.device).expand( + x.size(0), left_context_len + ) + processed_lens = states[-1] # (batch,) + # (batch, left_context_size) + processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1) + # Update processed lengths + new_processed_lens = processed_lens + x_lens + # (batch, left_context_size + chunk_size) + src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1) + + x = x.permute(1, 0, 2) + encoder_states = states[:-2] + logging.info(f"len_encoder_states={len(encoder_states)}") + ( + encoder_out, + encoder_out_lens, + new_encoder_states, + ) = self.encoder.streaming_forward( + x=x, + x_lens=x_lens, + states=encoder_states, + src_key_padding_mask=src_key_padding_mask, + ) + encoder_out = encoder_out.permute(1, 0, 2) + encoder_out = self.encoder_proj(encoder_out) + # Now encoder_out is of shape (N, T, joiner_dim) + + new_states = new_encoder_states + [ + new_cached_embed_left_pad, + new_processed_lens, + ] + + return encoder_out, new_states + + def get_init_states( + self, + batch_size: int = 1, + device: torch.device = torch.device("cpu"), + ) -> List[torch.Tensor]: + """ + Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6] + is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). + states[-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + states[-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + """ + states = self.encoder.get_init_states(batch_size, device) + + embed_states = self.encoder_embed.get_init_states(batch_size, device) + + states.append(embed_states) + + processed_lens = torch.zeros(batch_size, dtype=torch.int64, device=device) + states.append(processed_lens) + + return states + + +class OnnxDecoder(nn.Module): + """A wrapper for Decoder and the decoder_proj from the joiner""" + + def __init__(self, decoder: Decoder, decoder_proj: nn.Linear): + super().__init__() + self.decoder = decoder + self.decoder_proj = decoder_proj + + def forward(self, y: torch.Tensor) -> torch.Tensor: + """ + Args: + y: + A 2-D tensor of shape (N, context_size). + Returns + Return a 2-D tensor of shape (N, joiner_dim) + """ + need_pad = False + decoder_output = self.decoder(y, need_pad=need_pad) + decoder_output = decoder_output.squeeze(1) + output = self.decoder_proj(decoder_output) + + return output + + +class OnnxJoiner(nn.Module): + """A wrapper for the joiner""" + + def __init__(self, output_linear: nn.Linear): + super().__init__() + self.output_linear = output_linear + + def forward( + self, + encoder_out: torch.Tensor, + decoder_out: torch.Tensor, + ) -> torch.Tensor: + """ + Args: + encoder_out: + A 2-D tensor of shape (N, joiner_dim) + decoder_out: + A 2-D tensor of shape (N, joiner_dim) + Returns: + Return a 2-D tensor of shape (N, vocab_size) + """ + logit = encoder_out + decoder_out + logit = self.output_linear(torch.tanh(logit)) + return logit + + +def export_encoder_model_onnx( + encoder_model: OnnxEncoder, + encoder_filename: str, + opset_version: int = 11, +) -> None: + encoder_model.encoder.__class__.forward = ( + encoder_model.encoder.__class__.streaming_forward + ) + + decode_chunk_len = encoder_model.chunk_size * 2 + # The encoder_embed subsample features (T - 7) // 2 + # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling + T = decode_chunk_len + encoder_model.pad_length + + x = torch.rand(1, T, 80, dtype=torch.float32) + init_state = encoder_model.get_init_states() + num_encoders = len(encoder_model.encoder.encoder_dim) + logging.info(f"num_encoders: {num_encoders}") + logging.info(f"len(init_state): {len(init_state)}") + + inputs = {} + input_names = ["x"] + + outputs = {} + output_names = ["encoder_out"] + + def build_inputs_outputs(tensors, i): + assert len(tensors) == 6, len(tensors) + + # (downsample_left, batch_size, key_dim) + name = f"cached_key_{i}" + logging.info(f"{name}.shape: {tensors[0].shape}") + inputs[name] = {1: "N"} + outputs[f"new_{name}"] = {1: "N"} + input_names.append(name) + output_names.append(f"new_{name}") + + # (1, batch_size, downsample_left, nonlin_attn_head_dim) + name = f"cached_nonlin_attn_{i}" + logging.info(f"{name}.shape: {tensors[1].shape}") + inputs[name] = {1: "N"} + outputs[f"new_{name}"] = {1: "N"} + input_names.append(name) + output_names.append(f"new_{name}") + + # (downsample_left, batch_size, value_dim) + name = f"cached_val1_{i}" + logging.info(f"{name}.shape: {tensors[2].shape}") + inputs[name] = {1: "N"} + outputs[f"new_{name}"] = {1: "N"} + input_names.append(name) + output_names.append(f"new_{name}") + + # (downsample_left, batch_size, value_dim) + name = f"cached_val2_{i}" + logging.info(f"{name}.shape: {tensors[3].shape}") + inputs[name] = {1: "N"} + outputs[f"new_{name}"] = {1: "N"} + input_names.append(name) + output_names.append(f"new_{name}") + + # (batch_size, embed_dim, conv_left_pad) + name = f"cached_conv1_{i}" + logging.info(f"{name}.shape: {tensors[4].shape}") + inputs[name] = {0: "N"} + outputs[f"new_{name}"] = {0: "N"} + input_names.append(name) + output_names.append(f"new_{name}") + + # (batch_size, embed_dim, conv_left_pad) + name = f"cached_conv2_{i}" + logging.info(f"{name}.shape: {tensors[5].shape}") + inputs[name] = {0: "N"} + outputs[f"new_{name}"] = {0: "N"} + input_names.append(name) + output_names.append(f"new_{name}") + + num_encoder_layers = ",".join(map(str, encoder_model.encoder.num_encoder_layers)) + encoder_dims = ",".join(map(str, encoder_model.encoder.encoder_dim)) + cnn_module_kernels = ",".join(map(str, encoder_model.encoder.cnn_module_kernel)) + ds = encoder_model.encoder.downsampling_factor + left_context_len = encoder_model.left_context_len + left_context_len = [left_context_len // k for k in ds] + left_context_len = ",".join(map(str, left_context_len)) + query_head_dims = ",".join(map(str, encoder_model.encoder.query_head_dim)) + value_head_dims = ",".join(map(str, encoder_model.encoder.value_head_dim)) + num_heads = ",".join(map(str, encoder_model.encoder.num_heads)) + + meta_data = { + "model_type": "zipformer2", + "version": "1", + "model_author": "k2-fsa", + "comment": "streaming zipformer2", + "decode_chunk_len": str(decode_chunk_len), # 32 + "T": str(T), # 32+7+2*3=45 + "num_encoder_layers": num_encoder_layers, + "encoder_dims": encoder_dims, + "cnn_module_kernels": cnn_module_kernels, + "left_context_len": left_context_len, + "query_head_dims": query_head_dims, + "value_head_dims": value_head_dims, + "num_heads": num_heads, + } + logging.info(f"meta_data: {meta_data}") + + for i in range(len(init_state[:-2]) // 6): + build_inputs_outputs(init_state[i * 6 : (i + 1) * 6], i) + + # (batch_size, channels, left_pad, freq) + embed_states = init_state[-2] + name = "embed_states" + logging.info(f"{name}.shape: {embed_states.shape}") + inputs[name] = {0: "N"} + outputs[f"new_{name}"] = {0: "N"} + input_names.append(name) + output_names.append(f"new_{name}") + + # (batch_size,) + processed_lens = init_state[-1] + name = "processed_lens" + logging.info(f"{name}.shape: {processed_lens.shape}") + inputs[name] = {0: "N"} + outputs[f"new_{name}"] = {0: "N"} + input_names.append(name) + output_names.append(f"new_{name}") + + logging.info(inputs) + logging.info(outputs) + logging.info(input_names) + logging.info(output_names) + + torch.onnx.export( + encoder_model, + (x, init_state), + encoder_filename, + verbose=False, + opset_version=opset_version, + input_names=input_names, + output_names=output_names, + dynamic_axes={ + "x": {0: "N"}, + "encoder_out": {0: "N"}, + **inputs, + **outputs, + }, + ) + + add_meta_data(filename=encoder_filename, meta_data=meta_data) + + +def export_decoder_model_onnx( + decoder_model: OnnxDecoder, + decoder_filename: str, + opset_version: int = 11, +) -> None: + """Export the decoder model to ONNX format. + + The exported model has one input: + + - y: a torch.int64 tensor of shape (N, decoder_model.context_size) + + and has one output: + + - decoder_out: a torch.float32 tensor of shape (N, joiner_dim) + + Args: + decoder_model: + The decoder model to be exported. + decoder_filename: + Filename to save the exported ONNX model. + opset_version: + The opset version to use. + """ + context_size = decoder_model.decoder.context_size + vocab_size = decoder_model.decoder.vocab_size + + y = torch.zeros(10, context_size, dtype=torch.int64) + decoder_model = torch.jit.script(decoder_model) + torch.onnx.export( + decoder_model, + y, + decoder_filename, + verbose=False, + opset_version=opset_version, + input_names=["y"], + output_names=["decoder_out"], + dynamic_axes={ + "y": {0: "N"}, + "decoder_out": {0: "N"}, + }, + ) + + meta_data = { + "context_size": str(context_size), + "vocab_size": str(vocab_size), + } + add_meta_data(filename=decoder_filename, meta_data=meta_data) + + +def export_joiner_model_onnx( + joiner_model: nn.Module, + joiner_filename: str, + opset_version: int = 11, +) -> None: + """Export the joiner model to ONNX format. + The exported joiner model has two inputs: + + - encoder_out: a tensor of shape (N, joiner_dim) + - decoder_out: a tensor of shape (N, joiner_dim) + + and produces one output: + + - logit: a tensor of shape (N, vocab_size) + """ + joiner_dim = joiner_model.output_linear.weight.shape[1] + logging.info(f"joiner dim: {joiner_dim}") + + projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32) + projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32) + + torch.onnx.export( + joiner_model, + (projected_encoder_out, projected_decoder_out), + joiner_filename, + verbose=False, + opset_version=opset_version, + input_names=[ + "encoder_out", + "decoder_out", + ], + output_names=["logit"], + dynamic_axes={ + "encoder_out": {0: "N"}, + "decoder_out": {0: "N"}, + "logit": {0: "N"}, + }, + ) + meta_data = { + "joiner_dim": str(joiner_dim), + } + add_meta_data(filename=joiner_filename, meta_data=meta_data) + + +@torch.no_grad() +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}") + + token_table = k2.SymbolTable.from_file(params.tokens) + params.blank_id = token_table[""] + params.vocab_size = num_tokens(token_table) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + model.to(device) + + if not params.use_averaged_model: + 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 i >= 1: + 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)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to("cpu") + model.eval() + + convert_scaled_to_non_scaled(model, inplace=True) + + encoder = OnnxEncoder( + encoder=model.encoder, + encoder_embed=model.encoder_embed, + encoder_proj=model.joiner.encoder_proj, + ) + + decoder = OnnxDecoder( + decoder=model.decoder, + decoder_proj=model.joiner.decoder_proj, + ) + + joiner = OnnxJoiner(output_linear=model.joiner.output_linear) + + encoder_num_param = sum([p.numel() for p in encoder.parameters()]) + decoder_num_param = sum([p.numel() for p in decoder.parameters()]) + joiner_num_param = sum([p.numel() for p in joiner.parameters()]) + total_num_param = encoder_num_param + decoder_num_param + joiner_num_param + logging.info(f"encoder parameters: {encoder_num_param}") + logging.info(f"decoder parameters: {decoder_num_param}") + logging.info(f"joiner parameters: {joiner_num_param}") + logging.info(f"total parameters: {total_num_param}") + + if params.iter > 0: + suffix = f"iter-{params.iter}" + else: + suffix = f"epoch-{params.epoch}" + + suffix += f"-avg-{params.avg}" + suffix += f"-chunk-{params.chunk_size}" + suffix += f"-left-{params.left_context_frames}" + + opset_version = 13 + + logging.info("Exporting encoder") + encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx" + export_encoder_model_onnx( + encoder, + encoder_filename, + opset_version=opset_version, + ) + logging.info(f"Exported encoder to {encoder_filename}") + + logging.info("Exporting decoder") + decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx" + export_decoder_model_onnx( + decoder, + decoder_filename, + opset_version=opset_version, + ) + logging.info(f"Exported decoder to {decoder_filename}") + + logging.info("Exporting joiner") + joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx" + export_joiner_model_onnx( + joiner, + joiner_filename, + opset_version=opset_version, + ) + logging.info(f"Exported joiner to {joiner_filename}") + + # Generate int8 quantization models + # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection + + logging.info("Generate int8 quantization models") + + encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx" + quantize_dynamic( + model_input=encoder_filename, + model_output=encoder_filename_int8, + op_types_to_quantize=["MatMul"], + weight_type=QuantType.QInt8, + ) + + decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx" + quantize_dynamic( + model_input=decoder_filename, + model_output=decoder_filename_int8, + op_types_to_quantize=["MatMul", "Gather"], + weight_type=QuantType.QInt8, + ) + + joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx" + quantize_dynamic( + model_input=joiner_filename, + model_output=joiner_filename_int8, + op_types_to_quantize=["MatMul"], + weight_type=QuantType.QInt8, + ) + + +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/ami/ASR/zipformer/export-onnx.py b/egs/ami/ASR/zipformer/export-onnx.py new file mode 100755 index 000000000..3682f0b62 --- /dev/null +++ b/egs/ami/ASR/zipformer/export-onnx.py @@ -0,0 +1,620 @@ +#!/usr/bin/env python3 +# +# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang, Wei Kang) +# Copyright 2023 Danqing Fu (danqing.fu@gmail.com) + +""" +This script exports a transducer model from PyTorch to ONNX. + +We use the pre-trained model from +https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +as an example to show how to use this file. + +1. Download the pre-trained model + +cd egs/librispeech/ASR + +repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url +repo=$(basename $repo_url) + +pushd $repo +git lfs pull --include "exp/pretrained.pt" + +cd exp +ln -s pretrained.pt epoch-99.pt +popd + +2. Export the model to ONNX + +./zipformer/export-onnx.py \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ + --use-averaged-model 0 \ + --epoch 99 \ + --avg 1 \ + --exp-dir $repo/exp \ + --num-encoder-layers "2,2,3,4,3,2" \ + --downsampling-factor "1,2,4,8,4,2" \ + --feedforward-dim "512,768,1024,1536,1024,768" \ + --num-heads "4,4,4,8,4,4" \ + --encoder-dim "192,256,384,512,384,256" \ + --query-head-dim 32 \ + --value-head-dim 12 \ + --pos-head-dim 4 \ + --pos-dim 48 \ + --encoder-unmasked-dim "192,192,256,256,256,192" \ + --cnn-module-kernel "31,31,15,15,15,31" \ + --decoder-dim 512 \ + --joiner-dim 512 \ + --causal False \ + --chunk-size "16,32,64,-1" \ + --left-context-frames "64,128,256,-1" + +It will generate the following 3 files inside $repo/exp: + + - encoder-epoch-99-avg-1.onnx + - decoder-epoch-99-avg-1.onnx + - joiner-epoch-99-avg-1.onnx + +See ./onnx_pretrained.py and ./onnx_check.py for how to +use the exported ONNX models. +""" + +import argparse +import logging +from pathlib import Path +from typing import Dict, Tuple + +import k2 +import onnx +import torch +import torch.nn as nn +from decoder import Decoder +from onnxruntime.quantization import QuantType, quantize_dynamic +from scaling_converter import convert_scaled_to_non_scaled +from train import add_model_arguments, get_model, get_params +from zipformer import Zipformer2 + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import make_pad_mask, num_tokens, 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( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="""It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--tokens", + type=str, + default="data/lang_bpe_500/tokens.txt", + help="Path to the tokens.txt", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + add_model_arguments(parser) + + return parser + + +def add_meta_data(filename: str, meta_data: Dict[str, str]): + """Add meta data to an ONNX model. It is changed in-place. + + Args: + filename: + Filename of the ONNX model to be changed. + meta_data: + Key-value pairs. + """ + model = onnx.load(filename) + for key, value in meta_data.items(): + meta = model.metadata_props.add() + meta.key = key + meta.value = value + + onnx.save(model, filename) + + +class OnnxEncoder(nn.Module): + """A wrapper for Zipformer and the encoder_proj from the joiner""" + + def __init__( + self, encoder: Zipformer2, encoder_embed: nn.Module, encoder_proj: nn.Linear + ): + """ + Args: + encoder: + A Zipformer encoder. + encoder_proj: + The projection layer for encoder from the joiner. + """ + super().__init__() + self.encoder = encoder + self.encoder_embed = encoder_embed + self.encoder_proj = encoder_proj + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Please see the help information of Zipformer.forward + + Args: + x: + A 3-D tensor of shape (N, T, C) + x_lens: + A 1-D tensor of shape (N,). Its dtype is torch.int64 + Returns: + Return a tuple containing: + - encoder_out, A 3-D tensor of shape (N, T', joiner_dim) + - encoder_out_lens, A 1-D tensor of shape (N,) + """ + x, x_lens = self.encoder_embed(x, x_lens) + src_key_padding_mask = make_pad_mask(x_lens) + x = x.permute(1, 0, 2) + encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask) + encoder_out = encoder_out.permute(1, 0, 2) + encoder_out = self.encoder_proj(encoder_out) + # Now encoder_out is of shape (N, T, joiner_dim) + + return encoder_out, encoder_out_lens + + +class OnnxDecoder(nn.Module): + """A wrapper for Decoder and the decoder_proj from the joiner""" + + def __init__(self, decoder: Decoder, decoder_proj: nn.Linear): + super().__init__() + self.decoder = decoder + self.decoder_proj = decoder_proj + + def forward(self, y: torch.Tensor) -> torch.Tensor: + """ + Args: + y: + A 2-D tensor of shape (N, context_size). + Returns + Return a 2-D tensor of shape (N, joiner_dim) + """ + need_pad = False + decoder_output = self.decoder(y, need_pad=need_pad) + decoder_output = decoder_output.squeeze(1) + output = self.decoder_proj(decoder_output) + + return output + + +class OnnxJoiner(nn.Module): + """A wrapper for the joiner""" + + def __init__(self, output_linear: nn.Linear): + super().__init__() + self.output_linear = output_linear + + def forward( + self, + encoder_out: torch.Tensor, + decoder_out: torch.Tensor, + ) -> torch.Tensor: + """ + Args: + encoder_out: + A 2-D tensor of shape (N, joiner_dim) + decoder_out: + A 2-D tensor of shape (N, joiner_dim) + Returns: + Return a 2-D tensor of shape (N, vocab_size) + """ + logit = encoder_out + decoder_out + logit = self.output_linear(torch.tanh(logit)) + return logit + + +def export_encoder_model_onnx( + encoder_model: OnnxEncoder, + encoder_filename: str, + opset_version: int = 11, +) -> None: + """Export the given encoder model to ONNX format. + The exported model has two inputs: + + - x, a tensor of shape (N, T, C); dtype is torch.float32 + - x_lens, a tensor of shape (N,); dtype is torch.int64 + + and it has two outputs: + + - encoder_out, a tensor of shape (N, T', joiner_dim) + - encoder_out_lens, a tensor of shape (N,) + + Args: + encoder_model: + The input encoder model + encoder_filename: + The filename to save the exported ONNX model. + opset_version: + The opset version to use. + """ + x = torch.zeros(1, 100, 80, dtype=torch.float32) + x_lens = torch.tensor([100], dtype=torch.int64) + + encoder_model = torch.jit.trace(encoder_model, (x, x_lens)) + + torch.onnx.export( + encoder_model, + (x, x_lens), + encoder_filename, + verbose=False, + opset_version=opset_version, + input_names=["x", "x_lens"], + output_names=["encoder_out", "encoder_out_lens"], + dynamic_axes={ + "x": {0: "N", 1: "T"}, + "x_lens": {0: "N"}, + "encoder_out": {0: "N", 1: "T"}, + "encoder_out_lens": {0: "N"}, + }, + ) + + meta_data = { + "model_type": "zipformer2", + "version": "1", + "model_author": "k2-fsa", + "comment": "non-streaming zipformer2", + } + logging.info(f"meta_data: {meta_data}") + + add_meta_data(filename=encoder_filename, meta_data=meta_data) + + +def export_decoder_model_onnx( + decoder_model: OnnxDecoder, + decoder_filename: str, + opset_version: int = 11, +) -> None: + """Export the decoder model to ONNX format. + + The exported model has one input: + + - y: a torch.int64 tensor of shape (N, decoder_model.context_size) + + and has one output: + + - decoder_out: a torch.float32 tensor of shape (N, joiner_dim) + + Args: + decoder_model: + The decoder model to be exported. + decoder_filename: + Filename to save the exported ONNX model. + opset_version: + The opset version to use. + """ + context_size = decoder_model.decoder.context_size + vocab_size = decoder_model.decoder.vocab_size + + y = torch.zeros(10, context_size, dtype=torch.int64) + decoder_model = torch.jit.script(decoder_model) + torch.onnx.export( + decoder_model, + y, + decoder_filename, + verbose=False, + opset_version=opset_version, + input_names=["y"], + output_names=["decoder_out"], + dynamic_axes={ + "y": {0: "N"}, + "decoder_out": {0: "N"}, + }, + ) + + meta_data = { + "context_size": str(context_size), + "vocab_size": str(vocab_size), + } + add_meta_data(filename=decoder_filename, meta_data=meta_data) + + +def export_joiner_model_onnx( + joiner_model: nn.Module, + joiner_filename: str, + opset_version: int = 11, +) -> None: + """Export the joiner model to ONNX format. + The exported joiner model has two inputs: + + - encoder_out: a tensor of shape (N, joiner_dim) + - decoder_out: a tensor of shape (N, joiner_dim) + + and produces one output: + + - logit: a tensor of shape (N, vocab_size) + """ + joiner_dim = joiner_model.output_linear.weight.shape[1] + logging.info(f"joiner dim: {joiner_dim}") + + projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32) + projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32) + + torch.onnx.export( + joiner_model, + (projected_encoder_out, projected_decoder_out), + joiner_filename, + verbose=False, + opset_version=opset_version, + input_names=[ + "encoder_out", + "decoder_out", + ], + output_names=["logit"], + dynamic_axes={ + "encoder_out": {0: "N"}, + "decoder_out": {0: "N"}, + "logit": {0: "N"}, + }, + ) + meta_data = { + "joiner_dim": str(joiner_dim), + } + add_meta_data(filename=joiner_filename, meta_data=meta_data) + + +@torch.no_grad() +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}") + + token_table = k2.SymbolTable.from_file(params.tokens) + params.blank_id = token_table[""] + params.vocab_size = num_tokens(token_table) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + model.to(device) + + if not params.use_averaged_model: + 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 i >= 1: + 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)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to("cpu") + model.eval() + + convert_scaled_to_non_scaled(model, inplace=True, is_onnx=True) + + encoder = OnnxEncoder( + encoder=model.encoder, + encoder_embed=model.encoder_embed, + encoder_proj=model.joiner.encoder_proj, + ) + + decoder = OnnxDecoder( + decoder=model.decoder, + decoder_proj=model.joiner.decoder_proj, + ) + + joiner = OnnxJoiner(output_linear=model.joiner.output_linear) + + encoder_num_param = sum([p.numel() for p in encoder.parameters()]) + decoder_num_param = sum([p.numel() for p in decoder.parameters()]) + joiner_num_param = sum([p.numel() for p in joiner.parameters()]) + total_num_param = encoder_num_param + decoder_num_param + joiner_num_param + logging.info(f"encoder parameters: {encoder_num_param}") + logging.info(f"decoder parameters: {decoder_num_param}") + logging.info(f"joiner parameters: {joiner_num_param}") + logging.info(f"total parameters: {total_num_param}") + + if params.iter > 0: + suffix = f"iter-{params.iter}" + else: + suffix = f"epoch-{params.epoch}" + + suffix += f"-avg-{params.avg}" + + opset_version = 13 + + logging.info("Exporting encoder") + encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx" + export_encoder_model_onnx( + encoder, + encoder_filename, + opset_version=opset_version, + ) + logging.info(f"Exported encoder to {encoder_filename}") + + logging.info("Exporting decoder") + decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx" + export_decoder_model_onnx( + decoder, + decoder_filename, + opset_version=opset_version, + ) + logging.info(f"Exported decoder to {decoder_filename}") + + logging.info("Exporting joiner") + joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx" + export_joiner_model_onnx( + joiner, + joiner_filename, + opset_version=opset_version, + ) + logging.info(f"Exported joiner to {joiner_filename}") + + # Generate int8 quantization models + # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection + + logging.info("Generate int8 quantization models") + + encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx" + quantize_dynamic( + model_input=encoder_filename, + model_output=encoder_filename_int8, + op_types_to_quantize=["MatMul"], + weight_type=QuantType.QInt8, + ) + + decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx" + quantize_dynamic( + model_input=decoder_filename, + model_output=decoder_filename_int8, + op_types_to_quantize=["MatMul", "Gather"], + weight_type=QuantType.QInt8, + ) + + joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx" + quantize_dynamic( + model_input=joiner_filename, + model_output=joiner_filename_int8, + op_types_to_quantize=["MatMul"], + weight_type=QuantType.QInt8, + ) + + +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/ami/ASR/zipformer/export.py b/egs/ami/ASR/zipformer/export.py new file mode 100755 index 000000000..2b8d1aaf3 --- /dev/null +++ b/egs/ami/ASR/zipformer/export.py @@ -0,0 +1,526 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao, +# 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. + +# This script converts several saved checkpoints +# to a single one using model averaging. +""" + +Usage: + +Note: This is a example for librispeech dataset, if you are using different +dataset, you should change the argument values according to your dataset. + +(1) Export to torchscript model using torch.jit.script() + +- For non-streaming model: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --tokens data/lang_bpe_500/tokens.txt \ + --epoch 30 \ + --avg 9 \ + --jit 1 + +It will generate a file `jit_script.pt` in the given `exp_dir`. You can later +load it by `torch.jit.load("jit_script.pt")`. + +Check ./jit_pretrained.py for its usage. + +Check https://github.com/k2-fsa/sherpa +for how to use the exported models outside of icefall. + +- For streaming model: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --causal 1 \ + --chunk-size 16 \ + --left-context-frames 128 \ + --tokens data/lang_bpe_500/tokens.txt \ + --epoch 30 \ + --avg 9 \ + --jit 1 + +It will generate a file `jit_script_chunk_16_left_128.pt` in the given `exp_dir`. +You can later load it by `torch.jit.load("jit_script_chunk_16_left_128.pt")`. + +Check ./jit_pretrained_streaming.py for its usage. + +Check https://github.com/k2-fsa/sherpa +for how to use the exported models outside of icefall. + +(2) Export `model.state_dict()` + +- For non-streaming model: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --tokens data/lang_bpe_500/tokens.txt \ + --epoch 30 \ + --avg 9 + +- For streaming model: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --causal 1 \ + --tokens data/lang_bpe_500/tokens.txt \ + --epoch 30 \ + --avg 9 + +It will generate a file `pretrained.pt` in the given `exp_dir`. You can later +load it by `icefall.checkpoint.load_checkpoint()`. + +- For non-streaming model: + +To use the generated file with `zipformer/decode.py`, +you can do: + + cd /path/to/exp_dir + ln -s pretrained.pt epoch-9999.pt + + cd /path/to/egs/librispeech/ASR + ./zipformer/decode.py \ + --exp-dir ./zipformer/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 600 \ + --decoding-method greedy_search \ + --bpe-model data/lang_bpe_500/bpe.model + +- For streaming model: + +To use the generated file with `zipformer/decode.py` and `zipformer/streaming_decode.py`, you can do: + + cd /path/to/exp_dir + ln -s pretrained.pt epoch-9999.pt + + cd /path/to/egs/librispeech/ASR + + # simulated streaming decoding + ./zipformer/decode.py \ + --exp-dir ./zipformer/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 600 \ + --causal 1 \ + --chunk-size 16 \ + --left-context-frames 128 \ + --decoding-method greedy_search \ + --bpe-model data/lang_bpe_500/bpe.model + + # chunk-wise streaming decoding + ./zipformer/streaming_decode.py \ + --exp-dir ./zipformer/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 600 \ + --causal 1 \ + --chunk-size 16 \ + --left-context-frames 128 \ + --decoding-method greedy_search \ + --bpe-model data/lang_bpe_500/bpe.model + +Check ./pretrained.py for its usage. + +Note: If you don't want to train a model from scratch, we have +provided one for you. You can get it at + +- non-streaming model: +https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 + +- streaming model: +https://huggingface.co/Zengwei/icefall-asr-librispeech-streaming-zipformer-2023-05-17 + +with the following commands: + + sudo apt-get install git-lfs + git lfs install + git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 + git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-streaming-zipformer-2023-05-17 + # You will find the pre-trained models in exp dir +""" + +import argparse +import logging +from pathlib import Path +from typing import List, Tuple + +import k2 +import torch +from scaling_converter import convert_scaled_to_non_scaled +from torch import Tensor, nn +from train import add_model_arguments, get_model, get_params + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import make_pad_mask, num_tokens, str2bool + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + 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=9, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="""It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--tokens", + type=str, + default="data/lang_bpe_500/tokens.txt", + help="Path to the tokens.txt", + ) + + parser.add_argument( + "--jit", + type=str2bool, + default=False, + help="""True to save a model after applying torch.jit.script. + It will generate a file named jit_script.pt. + Check ./jit_pretrained.py for how to use it. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + add_model_arguments(parser) + + return parser + + +class EncoderModel(nn.Module): + """A wrapper for encoder and encoder_embed""" + + def __init__(self, encoder: nn.Module, encoder_embed: nn.Module) -> None: + super().__init__() + self.encoder = encoder + self.encoder_embed = encoder_embed + + def forward( + self, features: Tensor, feature_lengths: Tensor + ) -> Tuple[Tensor, Tensor]: + """ + Args: + features: (N, T, C) + feature_lengths: (N,) + """ + x, x_lens = self.encoder_embed(features, feature_lengths) + + src_key_padding_mask = make_pad_mask(x_lens) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + return encoder_out, encoder_out_lens + + +class StreamingEncoderModel(nn.Module): + """A wrapper for encoder and encoder_embed""" + + def __init__(self, encoder: nn.Module, encoder_embed: nn.Module) -> None: + super().__init__() + assert len(encoder.chunk_size) == 1, encoder.chunk_size + assert len(encoder.left_context_frames) == 1, encoder.left_context_frames + self.chunk_size = encoder.chunk_size[0] + self.left_context_len = encoder.left_context_frames[0] + + # The encoder_embed subsample features (T - 7) // 2 + # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling + self.pad_length = 7 + 2 * 3 + + self.encoder = encoder + self.encoder_embed = encoder_embed + + def forward( + self, features: Tensor, feature_lengths: Tensor, states: List[Tensor] + ) -> Tuple[Tensor, Tensor, List[Tensor]]: + """Streaming forward for encoder_embed and encoder. + + Args: + features: (N, T, C) + feature_lengths: (N,) + states: a list of Tensors + + Returns encoder outputs, output lengths, and updated states. + """ + chunk_size = self.chunk_size + left_context_len = self.left_context_len + + cached_embed_left_pad = states[-2] + x, x_lens, new_cached_embed_left_pad = self.encoder_embed.streaming_forward( + x=features, + x_lens=feature_lengths, + cached_left_pad=cached_embed_left_pad, + ) + assert x.size(1) == chunk_size, (x.size(1), chunk_size) + + src_key_padding_mask = make_pad_mask(x_lens) + + # processed_mask is used to mask out initial states + processed_mask = torch.arange(left_context_len, device=x.device).expand( + x.size(0), left_context_len + ) + processed_lens = states[-1] # (batch,) + # (batch, left_context_size) + processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1) + # Update processed lengths + new_processed_lens = processed_lens + x_lens + + # (batch, left_context_size + chunk_size) + src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1) + + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + encoder_states = states[:-2] + + ( + encoder_out, + encoder_out_lens, + new_encoder_states, + ) = self.encoder.streaming_forward( + x=x, + x_lens=x_lens, + states=encoder_states, + src_key_padding_mask=src_key_padding_mask, + ) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + new_states = new_encoder_states + [ + new_cached_embed_left_pad, + new_processed_lens, + ] + return encoder_out, encoder_out_lens, new_states + + @torch.jit.export + def get_init_states( + self, + batch_size: int = 1, + device: torch.device = torch.device("cpu"), + ) -> List[torch.Tensor]: + """ + Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6] + is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). + states[-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + states[-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + """ + states = self.encoder.get_init_states(batch_size, device) + + embed_states = self.encoder_embed.get_init_states(batch_size, device) + states.append(embed_states) + + processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device) + states.append(processed_lens) + + return states + + +@torch.no_grad() +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}") + + token_table = k2.SymbolTable.from_file(params.tokens) + params.blank_id = token_table[""] + params.vocab_size = num_tokens(token_table) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + 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.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 i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.eval() + + if params.jit is True: + convert_scaled_to_non_scaled(model, inplace=True) + # 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) + + # Wrap encoder and encoder_embed as a module + if params.causal: + model.encoder = StreamingEncoderModel(model.encoder, model.encoder_embed) + chunk_size = model.encoder.chunk_size + left_context_len = model.encoder.left_context_len + filename = f"jit_script_chunk_{chunk_size}_left_{left_context_len}.pt" + else: + model.encoder = EncoderModel(model.encoder, model.encoder_embed) + filename = "jit_script.pt" + + logging.info("Using torch.jit.script") + model = torch.jit.script(model) + model.save(str(params.exp_dir / filename)) + logging.info(f"Saved to {filename}") + else: + logging.info("Not using torchscript. Export model.state_dict()") + # 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/ami/ASR/zipformer/generate_averaged_model.py b/egs/ami/ASR/zipformer/generate_averaged_model.py new file mode 120000 index 000000000..5a015ee6c --- /dev/null +++ b/egs/ami/ASR/zipformer/generate_averaged_model.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/generate_averaged_model.py \ No newline at end of file diff --git a/egs/ami/ASR/zipformer/jit_pretrained.py b/egs/ami/ASR/zipformer/jit_pretrained.py new file mode 100755 index 000000000..a41fbc1c9 --- /dev/null +++ b/egs/ami/ASR/zipformer/jit_pretrained.py @@ -0,0 +1,280 @@ +#!/usr/bin/env python3 +# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, Zengwei Yao) +# +# 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 loads torchscript models, exported by `torch.jit.script()` +and uses them to decode waves. +You can use the following command to get the exported models: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --tokens data/lang_bpe_500/tokens.txt \ + --epoch 30 \ + --avg 9 \ + --jit 1 + +Usage of this script: + +./zipformer/jit_pretrained.py \ + --nn-model-filename ./zipformer/exp/cpu_jit.pt \ + --tokens ./data/lang_bpe_500/tokens.txt \ + /path/to/foo.wav \ + /path/to/bar.wav +""" + +import argparse +import logging +import math +from typing import List + +import k2 +import kaldifeat +import torch +import torchaudio +from torch.nn.utils.rnn import pad_sequence + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--nn-model-filename", + type=str, + required=True, + help="Path to the torchscript model cpu_jit.pt", + ) + + parser.add_argument( + "--tokens", + type=str, + help="""Path to tokens.txt.""", + ) + + 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.", + ) + + return parser + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float = 16000 +) -> 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}. Given: {sample_rate}" + # We use only the first channel + ans.append(wave[0].contiguous()) + return ans + + +def greedy_search( + model: torch.jit.ScriptModule, + 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: + A 3-D tensor of shape (N, T, C) + encoder_out_lens: + A 1-D tensor of shape (N,). + Returns: + Return the decoded results for each utterance. + """ + 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 = encoder_out.device + blank_id = model.decoder.blank_id + + 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) + + context_size = model.decoder.context_size + 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=torch.tensor([False]), + ).squeeze(1) + + offset = 0 + for batch_size in batch_size_list: + start = offset + end = offset + batch_size + current_encoder_out = packed_encoder_out.data[start:end] + current_encoder_out = current_encoder_out + # current_encoder_out's shape: (batch_size, encoder_out_dim) + offset = end + + decoder_out = decoder_out[:batch_size] + + logits = model.joiner( + current_encoder_out, + decoder_out, + ) + # logits'shape (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 != blank_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=torch.tensor([False]), + ) + decoder_out = decoder_out.squeeze(1) + + 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 + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + logging.info(vars(args)) + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + model = torch.jit.load(args.nn_model_filename) + + model.eval() + + model.to(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 = 16000 + opts.mel_opts.num_bins = 80 + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {args.sound_files}") + waves = read_sound_files( + filenames=args.sound_files, + ) + 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) + + encoder_out, encoder_out_lens = model.encoder( + features=features, + feature_lengths=feature_lengths, + ) + + hyps = greedy_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + + s = "\n" + + token_table = k2.SymbolTable.from_file(args.tokens) + + def token_ids_to_words(token_ids: List[int]) -> str: + text = "" + for i in token_ids: + text += token_table[i] + return text.replace("▁", " ").strip() + + for filename, hyp in zip(args.sound_files, hyps): + words = token_ids_to_words(hyp) + s += f"{filename}:\n{words}\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/ami/ASR/zipformer/jit_pretrained_ctc.py b/egs/ami/ASR/zipformer/jit_pretrained_ctc.py new file mode 100755 index 000000000..660a4bfc6 --- /dev/null +++ b/egs/ami/ASR/zipformer/jit_pretrained_ctc.py @@ -0,0 +1,436 @@ +#!/usr/bin/env python3 +# Copyright 2022-2023 Xiaomi Corp. (authors: Fangjun Kuang, +# Zengwei Yao) +# +# 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 loads a checkpoint and uses it to decode waves. +You can generate the checkpoint with the following command: + +- For non-streaming model: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --use-ctc 1 \ + --tokens data/lang_bpe_500/tokens.txt \ + --epoch 30 \ + --avg 9 \ + --jit 1 + +- For streaming model: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --use-ctc 1 \ + --causal 1 \ + --tokens data/lang_bpe_500/tokens.txt \ + --epoch 30 \ + --avg 9 \ + --jit 1 + +Usage of this script: + +(1) ctc-decoding +./zipformer/jit_pretrained_ctc.py \ + --model-filename ./zipformer/exp/jit_script.pt \ + --tokens data/lang_bpe_500/tokens.txt \ + --method ctc-decoding \ + --sample-rate 16000 \ + /path/to/foo.wav \ + /path/to/bar.wav + +(2) 1best +./zipformer/jit_pretrained_ctc.py \ + --model-filename ./zipformer/exp/jit_script.pt \ + --HLG data/lang_bpe_500/HLG.pt \ + --words-file data/lang_bpe_500/words.txt \ + --method 1best \ + --sample-rate 16000 \ + /path/to/foo.wav \ + /path/to/bar.wav + +(3) nbest-rescoring +./zipformer/jit_pretrained_ctc.py \ + --model-filename ./zipformer/exp/jit_script.pt \ + --HLG data/lang_bpe_500/HLG.pt \ + --words-file data/lang_bpe_500/words.txt \ + --G data/lm/G_4_gram.pt \ + --method nbest-rescoring \ + --sample-rate 16000 \ + /path/to/foo.wav \ + /path/to/bar.wav + +(4) whole-lattice-rescoring +./zipformer/jit_pretrained_ctc.py \ + --model-filename ./zipformer/exp/jit_script.pt \ + --HLG data/lang_bpe_500/HLG.pt \ + --words-file data/lang_bpe_500/words.txt \ + --G data/lm/G_4_gram.pt \ + --method whole-lattice-rescoring \ + --sample-rate 16000 \ + /path/to/foo.wav \ + /path/to/bar.wav +""" + +import argparse +import logging +import math +from typing import List + +import k2 +import kaldifeat +import torch +import torchaudio +from ctc_decode import get_decoding_params +from export import num_tokens +from torch.nn.utils.rnn import pad_sequence +from train import get_params + +from icefall.decode import ( + get_lattice, + one_best_decoding, + rescore_with_n_best_list, + rescore_with_whole_lattice, +) +from icefall.utils import get_texts + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--model-filename", + type=str, + required=True, + help="Path to the torchscript model.", + ) + + parser.add_argument( + "--words-file", + type=str, + help="""Path to words.txt. + Used only when method is not ctc-decoding. + """, + ) + + parser.add_argument( + "--HLG", + type=str, + help="""Path to HLG.pt. + Used only when method is not ctc-decoding. + """, + ) + + parser.add_argument( + "--tokens", + type=str, + help="""Path to tokens.txt. + Used only when method is ctc-decoding. + """, + ) + + parser.add_argument( + "--method", + type=str, + default="1best", + help="""Decoding method. + Possible values are: + (0) ctc-decoding - Use CTC decoding. It uses a token table, + i.e., lang_dir/token.txt, to convert + word pieces to words. It needs neither a lexicon + nor an n-gram LM. + (1) 1best - Use the best path as decoding output. Only + the transformer encoder output is used for decoding. + We call it HLG decoding. + (2) nbest-rescoring. Extract n paths from the decoding lattice, + rescore them with an LM, the path with + the highest score is the decoding result. + We call it HLG decoding + nbest n-gram LM rescoring. + (3) whole-lattice-rescoring - Use an LM to rescore the + decoding lattice and then use 1best to decode the + rescored lattice. + We call it HLG decoding + whole-lattice n-gram LM rescoring. + """, + ) + + parser.add_argument( + "--G", + type=str, + help="""An LM for rescoring. + Used only when method is + whole-lattice-rescoring or nbest-rescoring. + It's usually a 4-gram LM. + """, + ) + + parser.add_argument( + "--num-paths", + type=int, + default=100, + help=""" + Used only when method is attention-decoder. + It specifies the size of n-best list.""", + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=1.3, + help=""" + Used only when method is whole-lattice-rescoring and nbest-rescoring. + It specifies the scale for n-gram LM scores. + (Note: You need to tune it on a dataset.) + """, + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=1.0, + help=""" + Used only when method is nbest-rescoring. + It specifies the scale for lattice.scores when + extracting n-best lists. A smaller value results in + more unique number of paths with the risk of missing + the best path. + """, + ) + + parser.add_argument( + "--sample-rate", + type=int, + default=16000, + help="The sample rate of the input sound file", + ) + + 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.", + ) + + return parser + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float = 16000 +) -> 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}. Given: {sample_rate}" + # We use only the first channel + ans.append(wave[0].contiguous()) + return ans + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + + params = get_params() + # add decoding params + params.update(get_decoding_params()) + params.update(vars(args)) + + token_table = k2.SymbolTable.from_file(params.tokens) + params.vocab_size = num_tokens(token_table) + 1 + + logging.info(f"{params}") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + model = torch.jit.load(args.model_filename) + model.to(device) + model.eval() + + 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) + + encoder_out, encoder_out_lens = model.encoder(features, feature_lengths) + ctc_output = model.ctc_output(encoder_out) # (N, T, C) + + batch_size = ctc_output.shape[0] + supervision_segments = torch.tensor( + [ + [i, 0, feature_lengths[i].item() // params.subsampling_factor] + for i in range(batch_size) + ], + dtype=torch.int32, + ) + + if params.method == "ctc-decoding": + logging.info("Use CTC decoding") + max_token_id = params.vocab_size - 1 + + H = k2.ctc_topo( + max_token=max_token_id, + modified=False, + device=device, + ) + + lattice = get_lattice( + nnet_output=ctc_output, + decoding_graph=H, + supervision_segments=supervision_segments, + search_beam=params.search_beam, + output_beam=params.output_beam, + min_active_states=params.min_active_states, + max_active_states=params.max_active_states, + subsampling_factor=params.subsampling_factor, + ) + + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + token_ids = get_texts(best_path) + hyps = [[token_table[i] for i in ids] for ids in token_ids] + elif params.method in [ + "1best", + "nbest-rescoring", + "whole-lattice-rescoring", + ]: + logging.info(f"Loading HLG from {params.HLG}") + HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu")) + HLG = HLG.to(device) + if not hasattr(HLG, "lm_scores"): + # For whole-lattice-rescoring and attention-decoder + HLG.lm_scores = HLG.scores.clone() + + if params.method in [ + "nbest-rescoring", + "whole-lattice-rescoring", + ]: + logging.info(f"Loading G from {params.G}") + G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu")) + G = G.to(device) + if params.method == "whole-lattice-rescoring": + # Add epsilon self-loops to G as we will compose + # it with the whole lattice later + G = k2.add_epsilon_self_loops(G) + G = k2.arc_sort(G) + + # G.lm_scores is used to replace HLG.lm_scores during + # LM rescoring. + G.lm_scores = G.scores.clone() + + lattice = get_lattice( + nnet_output=ctc_output, + decoding_graph=HLG, + supervision_segments=supervision_segments, + search_beam=params.search_beam, + output_beam=params.output_beam, + min_active_states=params.min_active_states, + max_active_states=params.max_active_states, + subsampling_factor=params.subsampling_factor, + ) + + if params.method == "1best": + logging.info("Use HLG decoding") + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + if params.method == "nbest-rescoring": + logging.info("Use HLG decoding + LM rescoring") + best_path_dict = rescore_with_n_best_list( + lattice=lattice, + G=G, + num_paths=params.num_paths, + lm_scale_list=[params.ngram_lm_scale], + nbest_scale=params.nbest_scale, + ) + best_path = next(iter(best_path_dict.values())) + elif params.method == "whole-lattice-rescoring": + logging.info("Use HLG decoding + LM rescoring") + best_path_dict = rescore_with_whole_lattice( + lattice=lattice, + G_with_epsilon_loops=G, + lm_scale_list=[params.ngram_lm_scale], + ) + best_path = next(iter(best_path_dict.values())) + + hyps = get_texts(best_path) + word_sym_table = k2.SymbolTable.from_file(params.words_file) + hyps = [[word_sym_table[i] for i in ids] for ids in hyps] + else: + raise ValueError(f"Unsupported decoding method: {params.method}") + + s = "\n" + if params.method == "ctc-decoding": + for filename, hyp in zip(params.sound_files, hyps): + words = "".join(hyp) + words = words.replace("▁", " ").strip() + s += f"{filename}:\n{words}\n\n" + elif params.method in [ + "1best", + "nbest-rescoring", + "whole-lattice-rescoring", + ]: + for filename, hyp in zip(params.sound_files, hyps): + words = " ".join(hyp) + words = words.replace("▁", " ").strip() + 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/ami/ASR/zipformer/jit_pretrained_streaming.py b/egs/ami/ASR/zipformer/jit_pretrained_streaming.py new file mode 100755 index 000000000..d4ceacefd --- /dev/null +++ b/egs/ami/ASR/zipformer/jit_pretrained_streaming.py @@ -0,0 +1,273 @@ +#!/usr/bin/env python3 +# flake8: noqa +# Copyright 2022-2023 Xiaomi Corp. (authors: Fangjun Kuang, Zengwei Yao) +# +# 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 loads torchscript models exported by `torch.jit.script()` +and uses them to decode waves. +You can use the following command to get the exported models: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --causal 1 \ + --chunk-size 16 \ + --left-context-frames 128 \ + --tokens data/lang_bpe_500/tokens.txt \ + --epoch 30 \ + --avg 9 \ + --jit 1 + +Usage of this script: + +./zipformer/jit_pretrained_streaming.py \ + --nn-model-filename ./zipformer/exp-causal/jit_script_chunk_16_left_128.pt \ + --tokens ./data/lang_bpe_500/tokens.txt \ + /path/to/foo.wav \ +""" + +import argparse +import logging +import math +from typing import List, Optional + +import k2 +import kaldifeat +import torch +import torchaudio +from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature +from torch.nn.utils.rnn import pad_sequence + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--nn-model-filename", + type=str, + required=True, + help="Path to the torchscript model jit_script.pt", + ) + + parser.add_argument( + "--tokens", + type=str, + help="""Path to tokens.txt.""", + ) + + parser.add_argument( + "--sample-rate", + type=int, + default=16000, + help="The sample rate of the input sound file", + ) + + parser.add_argument( + "sound_file", + type=str, + 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.", + ) + + 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}. Given: {sample_rate}" + # We use only the first channel + ans.append(wave[0]) + return ans + + +def greedy_search( + decoder: torch.jit.ScriptModule, + joiner: torch.jit.ScriptModule, + encoder_out: torch.Tensor, + decoder_out: Optional[torch.Tensor] = None, + hyp: Optional[List[int]] = None, + device: torch.device = torch.device("cpu"), +): + assert encoder_out.ndim == 2 + context_size = decoder.context_size + blank_id = decoder.blank_id + + if decoder_out is None: + assert hyp is None, hyp + hyp = [blank_id] * context_size + decoder_input = torch.tensor(hyp, dtype=torch.int32, device=device).unsqueeze(0) + # decoder_input.shape (1,, 1 context_size) + decoder_out = decoder(decoder_input, torch.tensor([False])).squeeze(1) + else: + assert decoder_out.ndim == 2 + assert hyp is not None, hyp + + T = encoder_out.size(0) + for i in range(T): + cur_encoder_out = encoder_out[i : i + 1] + joiner_out = joiner(cur_encoder_out, decoder_out).squeeze(0) + y = joiner_out.argmax(dim=0).item() + + if y != blank_id: + hyp.append(y) + decoder_input = hyp[-context_size:] + + decoder_input = torch.tensor( + decoder_input, dtype=torch.int32, device=device + ).unsqueeze(0) + decoder_out = decoder(decoder_input, torch.tensor([False])).squeeze(1) + + return hyp, decoder_out + + +def create_streaming_feature_extractor(sample_rate) -> OnlineFeature: + """Create a CPU streaming feature extractor. + + At present, we assume it returns a fbank feature extractor with + fixed options. In the future, we will support passing in the options + from outside. + + Returns: + Return a CPU streaming feature extractor. + """ + opts = FbankOptions() + opts.device = "cpu" + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = sample_rate + opts.mel_opts.num_bins = 80 + return OnlineFbank(opts) + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + logging.info(vars(args)) + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + model = torch.jit.load(args.nn_model_filename) + model.eval() + model.to(device) + + encoder = model.encoder + decoder = model.decoder + joiner = model.joiner + + token_table = k2.SymbolTable.from_file(args.tokens) + context_size = decoder.context_size + + logging.info("Constructing Fbank computer") + online_fbank = create_streaming_feature_extractor(args.sample_rate) + + logging.info(f"Reading sound files: {args.sound_file}") + wave_samples = read_sound_files( + filenames=[args.sound_file], + expected_sample_rate=args.sample_rate, + )[0] + logging.info(wave_samples.shape) + + logging.info("Decoding started") + + chunk_length = encoder.chunk_size * 2 + T = chunk_length + encoder.pad_length + + logging.info(f"chunk_length: {chunk_length}") + logging.info(f"T: {T}") + + states = encoder.get_init_states(device=device) + + tail_padding = torch.zeros(int(0.3 * args.sample_rate), dtype=torch.float32) + + wave_samples = torch.cat([wave_samples, tail_padding]) + + chunk = int(0.25 * args.sample_rate) # 0.2 second + num_processed_frames = 0 + + hyp = None + decoder_out = None + + start = 0 + while start < wave_samples.numel(): + logging.info(f"{start}/{wave_samples.numel()}") + end = min(start + chunk, wave_samples.numel()) + samples = wave_samples[start:end] + start += chunk + online_fbank.accept_waveform( + sampling_rate=args.sample_rate, + waveform=samples, + ) + while online_fbank.num_frames_ready - num_processed_frames >= T: + frames = [] + for i in range(T): + frames.append(online_fbank.get_frame(num_processed_frames + i)) + frames = torch.cat(frames, dim=0).to(device).unsqueeze(0) + x_lens = torch.tensor([T], dtype=torch.int32, device=device) + encoder_out, out_lens, states = encoder( + features=frames, + feature_lengths=x_lens, + states=states, + ) + num_processed_frames += chunk_length + + hyp, decoder_out = greedy_search( + decoder, joiner, encoder_out.squeeze(0), decoder_out, hyp, device=device + ) + + text = "" + for i in hyp[context_size:]: + text += token_table[i] + text = text.replace("▁", " ").strip() + + logging.info(args.sound_file) + logging.info(text) + + logging.info("Decoding Done") + + +torch.set_num_threads(4) +torch.set_num_interop_threads(1) +torch._C._jit_set_profiling_executor(False) +torch._C._jit_set_profiling_mode(False) +torch._C._set_graph_executor_optimize(False) +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/ami/ASR/zipformer/joiner.py b/egs/ami/ASR/zipformer/joiner.py new file mode 120000 index 000000000..5b8a36332 --- /dev/null +++ b/egs/ami/ASR/zipformer/joiner.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/joiner.py \ No newline at end of file diff --git a/egs/ami/ASR/zipformer/model.py b/egs/ami/ASR/zipformer/model.py new file mode 120000 index 000000000..cd7e07d72 --- /dev/null +++ b/egs/ami/ASR/zipformer/model.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/model.py \ No newline at end of file diff --git a/egs/ami/ASR/zipformer/onnx_check.py b/egs/ami/ASR/zipformer/onnx_check.py new file mode 100755 index 000000000..93bd3a211 --- /dev/null +++ b/egs/ami/ASR/zipformer/onnx_check.py @@ -0,0 +1,240 @@ +#!/usr/bin/env python3 +# +# Copyright 2022 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 checks that exported onnx models produce the same output +with the given torchscript model for the same input. + +We use the pre-trained model from +https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +as an example to show how to use this file. + +1. Download the pre-trained model + +cd egs/librispeech/ASR + +repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url +repo=$(basename $repo_url) + +pushd $repo +git lfs pull --include "exp/pretrained.pt" + +cd exp +ln -s pretrained.pt epoch-99.pt +popd + +2. Export the model via torchscript (torch.jit.script()) + +./zipformer/export.py \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ + --use-averaged-model 0 \ + --epoch 99 \ + --avg 1 \ + --exp-dir $repo/exp/ \ + --jit 1 + +It will generate the following file in $repo/exp: + - jit_script.pt + +3. Export the model to ONNX + +./zipformer/export-onnx.py \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ + --use-averaged-model 0 \ + --epoch 99 \ + --avg 1 \ + --exp-dir $repo/exp/ + +It will generate the following 3 files inside $repo/exp: + + - encoder-epoch-99-avg-1.onnx + - decoder-epoch-99-avg-1.onnx + - joiner-epoch-99-avg-1.onnx + +4. Run this file + +./zipformer/onnx_check.py \ + --jit-filename $repo/exp/jit_script.pt \ + --onnx-encoder-filename $repo/exp/encoder-epoch-99-avg-1.onnx \ + --onnx-decoder-filename $repo/exp/decoder-epoch-99-avg-1.onnx \ + --onnx-joiner-filename $repo/exp/joiner-epoch-99-avg-1.onnx +""" + +import argparse +import logging + +import torch +from onnx_pretrained import OnnxModel + +from icefall import is_module_available + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--jit-filename", + required=True, + type=str, + help="Path to the torchscript model", + ) + + parser.add_argument( + "--onnx-encoder-filename", + required=True, + type=str, + help="Path to the onnx encoder model", + ) + + parser.add_argument( + "--onnx-decoder-filename", + required=True, + type=str, + help="Path to the onnx decoder model", + ) + + parser.add_argument( + "--onnx-joiner-filename", + required=True, + type=str, + help="Path to the onnx joiner model", + ) + + return parser + + +def test_encoder( + torch_model: torch.jit.ScriptModule, + onnx_model: OnnxModel, +): + C = 80 + for i in range(3): + N = torch.randint(low=1, high=20, size=(1,)).item() + T = torch.randint(low=30, high=50, size=(1,)).item() + logging.info(f"test_encoder: iter {i}, N={N}, T={T}") + + x = torch.rand(N, T, C) + x_lens = torch.randint(low=30, high=T + 1, size=(N,)) + x_lens[0] = T + + torch_encoder_out, torch_encoder_out_lens = torch_model.encoder(x, x_lens) + torch_encoder_out = torch_model.joiner.encoder_proj(torch_encoder_out) + + onnx_encoder_out, onnx_encoder_out_lens = onnx_model.run_encoder(x, x_lens) + + assert torch.allclose(torch_encoder_out, onnx_encoder_out, atol=1e-05), ( + (torch_encoder_out - onnx_encoder_out).abs().max() + ) + + +def test_decoder( + torch_model: torch.jit.ScriptModule, + onnx_model: OnnxModel, +): + context_size = onnx_model.context_size + vocab_size = onnx_model.vocab_size + for i in range(10): + N = torch.randint(1, 100, size=(1,)).item() + logging.info(f"test_decoder: iter {i}, N={N}") + x = torch.randint( + low=1, + high=vocab_size, + size=(N, context_size), + dtype=torch.int64, + ) + torch_decoder_out = torch_model.decoder(x, need_pad=torch.tensor([False])) + torch_decoder_out = torch_model.joiner.decoder_proj(torch_decoder_out) + torch_decoder_out = torch_decoder_out.squeeze(1) + + onnx_decoder_out = onnx_model.run_decoder(x) + assert torch.allclose(torch_decoder_out, onnx_decoder_out, atol=1e-4), ( + (torch_decoder_out - onnx_decoder_out).abs().max() + ) + + +def test_joiner( + torch_model: torch.jit.ScriptModule, + onnx_model: OnnxModel, +): + encoder_dim = torch_model.joiner.encoder_proj.weight.shape[1] + decoder_dim = torch_model.joiner.decoder_proj.weight.shape[1] + for i in range(10): + N = torch.randint(1, 100, size=(1,)).item() + logging.info(f"test_joiner: iter {i}, N={N}") + encoder_out = torch.rand(N, encoder_dim) + decoder_out = torch.rand(N, decoder_dim) + + projected_encoder_out = torch_model.joiner.encoder_proj(encoder_out) + projected_decoder_out = torch_model.joiner.decoder_proj(decoder_out) + + torch_joiner_out = torch_model.joiner(encoder_out, decoder_out) + onnx_joiner_out = onnx_model.run_joiner( + projected_encoder_out, projected_decoder_out + ) + + assert torch.allclose(torch_joiner_out, onnx_joiner_out, atol=1e-4), ( + (torch_joiner_out - onnx_joiner_out).abs().max() + ) + + +@torch.no_grad() +def main(): + args = get_parser().parse_args() + logging.info(vars(args)) + + torch_model = torch.jit.load(args.jit_filename) + + onnx_model = OnnxModel( + encoder_model_filename=args.onnx_encoder_filename, + decoder_model_filename=args.onnx_decoder_filename, + joiner_model_filename=args.onnx_joiner_filename, + ) + + logging.info("Test encoder") + test_encoder(torch_model, onnx_model) + + logging.info("Test decoder") + test_decoder(torch_model, onnx_model) + + logging.info("Test joiner") + test_joiner(torch_model, onnx_model) + logging.info("Finished checking ONNX models") + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +# See https://github.com/pytorch/pytorch/issues/38342 +# and https://github.com/pytorch/pytorch/issues/33354 +# +# If we don't do this, the delay increases whenever there is +# a new request that changes the actual batch size. +# If you use `py-spy dump --pid --native`, you will +# see a lot of time is spent in re-compiling the torch script model. +torch._C._jit_set_profiling_executor(False) +torch._C._jit_set_profiling_mode(False) +torch._C._set_graph_executor_optimize(False) +if __name__ == "__main__": + torch.manual_seed(20220727) + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/ami/ASR/zipformer/onnx_decode.py b/egs/ami/ASR/zipformer/onnx_decode.py new file mode 100755 index 000000000..356c2a830 --- /dev/null +++ b/egs/ami/ASR/zipformer/onnx_decode.py @@ -0,0 +1,325 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao, +# Xiaoyu Yang) +# +# 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 loads ONNX exported models and uses them to decode the test sets. + +We use the pre-trained model from +https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +as an example to show how to use this file. + +1. Download the pre-trained model + +cd egs/librispeech/ASR + +repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url +repo=$(basename $repo_url) + +pushd $repo +git lfs pull --include "data/lang_bpe_500/bpe.model" +git lfs pull --include "exp/pretrained.pt" + +cd exp +ln -s pretrained.pt epoch-99.pt +popd + +2. Export the model to ONNX + +./zipformer/export-onnx.py \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ + --use-averaged-model 0 \ + --epoch 99 \ + --avg 1 \ + --exp-dir $repo/exp \ + --causal False + +It will generate the following 3 files inside $repo/exp: + + - encoder-epoch-99-avg-1.onnx + - decoder-epoch-99-avg-1.onnx + - joiner-epoch-99-avg-1.onnx + +2. Run this file + +./zipformer/onnx_decode.py \ + --exp-dir $repo/exp \ + --max-duration 600 \ + --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \ + --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \ + --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ +""" + + +import argparse +import logging +import time +from pathlib import Path +from typing import List, Tuple + +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule + +from onnx_pretrained import greedy_search, OnnxModel + +from icefall.utils import setup_logger, store_transcripts, write_error_stats +from k2 import SymbolTable + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--encoder-model-filename", + type=str, + required=True, + help="Path to the encoder onnx model. ", + ) + + parser.add_argument( + "--decoder-model-filename", + type=str, + required=True, + help="Path to the decoder onnx model. ", + ) + + parser.add_argument( + "--joiner-model-filename", + type=str, + required=True, + help="Path to the joiner onnx model. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--tokens", + type=str, + help="""Path to tokens.txt.""", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="Valid values are greedy_search and modified_beam_search", + ) + + return parser + + +def decode_one_batch( + model: OnnxModel, token_table: SymbolTable, batch: dict +) -> List[List[str]]: + """Decode one batch and return the result. + Currently it only greedy_search is supported. + + Args: + model: + The neural model. + token_table: + The token table. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + + Returns: + Return the decoded results for each utterance. + """ + feature = batch["inputs"] + assert feature.ndim == 3 + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(dtype=torch.int64) + + encoder_out, encoder_out_lens = model.run_encoder(x=feature, x_lens=feature_lens) + + hyps = greedy_search( + model=model, encoder_out=encoder_out, encoder_out_lens=encoder_out_lens + ) + + def token_ids_to_words(token_ids: List[int]) -> str: + text = "" + for i in token_ids: + text += token_table[i] + return text.replace("▁", " ").strip() + + hyps = [token_ids_to_words(h).split() for h in hyps] + return hyps + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + model: nn.Module, + token_table: SymbolTable, +) -> Tuple[List[Tuple[str, List[str], List[str]]], float]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + model: + The neural model. + token_table: + The token table. + + Returns: + - A list of tuples. Each tuple contains three elements: + - cut_id, + - reference transcript, + - predicted result. + - The total duration (in seconds) of the dataset. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + log_interval = 10 + total_duration = 0 + + results = [] + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + total_duration += sum([cut.duration for cut in batch["supervisions"]["cut"]]) + + hyps = decode_one_batch(model=model, token_table=token_table, batch=batch) + + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results.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, total_duration + + +def save_results( + res_dir: Path, + test_set_name: str, + results: List[Tuple[str, List[str], List[str]]], +): + recog_path = res_dir / f"recogs-{test_set_name}.txt" + 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 = res_dir / f"errs-{test_set_name}.txt" + with open(errs_filename, "w") as f: + wer = write_error_stats(f, f"{test_set_name}", results, enable_log=True) + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + errs_info = res_dir / f"wer-summary-{test_set_name}.txt" + with open(errs_info, "w") as f: + print("WER", file=f) + print(wer, file=f) + + s = "\nFor {}, WER is {}:\n".format(test_set_name, wer) + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + + assert ( + args.decoding_method == "greedy_search" + ), "Only supports greedy_search currently." + res_dir = Path(args.exp_dir) / f"onnx-{args.decoding_method}" + + setup_logger(f"{res_dir}/log-decode") + logging.info("Decoding started") + + device = torch.device("cpu") + logging.info(f"Device: {device}") + + token_table = SymbolTable.from_file(args.tokens) + + logging.info(vars(args)) + + logging.info("About to create model") + model = OnnxModel( + encoder_model_filename=args.encoder_model_filename, + decoder_model_filename=args.decoder_model_filename, + joiner_model_filename=args.joiner_model_filename, + ) + + # we need cut ids to display recognition results. + args.return_cuts = True + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_other_dl = librispeech.test_dataloaders(test_other_cuts) + + test_sets = ["test-clean", "test-other"] + test_dl = [test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + start_time = time.time() + results, total_duration = decode_dataset( + dl=test_dl, model=model, token_table=token_table + ) + end_time = time.time() + elapsed_seconds = end_time - start_time + rtf = elapsed_seconds / total_duration + + logging.info(f"Elapsed time: {elapsed_seconds:.3f} s") + logging.info(f"Wave duration: {total_duration:.3f} s") + logging.info( + f"Real time factor (RTF): {elapsed_seconds:.3f}/{total_duration:.3f} = {rtf:.3f}" + ) + + save_results(res_dir=res_dir, test_set_name=test_set, results=results) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/ami/ASR/zipformer/onnx_pretrained-streaming.py b/egs/ami/ASR/zipformer/onnx_pretrained-streaming.py new file mode 100755 index 000000000..e62491444 --- /dev/null +++ b/egs/ami/ASR/zipformer/onnx_pretrained-streaming.py @@ -0,0 +1,546 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Fangjun Kuang) +# Copyright 2023 Danqing Fu (danqing.fu@gmail.com) + +""" +This script loads ONNX models exported by ./export-onnx-streaming.py +and uses them to decode waves. + +We use the pre-trained model from +https://huggingface.co/Zengwei/icefall-asr-librispeech-streaming-zipformer-2023-05-17 +as an example to show how to use this file. + +1. Download the pre-trained model + +cd egs/librispeech/ASR + +repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-streaming-zipformer-2023-05-17 +GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url +repo=$(basename $repo_url) + +pushd $repo +git lfs pull --include "exp/pretrained.pt" + +cd exp +ln -s pretrained.pt epoch-99.pt +popd + +2. Export the model to ONNX + +./zipformer/export-onnx-streaming.py \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ + --use-averaged-model 0 \ + --epoch 99 \ + --avg 1 \ + --exp-dir $repo/exp \ + --num-encoder-layers "2,2,3,4,3,2" \ + --downsampling-factor "1,2,4,8,4,2" \ + --feedforward-dim "512,768,1024,1536,1024,768" \ + --num-heads "4,4,4,8,4,4" \ + --encoder-dim "192,256,384,512,384,256" \ + --query-head-dim 32 \ + --value-head-dim 12 \ + --pos-head-dim 4 \ + --pos-dim 48 \ + --encoder-unmasked-dim "192,192,256,256,256,192" \ + --cnn-module-kernel "31,31,15,15,15,31" \ + --decoder-dim 512 \ + --joiner-dim 512 \ + --causal True \ + --chunk-size 16 \ + --left-context-frames 64 + +It will generate the following 3 files inside $repo/exp: + + - encoder-epoch-99-avg-1.onnx + - decoder-epoch-99-avg-1.onnx + - joiner-epoch-99-avg-1.onnx + +3. Run this file with the exported ONNX models + +./zipformer/onnx_pretrained-streaming.py \ + --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \ + --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \ + --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ + $repo/test_wavs/1089-134686-0001.wav + +Note: Even though this script only supports decoding a single file, +the exported ONNX models do support batch processing. +""" + +import argparse +import logging +from typing import Dict, List, Optional, Tuple + +import k2 +import numpy as np +import onnxruntime as ort +import torch +import torchaudio +from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--encoder-model-filename", + type=str, + required=True, + help="Path to the encoder onnx model. ", + ) + + parser.add_argument( + "--decoder-model-filename", + type=str, + required=True, + help="Path to the decoder onnx model. ", + ) + + parser.add_argument( + "--joiner-model-filename", + type=str, + required=True, + help="Path to the joiner onnx model. ", + ) + + parser.add_argument( + "--tokens", + type=str, + help="""Path to tokens.txt.""", + ) + + parser.add_argument( + "sound_file", + type=str, + help="The input sound file to transcribe. " + "Supported formats are those supported by torchaudio.load(). " + "For example, wav and flac are supported. " + "The sample rate has to be 16kHz.", + ) + + return parser + + +class OnnxModel: + def __init__( + self, + encoder_model_filename: str, + decoder_model_filename: str, + joiner_model_filename: str, + ): + session_opts = ort.SessionOptions() + session_opts.inter_op_num_threads = 1 + session_opts.intra_op_num_threads = 1 + + self.session_opts = session_opts + + self.init_encoder(encoder_model_filename) + self.init_decoder(decoder_model_filename) + self.init_joiner(joiner_model_filename) + + def init_encoder(self, encoder_model_filename: str): + self.encoder = ort.InferenceSession( + encoder_model_filename, + sess_options=self.session_opts, + providers=["CPUExecutionProvider"], + ) + self.init_encoder_states() + + def init_encoder_states(self, batch_size: int = 1): + encoder_meta = self.encoder.get_modelmeta().custom_metadata_map + logging.info(f"encoder_meta={encoder_meta}") + + model_type = encoder_meta["model_type"] + assert model_type == "zipformer2", model_type + + decode_chunk_len = int(encoder_meta["decode_chunk_len"]) + T = int(encoder_meta["T"]) + + num_encoder_layers = encoder_meta["num_encoder_layers"] + encoder_dims = encoder_meta["encoder_dims"] + cnn_module_kernels = encoder_meta["cnn_module_kernels"] + left_context_len = encoder_meta["left_context_len"] + query_head_dims = encoder_meta["query_head_dims"] + value_head_dims = encoder_meta["value_head_dims"] + num_heads = encoder_meta["num_heads"] + + def to_int_list(s): + return list(map(int, s.split(","))) + + num_encoder_layers = to_int_list(num_encoder_layers) + encoder_dims = to_int_list(encoder_dims) + cnn_module_kernels = to_int_list(cnn_module_kernels) + left_context_len = to_int_list(left_context_len) + query_head_dims = to_int_list(query_head_dims) + value_head_dims = to_int_list(value_head_dims) + num_heads = to_int_list(num_heads) + + logging.info(f"decode_chunk_len: {decode_chunk_len}") + logging.info(f"T: {T}") + logging.info(f"num_encoder_layers: {num_encoder_layers}") + logging.info(f"encoder_dims: {encoder_dims}") + logging.info(f"cnn_module_kernels: {cnn_module_kernels}") + logging.info(f"left_context_len: {left_context_len}") + logging.info(f"query_head_dims: {query_head_dims}") + logging.info(f"value_head_dims: {value_head_dims}") + logging.info(f"num_heads: {num_heads}") + + num_encoders = len(num_encoder_layers) + + self.states = [] + for i in range(num_encoders): + num_layers = num_encoder_layers[i] + key_dim = query_head_dims[i] * num_heads[i] + embed_dim = encoder_dims[i] + nonlin_attn_head_dim = 3 * embed_dim // 4 + value_dim = value_head_dims[i] * num_heads[i] + conv_left_pad = cnn_module_kernels[i] // 2 + + for layer in range(num_layers): + cached_key = torch.zeros( + left_context_len[i], batch_size, key_dim + ).numpy() + cached_nonlin_attn = torch.zeros( + 1, batch_size, left_context_len[i], nonlin_attn_head_dim + ).numpy() + cached_val1 = torch.zeros( + left_context_len[i], batch_size, value_dim + ).numpy() + cached_val2 = torch.zeros( + left_context_len[i], batch_size, value_dim + ).numpy() + cached_conv1 = torch.zeros(batch_size, embed_dim, conv_left_pad).numpy() + cached_conv2 = torch.zeros(batch_size, embed_dim, conv_left_pad).numpy() + self.states += [ + cached_key, + cached_nonlin_attn, + cached_val1, + cached_val2, + cached_conv1, + cached_conv2, + ] + embed_states = torch.zeros(batch_size, 128, 3, 19).numpy() + self.states.append(embed_states) + processed_lens = torch.zeros(batch_size, dtype=torch.int64).numpy() + self.states.append(processed_lens) + + self.num_encoders = num_encoders + + self.segment = T + self.offset = decode_chunk_len + + def init_decoder(self, decoder_model_filename: str): + self.decoder = ort.InferenceSession( + decoder_model_filename, + sess_options=self.session_opts, + providers=["CPUExecutionProvider"], + ) + + decoder_meta = self.decoder.get_modelmeta().custom_metadata_map + self.context_size = int(decoder_meta["context_size"]) + self.vocab_size = int(decoder_meta["vocab_size"]) + + logging.info(f"context_size: {self.context_size}") + logging.info(f"vocab_size: {self.vocab_size}") + + def init_joiner(self, joiner_model_filename: str): + self.joiner = ort.InferenceSession( + joiner_model_filename, + sess_options=self.session_opts, + providers=["CPUExecutionProvider"], + ) + + joiner_meta = self.joiner.get_modelmeta().custom_metadata_map + self.joiner_dim = int(joiner_meta["joiner_dim"]) + + logging.info(f"joiner_dim: {self.joiner_dim}") + + def _build_encoder_input_output( + self, + x: torch.Tensor, + ) -> Tuple[Dict[str, np.ndarray], List[str]]: + encoder_input = {"x": x.numpy()} + encoder_output = ["encoder_out"] + + def build_inputs_outputs(tensors, i): + assert len(tensors) == 6, len(tensors) + + # (downsample_left, batch_size, key_dim) + name = f"cached_key_{i}" + encoder_input[name] = tensors[0] + encoder_output.append(f"new_{name}") + + # (1, batch_size, downsample_left, nonlin_attn_head_dim) + name = f"cached_nonlin_attn_{i}" + encoder_input[name] = tensors[1] + encoder_output.append(f"new_{name}") + + # (downsample_left, batch_size, value_dim) + name = f"cached_val1_{i}" + encoder_input[name] = tensors[2] + encoder_output.append(f"new_{name}") + + # (downsample_left, batch_size, value_dim) + name = f"cached_val2_{i}" + encoder_input[name] = tensors[3] + encoder_output.append(f"new_{name}") + + # (batch_size, embed_dim, conv_left_pad) + name = f"cached_conv1_{i}" + encoder_input[name] = tensors[4] + encoder_output.append(f"new_{name}") + + # (batch_size, embed_dim, conv_left_pad) + name = f"cached_conv2_{i}" + encoder_input[name] = tensors[5] + encoder_output.append(f"new_{name}") + + for i in range(len(self.states[:-2]) // 6): + build_inputs_outputs(self.states[i * 6 : (i + 1) * 6], i) + + # (batch_size, channels, left_pad, freq) + name = "embed_states" + embed_states = self.states[-2] + encoder_input[name] = embed_states + encoder_output.append(f"new_{name}") + + # (batch_size,) + name = "processed_lens" + processed_lens = self.states[-1] + encoder_input[name] = processed_lens + encoder_output.append(f"new_{name}") + + return encoder_input, encoder_output + + def _update_states(self, states: List[np.ndarray]): + self.states = states + + def run_encoder(self, x: torch.Tensor) -> torch.Tensor: + """ + Args: + x: + A 3-D tensor of shape (N, T, C) + Returns: + Return a 3-D tensor of shape (N, T', joiner_dim) where + T' is usually equal to ((T-7)//2+1)//2 + """ + encoder_input, encoder_output_names = self._build_encoder_input_output(x) + + out = self.encoder.run(encoder_output_names, encoder_input) + + self._update_states(out[1:]) + + return torch.from_numpy(out[0]) + + def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor: + """ + Args: + decoder_input: + A 2-D tensor of shape (N, context_size) + Returns: + Return a 2-D tensor of shape (N, joiner_dim) + """ + out = self.decoder.run( + [self.decoder.get_outputs()[0].name], + {self.decoder.get_inputs()[0].name: decoder_input.numpy()}, + )[0] + + return torch.from_numpy(out) + + def run_joiner( + self, encoder_out: torch.Tensor, decoder_out: torch.Tensor + ) -> torch.Tensor: + """ + Args: + encoder_out: + A 2-D tensor of shape (N, joiner_dim) + decoder_out: + A 2-D tensor of shape (N, joiner_dim) + Returns: + Return a 2-D tensor of shape (N, vocab_size) + """ + out = self.joiner.run( + [self.joiner.get_outputs()[0].name], + { + self.joiner.get_inputs()[0].name: encoder_out.numpy(), + self.joiner.get_inputs()[1].name: decoder_out.numpy(), + }, + )[0] + + return torch.from_numpy(out) + + +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}. Given: {sample_rate}" + # We use only the first channel + ans.append(wave[0].contiguous()) + return ans + + +def create_streaming_feature_extractor() -> OnlineFeature: + """Create a CPU streaming feature extractor. + + At present, we assume it returns a fbank feature extractor with + fixed options. In the future, we will support passing in the options + from outside. + + Returns: + Return a CPU streaming feature extractor. + """ + opts = FbankOptions() + opts.device = "cpu" + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = 16000 + opts.mel_opts.num_bins = 80 + return OnlineFbank(opts) + + +def greedy_search( + model: OnnxModel, + encoder_out: torch.Tensor, + context_size: int, + decoder_out: Optional[torch.Tensor] = None, + hyp: Optional[List[int]] = None, +) -> List[int]: + """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1. + Args: + model: + The transducer model. + encoder_out: + A 3-D tensor of shape (1, T, joiner_dim) + context_size: + The context size of the decoder model. + decoder_out: + Optional. Decoder output of the previous chunk. + hyp: + Decoding results for previous chunks. + Returns: + Return the decoded results so far. + """ + + blank_id = 0 + + if decoder_out is None: + assert hyp is None, hyp + hyp = [blank_id] * context_size + decoder_input = torch.tensor([hyp], dtype=torch.int64) + decoder_out = model.run_decoder(decoder_input) + else: + assert hyp is not None, hyp + + encoder_out = encoder_out.squeeze(0) + T = encoder_out.size(0) + for t in range(T): + cur_encoder_out = encoder_out[t : t + 1] + joiner_out = model.run_joiner(cur_encoder_out, decoder_out).squeeze(0) + y = joiner_out.argmax(dim=0).item() + if y != blank_id: + hyp.append(y) + decoder_input = hyp[-context_size:] + decoder_input = torch.tensor([decoder_input], dtype=torch.int64) + decoder_out = model.run_decoder(decoder_input) + + return hyp, decoder_out + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + logging.info(vars(args)) + + model = OnnxModel( + encoder_model_filename=args.encoder_model_filename, + decoder_model_filename=args.decoder_model_filename, + joiner_model_filename=args.joiner_model_filename, + ) + + sample_rate = 16000 + + logging.info("Constructing Fbank computer") + online_fbank = create_streaming_feature_extractor() + + logging.info(f"Reading sound files: {args.sound_file}") + waves = read_sound_files( + filenames=[args.sound_file], + expected_sample_rate=sample_rate, + )[0] + + tail_padding = torch.zeros(int(0.3 * sample_rate), dtype=torch.float32) + wave_samples = torch.cat([waves, tail_padding]) + + num_processed_frames = 0 + segment = model.segment + offset = model.offset + + context_size = model.context_size + hyp = None + decoder_out = None + + chunk = int(1 * sample_rate) # 1 second + start = 0 + while start < wave_samples.numel(): + end = min(start + chunk, wave_samples.numel()) + samples = wave_samples[start:end] + start += chunk + + online_fbank.accept_waveform( + sampling_rate=sample_rate, + waveform=samples, + ) + + while online_fbank.num_frames_ready - num_processed_frames >= segment: + frames = [] + for i in range(segment): + frames.append(online_fbank.get_frame(num_processed_frames + i)) + num_processed_frames += offset + frames = torch.cat(frames, dim=0) + frames = frames.unsqueeze(0) + encoder_out = model.run_encoder(frames) + hyp, decoder_out = greedy_search( + model, + encoder_out, + context_size, + decoder_out, + hyp, + ) + + token_table = k2.SymbolTable.from_file(args.tokens) + + text = "" + for i in hyp[context_size:]: + text += token_table[i] + text = text.replace("▁", " ").strip() + + logging.info(args.sound_file) + logging.info(text) + + 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/ami/ASR/zipformer/onnx_pretrained.py b/egs/ami/ASR/zipformer/onnx_pretrained.py new file mode 100755 index 000000000..334376093 --- /dev/null +++ b/egs/ami/ASR/zipformer/onnx_pretrained.py @@ -0,0 +1,421 @@ +#!/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. +""" +This script loads ONNX models and uses them to decode waves. +You can use the following command to get the exported models: + +We use the pre-trained model from +https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +as an example to show how to use this file. + +1. Download the pre-trained model + +cd egs/librispeech/ASR + +repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url +repo=$(basename $repo_url) + +pushd $repo +git lfs pull --include "exp/pretrained.pt" + +cd exp +ln -s pretrained.pt epoch-99.pt +popd + +2. Export the model to ONNX + +./zipformer/export-onnx.py \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ + --use-averaged-model 0 \ + --epoch 99 \ + --avg 1 \ + --exp-dir $repo/exp \ + --causal False + +It will generate the following 3 files inside $repo/exp: + + - encoder-epoch-99-avg-1.onnx + - decoder-epoch-99-avg-1.onnx + - joiner-epoch-99-avg-1.onnx + +3. Run this file + +./zipformer/onnx_pretrained.py \ + --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \ + --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \ + --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ + $repo/test_wavs/1089-134686-0001.wav \ + $repo/test_wavs/1221-135766-0001.wav \ + $repo/test_wavs/1221-135766-0002.wav +""" + +import argparse +import logging +import math +from typing import List, Tuple + +import k2 +import kaldifeat +import onnxruntime as ort +import torch +import torchaudio +from torch.nn.utils.rnn import pad_sequence + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--encoder-model-filename", + type=str, + required=True, + help="Path to the encoder onnx model. ", + ) + + parser.add_argument( + "--decoder-model-filename", + type=str, + required=True, + help="Path to the decoder onnx model. ", + ) + + parser.add_argument( + "--joiner-model-filename", + type=str, + required=True, + help="Path to the joiner onnx model. ", + ) + + parser.add_argument( + "--tokens", + type=str, + help="""Path to tokens.txt.""", + ) + + 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", + ) + + return parser + + +class OnnxModel: + def __init__( + self, + encoder_model_filename: str, + decoder_model_filename: str, + joiner_model_filename: str, + ): + session_opts = ort.SessionOptions() + session_opts.inter_op_num_threads = 1 + session_opts.intra_op_num_threads = 4 + + self.session_opts = session_opts + + self.init_encoder(encoder_model_filename) + self.init_decoder(decoder_model_filename) + self.init_joiner(joiner_model_filename) + + def init_encoder(self, encoder_model_filename: str): + self.encoder = ort.InferenceSession( + encoder_model_filename, + sess_options=self.session_opts, + providers=["CPUExecutionProvider"], + ) + + def init_decoder(self, decoder_model_filename: str): + self.decoder = ort.InferenceSession( + decoder_model_filename, + sess_options=self.session_opts, + providers=["CPUExecutionProvider"], + ) + + decoder_meta = self.decoder.get_modelmeta().custom_metadata_map + self.context_size = int(decoder_meta["context_size"]) + self.vocab_size = int(decoder_meta["vocab_size"]) + + logging.info(f"context_size: {self.context_size}") + logging.info(f"vocab_size: {self.vocab_size}") + + def init_joiner(self, joiner_model_filename: str): + self.joiner = ort.InferenceSession( + joiner_model_filename, + sess_options=self.session_opts, + providers=["CPUExecutionProvider"], + ) + + joiner_meta = self.joiner.get_modelmeta().custom_metadata_map + self.joiner_dim = int(joiner_meta["joiner_dim"]) + + logging.info(f"joiner_dim: {self.joiner_dim}") + + def run_encoder( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Args: + x: + A 3-D tensor of shape (N, T, C) + x_lens: + A 2-D tensor of shape (N,). Its dtype is torch.int64 + Returns: + Return a tuple containing: + - encoder_out, its shape is (N, T', joiner_dim) + - encoder_out_lens, its shape is (N,) + """ + out = self.encoder.run( + [ + self.encoder.get_outputs()[0].name, + self.encoder.get_outputs()[1].name, + ], + { + self.encoder.get_inputs()[0].name: x.numpy(), + self.encoder.get_inputs()[1].name: x_lens.numpy(), + }, + ) + return torch.from_numpy(out[0]), torch.from_numpy(out[1]) + + def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor: + """ + Args: + decoder_input: + A 2-D tensor of shape (N, context_size) + Returns: + Return a 2-D tensor of shape (N, joiner_dim) + """ + out = self.decoder.run( + [self.decoder.get_outputs()[0].name], + {self.decoder.get_inputs()[0].name: decoder_input.numpy()}, + )[0] + + return torch.from_numpy(out) + + def run_joiner( + self, encoder_out: torch.Tensor, decoder_out: torch.Tensor + ) -> torch.Tensor: + """ + Args: + encoder_out: + A 2-D tensor of shape (N, joiner_dim) + decoder_out: + A 2-D tensor of shape (N, joiner_dim) + Returns: + Return a 2-D tensor of shape (N, vocab_size) + """ + out = self.joiner.run( + [self.joiner.get_outputs()[0].name], + { + self.joiner.get_inputs()[0].name: encoder_out.numpy(), + self.joiner.get_inputs()[1].name: decoder_out.numpy(), + }, + )[0] + + return torch.from_numpy(out) + + +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}. Given: {sample_rate}" + # We use only the first channel + ans.append(wave[0]) + return ans + + +def greedy_search( + model: OnnxModel, + 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: + A 3-D tensor of shape (N, T, joiner_dim) + encoder_out_lens: + A 1-D tensor of shape (N,). + Returns: + Return the decoded results for each 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 = 0 # hard-code to 0 + + 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) + + context_size = model.context_size + hyps = [[blank_id] * context_size for _ in range(N)] + + decoder_input = torch.tensor( + hyps, + dtype=torch.int64, + ) # (N, context_size) + + decoder_out = model.run_decoder(decoder_input) + + offset = 0 + for batch_size in batch_size_list: + start = offset + end = offset + batch_size + current_encoder_out = packed_encoder_out.data[start:end] + # current_encoder_out's shape: (batch_size, joiner_dim) + offset = end + + decoder_out = decoder_out[:batch_size] + logits = model.run_joiner(current_encoder_out, decoder_out) + + # logits'shape (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 != blank_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, + dtype=torch.int64, + ) + decoder_out = model.run_decoder(decoder_input) + + 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 + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + logging.info(vars(args)) + model = OnnxModel( + encoder_model_filename=args.encoder_model_filename, + decoder_model_filename=args.decoder_model_filename, + joiner_model_filename=args.joiner_model_filename, + ) + + logging.info("Constructing Fbank computer") + opts = kaldifeat.FbankOptions() + opts.device = "cpu" + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = args.sample_rate + opts.mel_opts.num_bins = 80 + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {args.sound_files}") + waves = read_sound_files( + filenames=args.sound_files, + expected_sample_rate=args.sample_rate, + ) + + 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, dtype=torch.int64) + encoder_out, encoder_out_lens = model.run_encoder(features, feature_lengths) + + hyps = greedy_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + s = "\n" + + token_table = k2.SymbolTable.from_file(args.tokens) + + def token_ids_to_words(token_ids: List[int]) -> str: + text = "" + for i in token_ids: + text += token_table[i] + return text.replace("▁", " ").strip() + + for filename, hyp in zip(args.sound_files, hyps): + words = token_ids_to_words(hyp) + s += f"{filename}:\n{words}\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/ami/ASR/zipformer/optim.py b/egs/ami/ASR/zipformer/optim.py new file mode 120000 index 000000000..5eaa3cffd --- /dev/null +++ b/egs/ami/ASR/zipformer/optim.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/optim.py \ No newline at end of file diff --git a/egs/ami/ASR/zipformer/pretrained.py b/egs/ami/ASR/zipformer/pretrained.py new file mode 100755 index 000000000..3104b6084 --- /dev/null +++ b/egs/ami/ASR/zipformer/pretrained.py @@ -0,0 +1,381 @@ +#!/usr/bin/env python3 +# Copyright 2021-2023 Xiaomi Corp. (authors: Fangjun Kuang, Zengwei Yao) +# +# 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 loads a checkpoint and uses it to decode waves. +You can generate the checkpoint with the following command: + +Note: This is a example for librispeech dataset, if you are using different +dataset, you should change the argument values according to your dataset. + +- For non-streaming model: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --tokens data/lang_bpe_500/tokens.txt \ + --epoch 30 \ + --avg 9 + +- For streaming model: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --causal 1 \ + --tokens data/lang_bpe_500/tokens.txt \ + --epoch 30 \ + --avg 9 + +Usage of this script: + +- For non-streaming model: + +(1) greedy search +./zipformer/pretrained.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --tokens data/lang_bpe_500/tokens.txt \ + --method greedy_search \ + /path/to/foo.wav \ + /path/to/bar.wav + +(2) modified beam search +./zipformer/pretrained.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --tokens ./data/lang_bpe_500/tokens.txt \ + --method modified_beam_search \ + /path/to/foo.wav \ + /path/to/bar.wav + +(3) fast beam search +./zipformer/pretrained.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --tokens ./data/lang_bpe_500/tokens.txt \ + --method fast_beam_search \ + /path/to/foo.wav \ + /path/to/bar.wav + +- For streaming model: + +(1) greedy search +./zipformer/pretrained.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --causal 1 \ + --chunk-size 16 \ + --left-context-frames 128 \ + --tokens ./data/lang_bpe_500/tokens.txt \ + --method greedy_search \ + /path/to/foo.wav \ + /path/to/bar.wav + +(2) modified beam search +./zipformer/pretrained.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --causal 1 \ + --chunk-size 16 \ + --left-context-frames 128 \ + --tokens ./data/lang_bpe_500/tokens.txt \ + --method modified_beam_search \ + /path/to/foo.wav \ + /path/to/bar.wav + +(3) fast beam search +./zipformer/pretrained.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --causal 1 \ + --chunk-size 16 \ + --left-context-frames 128 \ + --tokens ./data/lang_bpe_500/tokens.txt \ + --method fast_beam_search \ + /path/to/foo.wav \ + /path/to/bar.wav + + +You can also use `./zipformer/exp/epoch-xx.pt`. + +Note: ./zipformer/exp/pretrained.pt is generated by ./zipformer/export.py +""" + + +import argparse +import logging +import math +from typing import List + +import k2 +import kaldifeat +import torch +import torchaudio +from beam_search import ( + fast_beam_search_one_best, + greedy_search_batch, + modified_beam_search, +) +from export import num_tokens +from torch.nn.utils.rnn import pad_sequence +from train import add_model_arguments, get_model, get_params + +from icefall.utils import make_pad_mask + + +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( + "--tokens", + type=str, + help="""Path to tokens.txt.""", + ) + + parser.add_argument( + "--method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_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. + """, + ) + + 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}. Given: {sample_rate}" + # We use only the first channel + ans.append(wave[0].contiguous()) + return ans + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + + params = get_params() + + params.update(vars(args)) + + token_table = k2.SymbolTable.from_file(params.tokens) + + params.blank_id = token_table[""] + params.unk_id = token_table[""] + params.vocab_size = num_tokens(token_table) + 1 + + logging.info(f"{params}") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + if params.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + + logging.info("Creating model") + model = get_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() + + 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) + + # model forward + encoder_out, encoder_out_lens = model.forward_encoder(features, feature_lengths) + + hyps = [] + msg = f"Using {params.method}" + logging.info(msg) + + def token_ids_to_words(token_ids: List[int]) -> str: + text = "" + for i in token_ids: + text += token_table[i] + return text.replace("▁", " ").strip() + + 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 hyp_tokens: + hyps.append(token_ids_to_words(hyp)) + 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 hyp_tokens: + hyps.append(token_ids_to_words(hyp)) + 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 hyp_tokens: + hyps.append(token_ids_to_words(hyp)) + else: + raise ValueError(f"Unsupported method: {params.method}") + + s = "\n" + for filename, hyp in zip(params.sound_files, hyps): + s += f"{filename}:\n{hyp}\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/ami/ASR/zipformer/pretrained_ctc.py b/egs/ami/ASR/zipformer/pretrained_ctc.py new file mode 100755 index 000000000..9dff2e6fc --- /dev/null +++ b/egs/ami/ASR/zipformer/pretrained_ctc.py @@ -0,0 +1,455 @@ +#!/usr/bin/env python3 +# Copyright 2022-2023 Xiaomi Corp. (authors: Fangjun Kuang, +# Zengwei Yao) +# +# 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 loads a checkpoint and uses it to decode waves. +You can generate the checkpoint with the following command: + +- For non-streaming model: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --use-ctc 1 \ + --tokens data/lang_bpe_500/tokens.txt \ + --epoch 30 \ + --avg 9 + +- For streaming model: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --use-ctc 1 \ + --causal 1 \ + --tokens data/lang_bpe_500/tokens.txt \ + --epoch 30 \ + --avg 9 + +Usage of this script: + +(1) ctc-decoding +./zipformer/pretrained_ctc.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --tokens data/lang_bpe_500/tokens.txt \ + --method ctc-decoding \ + --sample-rate 16000 \ + /path/to/foo.wav \ + /path/to/bar.wav + +(2) 1best +./zipformer/pretrained_ctc.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --HLG data/lang_bpe_500/HLG.pt \ + --words-file data/lang_bpe_500/words.txt \ + --method 1best \ + --sample-rate 16000 \ + /path/to/foo.wav \ + /path/to/bar.wav + +(3) nbest-rescoring +./zipformer/pretrained_ctc.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --HLG data/lang_bpe_500/HLG.pt \ + --words-file data/lang_bpe_500/words.txt \ + --G data/lm/G_4_gram.pt \ + --method nbest-rescoring \ + --sample-rate 16000 \ + /path/to/foo.wav \ + /path/to/bar.wav + + +(4) whole-lattice-rescoring +./zipformer/pretrained_ctc.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --HLG data/lang_bpe_500/HLG.pt \ + --words-file data/lang_bpe_500/words.txt \ + --G data/lm/G_4_gram.pt \ + --method whole-lattice-rescoring \ + --sample-rate 16000 \ + /path/to/foo.wav \ + /path/to/bar.wav +""" + +import argparse +import logging +import math +from typing import List + +import k2 +import kaldifeat +import torch +import torchaudio +from ctc_decode import get_decoding_params +from export import num_tokens +from torch.nn.utils.rnn import pad_sequence +from train import add_model_arguments, get_model, get_params + +from icefall.decode import ( + get_lattice, + one_best_decoding, + rescore_with_n_best_list, + rescore_with_whole_lattice, +) +from icefall.utils import get_texts + + +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( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + + parser.add_argument( + "--words-file", + type=str, + help="""Path to words.txt. + Used only when method is not ctc-decoding. + """, + ) + + parser.add_argument( + "--HLG", + type=str, + help="""Path to HLG.pt. + Used only when method is not ctc-decoding. + """, + ) + + parser.add_argument( + "--tokens", + type=str, + help="""Path to tokens.txt. + Used only when method is ctc-decoding. + """, + ) + + parser.add_argument( + "--method", + type=str, + default="1best", + help="""Decoding method. + Possible values are: + (0) ctc-decoding - Use CTC decoding. It uses a token table, + i.e., lang_dir/tokens.txt, to convert + word pieces to words. It needs neither a lexicon + nor an n-gram LM. + (1) 1best - Use the best path as decoding output. Only + the transformer encoder output is used for decoding. + We call it HLG decoding. + (2) nbest-rescoring. Extract n paths from the decoding lattice, + rescore them with an LM, the path with + the highest score is the decoding result. + We call it HLG decoding + nbest n-gram LM rescoring. + (3) whole-lattice-rescoring - Use an LM to rescore the + decoding lattice and then use 1best to decode the + rescored lattice. + We call it HLG decoding + whole-lattice n-gram LM rescoring. + """, + ) + + parser.add_argument( + "--G", + type=str, + help="""An LM for rescoring. + Used only when method is + whole-lattice-rescoring or nbest-rescoring. + It's usually a 4-gram LM. + """, + ) + + parser.add_argument( + "--num-paths", + type=int, + default=100, + help=""" + Used only when method is attention-decoder. + It specifies the size of n-best list.""", + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=1.3, + help=""" + Used only when method is whole-lattice-rescoring and nbest-rescoring. + It specifies the scale for n-gram LM scores. + (Note: You need to tune it on a dataset.) + """, + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=1.0, + help=""" + Used only when method is nbest-rescoring. + It specifies the scale for lattice.scores when + extracting n-best lists. A smaller value results in + more unique number of paths with the risk of missing + the best path. + """, + ) + + parser.add_argument( + "--sample-rate", + type=int, + default=16000, + help="The sample rate of the input sound file", + ) + + 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.", + ) + + add_model_arguments(parser) + + return parser + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float = 16000 +) -> 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].contiguous()) + return ans + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + + params = get_params() + # add decoding params + params.update(get_decoding_params()) + params.update(vars(args)) + + token_table = k2.SymbolTable.from_file(params.tokens) + params.vocab_size = num_tokens(token_table) + 1 # +1 for blank + params.blank_id = token_table[""] + assert params.blank_id == 0 + + 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_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() + + 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) + + encoder_out, encoder_out_lens = model.forward_encoder(features, feature_lengths) + ctc_output = model.ctc_output(encoder_out) # (N, T, C) + + batch_size = ctc_output.shape[0] + supervision_segments = torch.tensor( + [ + [i, 0, feature_lengths[i].item() // params.subsampling_factor] + for i in range(batch_size) + ], + dtype=torch.int32, + ) + + if params.method == "ctc-decoding": + logging.info("Use CTC decoding") + max_token_id = params.vocab_size - 1 + + H = k2.ctc_topo( + max_token=max_token_id, + modified=False, + device=device, + ) + + lattice = get_lattice( + nnet_output=ctc_output, + decoding_graph=H, + supervision_segments=supervision_segments, + search_beam=params.search_beam, + output_beam=params.output_beam, + min_active_states=params.min_active_states, + max_active_states=params.max_active_states, + subsampling_factor=params.subsampling_factor, + ) + + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + token_ids = get_texts(best_path) + hyps = [[token_table[i] for i in ids] for ids in token_ids] + elif params.method in [ + "1best", + "nbest-rescoring", + "whole-lattice-rescoring", + ]: + logging.info(f"Loading HLG from {params.HLG}") + HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu")) + HLG = HLG.to(device) + if not hasattr(HLG, "lm_scores"): + # For whole-lattice-rescoring and attention-decoder + HLG.lm_scores = HLG.scores.clone() + + if params.method in [ + "nbest-rescoring", + "whole-lattice-rescoring", + ]: + logging.info(f"Loading G from {params.G}") + G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu")) + G = G.to(device) + if params.method == "whole-lattice-rescoring": + # Add epsilon self-loops to G as we will compose + # it with the whole lattice later + G = k2.add_epsilon_self_loops(G) + G = k2.arc_sort(G) + + # G.lm_scores is used to replace HLG.lm_scores during + # LM rescoring. + G.lm_scores = G.scores.clone() + + lattice = get_lattice( + nnet_output=ctc_output, + decoding_graph=HLG, + supervision_segments=supervision_segments, + search_beam=params.search_beam, + output_beam=params.output_beam, + min_active_states=params.min_active_states, + max_active_states=params.max_active_states, + subsampling_factor=params.subsampling_factor, + ) + + if params.method == "1best": + logging.info("Use HLG decoding") + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + if params.method == "nbest-rescoring": + logging.info("Use HLG decoding + LM rescoring") + best_path_dict = rescore_with_n_best_list( + lattice=lattice, + G=G, + num_paths=params.num_paths, + lm_scale_list=[params.ngram_lm_scale], + nbest_scale=params.nbest_scale, + ) + best_path = next(iter(best_path_dict.values())) + elif params.method == "whole-lattice-rescoring": + logging.info("Use HLG decoding + LM rescoring") + best_path_dict = rescore_with_whole_lattice( + lattice=lattice, + G_with_epsilon_loops=G, + lm_scale_list=[params.ngram_lm_scale], + ) + best_path = next(iter(best_path_dict.values())) + + hyps = get_texts(best_path) + word_sym_table = k2.SymbolTable.from_file(params.words_file) + hyps = [[word_sym_table[i] for i in ids] for ids in hyps] + else: + raise ValueError(f"Unsupported decoding method: {params.method}") + + s = "\n" + if params.method == "ctc-decoding": + for filename, hyp in zip(params.sound_files, hyps): + words = "".join(hyp) + words = words.replace("▁", " ").strip() + s += f"{filename}:\n{words}\n\n" + elif params.method in [ + "1best", + "nbest-rescoring", + "whole-lattice-rescoring", + ]: + for filename, hyp in zip(params.sound_files, hyps): + words = " ".join(hyp) + words = words.replace("▁", " ").strip() + 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/ami/ASR/zipformer/profile.py b/egs/ami/ASR/zipformer/profile.py new file mode 120000 index 000000000..c93adbd14 --- /dev/null +++ b/egs/ami/ASR/zipformer/profile.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/profile.py \ No newline at end of file diff --git a/egs/ami/ASR/zipformer/scaling.py b/egs/ami/ASR/zipformer/scaling.py new file mode 120000 index 000000000..6f398f431 --- /dev/null +++ b/egs/ami/ASR/zipformer/scaling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling.py \ No newline at end of file diff --git a/egs/ami/ASR/zipformer/scaling_converter.py b/egs/ami/ASR/zipformer/scaling_converter.py new file mode 120000 index 000000000..b0ecee05e --- /dev/null +++ b/egs/ami/ASR/zipformer/scaling_converter.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling_converter.py \ No newline at end of file diff --git a/egs/ami/ASR/zipformer/streaming_beam_search.py b/egs/ami/ASR/zipformer/streaming_beam_search.py new file mode 120000 index 000000000..b1ed54557 --- /dev/null +++ b/egs/ami/ASR/zipformer/streaming_beam_search.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/streaming_beam_search.py \ No newline at end of file diff --git a/egs/ami/ASR/zipformer/streaming_decode.py b/egs/ami/ASR/zipformer/streaming_decode.py new file mode 100755 index 000000000..904caf8af --- /dev/null +++ b/egs/ami/ASR/zipformer/streaming_decode.py @@ -0,0 +1,853 @@ +#!/usr/bin/env python3 +# Copyright 2022-2023 Xiaomi Corporation (Authors: Wei Kang, +# Fangjun Kuang, +# Zengwei Yao) +# +# 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: +./zipformer/streaming_decode.py \ + --epoch 28 \ + --avg 15 \ + --causal 1 \ + --chunk-size 32 \ + --left-context-frames 256 \ + --exp-dir ./zipformer/exp \ + --decoding-method greedy_search \ + --num-decode-streams 2000 +""" + +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 +from asr_datamodule import LibriSpeechAsrDataModule +from decode_stream import DecodeStream +from kaldifeat import Fbank, FbankOptions +from lhotse import CutSet +from streaming_beam_search import ( + fast_beam_search_one_best, + greedy_search, + modified_beam_search, +) +from torch import Tensor, nn +from torch.nn.utils.rnn import pad_sequence +from train import add_model_arguments, get_params, get_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import ( + AttributeDict, + make_pad_mask, + 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 1. + 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( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/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="""Supported decoding methods are: + greedy_search + modified_beam_search + fast_beam_search + """, + ) + + parser.add_argument( + "--num_active_paths", + type=int, + default=4, + help="""An interger indicating how many candidates we will keep for each + frame. Used only when --decoding-method is 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 --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( + "--num-decode-streams", + type=int, + default=2000, + help="The number of streams that can be decoded parallel.", + ) + + add_model_arguments(parser) + + return parser + + +def get_init_states( + model: nn.Module, + batch_size: int = 1, + device: torch.device = torch.device("cpu"), +) -> List[torch.Tensor]: + """ + Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6] + is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). + states[-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + states[-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + """ + states = model.encoder.get_init_states(batch_size, device) + + embed_states = model.encoder_embed.get_init_states(batch_size, device) + states.append(embed_states) + + processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device) + states.append(processed_lens) + + return states + + +def stack_states(state_list: List[List[torch.Tensor]]) -> List[torch.Tensor]: + """Stack list of zipformer states that correspond to separate utterances + into a single emformer state, so that it can be used as an input for + zipformer when those utterances are formed into a batch. + + Args: + state_list: + Each element in state_list corresponding to the internal state + of the zipformer model for a single utterance. For element-n, + state_list[n] is a list of cached tensors of all encoder layers. For layer-i, + state_list[n][i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, + cached_val2, cached_conv1, cached_conv2). + state_list[n][-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + state_list[n][-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + + Note: + It is the inverse of :func:`unstack_states`. + """ + batch_size = len(state_list) + assert (len(state_list[0]) - 2) % 6 == 0, len(state_list[0]) + tot_num_layers = (len(state_list[0]) - 2) // 6 + + batch_states = [] + for layer in range(tot_num_layers): + layer_offset = layer * 6 + # cached_key: (left_context_len, batch_size, key_dim) + cached_key = torch.cat( + [state_list[i][layer_offset] for i in range(batch_size)], dim=1 + ) + # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim) + cached_nonlin_attn = torch.cat( + [state_list[i][layer_offset + 1] for i in range(batch_size)], dim=1 + ) + # cached_val1: (left_context_len, batch_size, value_dim) + cached_val1 = torch.cat( + [state_list[i][layer_offset + 2] for i in range(batch_size)], dim=1 + ) + # cached_val2: (left_context_len, batch_size, value_dim) + cached_val2 = torch.cat( + [state_list[i][layer_offset + 3] for i in range(batch_size)], dim=1 + ) + # cached_conv1: (#batch, channels, left_pad) + cached_conv1 = torch.cat( + [state_list[i][layer_offset + 4] for i in range(batch_size)], dim=0 + ) + # cached_conv2: (#batch, channels, left_pad) + cached_conv2 = torch.cat( + [state_list[i][layer_offset + 5] for i in range(batch_size)], dim=0 + ) + batch_states += [ + cached_key, + cached_nonlin_attn, + cached_val1, + cached_val2, + cached_conv1, + cached_conv2, + ] + + cached_embed_left_pad = torch.cat( + [state_list[i][-2] for i in range(batch_size)], dim=0 + ) + batch_states.append(cached_embed_left_pad) + + processed_lens = torch.cat([state_list[i][-1] for i in range(batch_size)], dim=0) + batch_states.append(processed_lens) + + return batch_states + + +def unstack_states(batch_states: List[Tensor]) -> List[List[Tensor]]: + """Unstack the zipformer state corresponding to a batch of utterances + into a list of states, where the i-th entry is the state from the i-th + utterance in the batch. + + Note: + It is the inverse of :func:`stack_states`. + + Args: + batch_states: A list of cached tensors of all encoder layers. For layer-i, + states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, + cached_conv1, cached_conv2). + state_list[-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + states[-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + + Returns: + state_list: A list of list. Each element in state_list corresponding to the internal state + of the zipformer model for a single utterance. + """ + assert (len(batch_states) - 2) % 6 == 0, len(batch_states) + tot_num_layers = (len(batch_states) - 2) // 6 + + processed_lens = batch_states[-1] + batch_size = processed_lens.shape[0] + + state_list = [[] for _ in range(batch_size)] + + for layer in range(tot_num_layers): + layer_offset = layer * 6 + # cached_key: (left_context_len, batch_size, key_dim) + cached_key_list = batch_states[layer_offset].chunk(chunks=batch_size, dim=1) + # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim) + cached_nonlin_attn_list = batch_states[layer_offset + 1].chunk( + chunks=batch_size, dim=1 + ) + # cached_val1: (left_context_len, batch_size, value_dim) + cached_val1_list = batch_states[layer_offset + 2].chunk( + chunks=batch_size, dim=1 + ) + # cached_val2: (left_context_len, batch_size, value_dim) + cached_val2_list = batch_states[layer_offset + 3].chunk( + chunks=batch_size, dim=1 + ) + # cached_conv1: (#batch, channels, left_pad) + cached_conv1_list = batch_states[layer_offset + 4].chunk( + chunks=batch_size, dim=0 + ) + # cached_conv2: (#batch, channels, left_pad) + cached_conv2_list = batch_states[layer_offset + 5].chunk( + chunks=batch_size, dim=0 + ) + for i in range(batch_size): + state_list[i] += [ + cached_key_list[i], + cached_nonlin_attn_list[i], + cached_val1_list[i], + cached_val2_list[i], + cached_conv1_list[i], + cached_conv2_list[i], + ] + + cached_embed_left_pad_list = batch_states[-2].chunk(chunks=batch_size, dim=0) + for i in range(batch_size): + state_list[i].append(cached_embed_left_pad_list[i]) + + processed_lens_list = batch_states[-1].chunk(chunks=batch_size, dim=0) + for i in range(batch_size): + state_list[i].append(processed_lens_list[i]) + + return state_list + + +def streaming_forward( + features: Tensor, + feature_lens: Tensor, + model: nn.Module, + states: List[Tensor], + chunk_size: int, + left_context_len: int, +) -> Tuple[Tensor, Tensor, List[Tensor]]: + """ + Returns encoder outputs, output lengths, and updated states. + """ + cached_embed_left_pad = states[-2] + (x, x_lens, new_cached_embed_left_pad,) = model.encoder_embed.streaming_forward( + x=features, + x_lens=feature_lens, + cached_left_pad=cached_embed_left_pad, + ) + assert x.size(1) == chunk_size, (x.size(1), chunk_size) + + src_key_padding_mask = make_pad_mask(x_lens) + + # processed_mask is used to mask out initial states + processed_mask = torch.arange(left_context_len, device=x.device).expand( + x.size(0), left_context_len + ) + processed_lens = states[-1] # (batch,) + # (batch, left_context_size) + processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1) + # Update processed lengths + new_processed_lens = processed_lens + x_lens + + # (batch, left_context_size + chunk_size) + src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1) + + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + encoder_states = states[:-2] + ( + encoder_out, + encoder_out_lens, + new_encoder_states, + ) = model.encoder.streaming_forward( + x=x, + x_lens=x_lens, + states=encoder_states, + src_key_padding_mask=src_key_padding_mask, + ) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + new_states = new_encoder_states + [ + new_cached_embed_left_pad, + new_processed_lens, + ] + return encoder_out, encoder_out_lens, new_states + + +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 + chunk_size = int(params.chunk_size) + left_context_len = int(params.left_context_frames) + + features = [] + feature_lens = [] + states = [] + processed_lens = [] # Used in fast-beam-search + + for stream in decode_streams: + feat, feat_len = stream.get_feature_frames(chunk_size * 2) + features.append(feat) + feature_lens.append(feat_len) + states.append(stream.states) + processed_lens.append(stream.done_frames) + + feature_lens = torch.tensor(feature_lens, device=device) + features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS) + + # Make sure the length after encoder_embed is at least 1. + # The encoder_embed subsample features (T - 7) // 2 + # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling + tail_length = chunk_size * 2 + 7 + 2 * 3 + if features.size(1) < tail_length: + pad_length = tail_length - features.size(1) + feature_lens += pad_length + features = torch.nn.functional.pad( + features, + (0, 0, 0, pad_length), + mode="constant", + value=LOG_EPS, + ) + + states = stack_states(states) + + encoder_out, encoder_out_lens, new_states = streaming_forward( + features=features, + feature_lens=feature_lens, + model=model, + states=states, + chunk_size=chunk_size, + left_context_len=left_context_len, + ) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + if params.decoding_method == "greedy_search": + greedy_search(model=model, encoder_out=encoder_out, streams=decode_streams) + elif params.decoding_method == "fast_beam_search": + processed_lens = torch.tensor(processed_lens, device=device) + processed_lens = processed_lens + encoder_out_lens + fast_beam_search_one_best( + model=model, + encoder_out=encoder_out, + processed_lens=processed_lens, + streams=decode_streams, + beam=params.beam, + max_states=params.max_states, + max_contexts=params.max_contexts, + ) + elif params.decoding_method == "modified_beam_search": + modified_beam_search( + model=model, + streams=decode_streams, + encoder_out=encoder_out, + num_active_paths=params.num_active_paths, + ) + else: + raise ValueError(f"Unsupported decoding method: {params.decoding_method}") + + states = unstack_states(new_states) + + finished_streams = [] + for i in range(len(decode_streams)): + decode_streams[i].states = states[i] + decode_streams[i].done_frames += encoder_out_lens[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 = 100 + + decode_results = [] + # Contain decode streams currently running. + decode_streams = [] + for num, cut in enumerate(cuts): + # each utterance has a DecodeStream. + initial_states = get_init_states(model=model, batch_size=1, device=device) + decode_stream = DecodeStream( + params=params, + cut_id=cut.id, + 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, tail_pad_len=30) + 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): + decode_results.append( + ( + decode_streams[i].id, + decode_streams[i].ground_truth.split(), + sp.decode(decode_streams[i].decoding_result()).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): + decode_results.append( + ( + decode_streams[i].id, + decode_streams[i].ground_truth.split(), + sp.decode(decode_streams[i].decoding_result()).split(), + ) + ) + del decode_streams[i] + + if params.decoding_method == "greedy_search": + key = "greedy_search" + elif params.decoding_method == "fast_beam_search": + key = ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}" + ) + elif params.decoding_method == "modified_beam_search": + key = f"num_active_paths_{params.num_active_paths}" + else: + raise ValueError(f"Unsupported decoding method: {params.decoding_method}") + 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" + ) + 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() + LibriSpeechAsrDataModule.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}" + + assert params.causal, params.causal + assert "," not in params.chunk_size, "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + # 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}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + 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() + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + 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)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + 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}") + + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_sets = ["test-clean", "test-other"] + test_cuts = [test_clean_cuts, test_other_cuts] + + 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/ami/ASR/zipformer/subsampling.py b/egs/ami/ASR/zipformer/subsampling.py new file mode 120000 index 000000000..01ae9002c --- /dev/null +++ b/egs/ami/ASR/zipformer/subsampling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/subsampling.py \ No newline at end of file diff --git a/egs/ami/ASR/zipformer/test_scaling.py b/egs/ami/ASR/zipformer/test_scaling.py new file mode 120000 index 000000000..715798436 --- /dev/null +++ b/egs/ami/ASR/zipformer/test_scaling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/test_scaling.py \ No newline at end of file diff --git a/egs/ami/ASR/zipformer/test_subsampling.py b/egs/ami/ASR/zipformer/test_subsampling.py new file mode 120000 index 000000000..bf0ee3d11 --- /dev/null +++ b/egs/ami/ASR/zipformer/test_subsampling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/test_subsampling.py \ No newline at end of file diff --git a/egs/ami/ASR/zipformer/train.py b/egs/ami/ASR/zipformer/train.py new file mode 100755 index 000000000..7d42e87b4 --- /dev/null +++ b/egs/ami/ASR/zipformer/train.py @@ -0,0 +1,1344 @@ +#!/usr/bin/env python3 +# Copyright 2021-2023 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# 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. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +# For non-streaming model training: +./zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --full-libri 1 \ + --max-duration 1000 + +# For streaming model training: +./zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --causal 1 \ + --full-libri 1 \ + --max-duration 1000 + +It supports training with: + - transducer loss (default), with `--use-transducer True --use-ctc False` + - ctc loss (not recommended), with `--use-transducer False --use-ctc True` + - transducer loss & ctc loss, with `--use-transducer True --use-ctc True` +""" + + +import argparse +import copy +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 AmiAsrDataModule +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 AsrModel +from optim import Eden, ScaledAdam +from scaling import ScheduledFloat +from subsampling import Conv2dSubsampling +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 zipformer import Zipformer2 + +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, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,3,4,3,2", + help="Number of zipformer encoder layers per stack, comma separated.", + ) + + parser.add_argument( + "--downsampling-factor", + type=str, + default="1,2,4,8,4,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--feedforward-dim", + type=str, + default="512,768,1024,1536,1024,768", + help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.", + ) + + parser.add_argument( + "--num-heads", + type=str, + default="4,4,4,8,4,4", + help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", + ) + + parser.add_argument( + "--encoder-dim", + type=str, + default="192,256,384,512,384,256", + help="Embedding dimension in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--query-head-dim", + type=str, + default="32", + help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--value-head-dim", + type=str, + default="12", + help="Value dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-head-dim", + type=str, + default="4", + help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-dim", + type=int, + default="48", + help="Positional-encoding embedding dimension", + ) + + parser.add_argument( + "--encoder-unmasked-dim", + type=str, + default="192,192,256,256,256,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "A single int or comma-separated list. Must be <= each corresponding encoder_dim.", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=str, + default="31,31,15,15,15,31", + help="Sizes of convolutional kernels in convolution modules in each encoder stack: " + "a single int or comma-separated list.", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--causal", + type=str2bool, + default=False, + help="If True, use causal version of model.", + ) + + parser.add_argument( + "--chunk-size", + type=str, + default="16,32,64,-1", + help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. " + " Must be just -1 if --causal=False", + ) + + parser.add_argument( + "--left-context-frames", + type=str, + default="64,128,256,-1", + help="Maximum left-contexts for causal training, measured in frames which will " + "be converted to a number of chunks. If splitting into chunks, " + "chunk left-context frames will be chosen randomly from this list; else not relevant.", + ) + + parser.add_argument( + "--use-transducer", + type=str2bool, + default=True, + help="If True, use Transducer head.", + ) + + parser.add_argument( + "--use-ctc", + type=str2bool, + default=False, + help="If True, use CTC head.", + ) + + +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=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/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="zipformer/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( + "--base-lr", type=float, default=0.045, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=7500, + 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=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + 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( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC 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( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=4000, + 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 1. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + 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( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + 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 warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + 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 zipformer + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed. + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_embed(params: AttributeDict) -> nn.Module: + # encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, (T - 7) // 2, encoder_dims). + # That is, it does two things simultaneously: + # (1) subsampling: T -> (T - 7) // 2 + # (2) embedding: num_features -> encoder_dims + # In the normal configuration, we will downsample once more at the end + # by a factor of 2, and most of the encoder stacks will run at a lower + # sampling rate. + encoder_embed = Conv2dSubsampling( + in_channels=params.feature_dim, + out_channels=_to_int_tuple(params.encoder_dim)[0], + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + ) + return encoder_embed + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Zipformer2( + output_downsampling_factor=2, + downsampling_factor=_to_int_tuple(params.downsampling_factor), + num_encoder_layers=_to_int_tuple(params.num_encoder_layers), + encoder_dim=_to_int_tuple(params.encoder_dim), + encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim), + query_head_dim=_to_int_tuple(params.query_head_dim), + pos_head_dim=_to_int_tuple(params.pos_head_dim), + value_head_dim=_to_int_tuple(params.value_head_dim), + pos_dim=params.pos_dim, + num_heads=_to_int_tuple(params.num_heads), + feedforward_dim=_to_int_tuple(params.feedforward_dim), + cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel), + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + warmup_batches=4000.0, + causal=params.causal, + chunk_size=_to_int_tuple(params.chunk_size), + left_context_frames=_to_int_tuple(params.left_context_frames), + ) + 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=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_model(params: AttributeDict) -> nn.Module: + assert params.use_transducer or params.use_ctc, ( + f"At least one of them should be True, " + f"but got params.use_transducer={params.use_transducer}, " + f"params.use_ctc={params.use_ctc}" + ) + + encoder_embed = get_encoder_embed(params) + encoder = get_encoder_model(params) + + if params.use_transducer: + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + else: + decoder = None + joiner = None + + model = AsrModel( + encoder_embed=encoder_embed, + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + vocab_size=params.vocab_size, + use_transducer=params.use_transducer, + use_ctc=params.use_ctc, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + 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 larger than 1, 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. + model_avg: + The stored model averaged from the start of training. + 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 > 1: + 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, + model_avg=model_avg, + 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"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + 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. + model_avg: + The stored model averaged from the start of training. + 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, + model_avg=model_avg, + 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: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute 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 Zipformer 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 if isinstance(model, DDP) else next(model.parameters()).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) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_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, + ) + + loss = 0.0 + + if params.use_transducer: + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + if params.use_ctc: + loss += params.ctc_loss_scale * ctc_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() + if params.use_transducer: + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + if params.use_ctc: + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + 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: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + 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. + model_avg: + The stored model averaged from the start of 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() + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + + 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, + ) + # 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 + save_bad_model() + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + 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}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + 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 params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + 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}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + 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)) + + 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 not params.use_transducer: + params.ctc_loss_scale = 1.0 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + 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: + opts = diagnostics.TensorDiagnosticOptions( + 512 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + ami = AmiAsrDataModule(args) + + train_cuts = ami.train_cuts(sp=sp) + + 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 = ami.train_dataloaders(train_cuts, sampler_state_dict=sampler_state_dict) + + valid_cuts = ami.dev_ihm_cuts() + valid_dl = ami.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, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + 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 + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + 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, + model_avg=model_avg, + 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, + model_avg=model_avg, + 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: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 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, + ) + loss.backward() + 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 + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + AmiAsrDataModule.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/ami/ASR/zipformer/zipformer.py b/egs/ami/ASR/zipformer/zipformer.py new file mode 120000 index 000000000..23011dda7 --- /dev/null +++ b/egs/ami/ASR/zipformer/zipformer.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/zipformer.py \ No newline at end of file From e20af6f28b9f76b1520f08451c3875f4b2d043fe Mon Sep 17 00:00:00 2001 From: jinzr <60612200+JinZr@users.noreply.github.com> Date: Sun, 15 Oct 2023 22:56:06 +0800 Subject: [PATCH 2/5] Update decode.py --- egs/ami/ASR/zipformer/decode.py | 331 +++++--------------------------- 1 file changed, 43 insertions(+), 288 deletions(-) diff --git a/egs/ami/ASR/zipformer/decode.py b/egs/ami/ASR/zipformer/decode.py index 3531d657f..be36d3609 100755 --- a/egs/ami/ASR/zipformer/decode.py +++ b/egs/ami/ASR/zipformer/decode.py @@ -106,24 +106,18 @@ import k2 import sentencepiece as spm import torch import torch.nn as nn -from asr_datamodule import LibriSpeechAsrDataModule +from asr_datamodule import AmiAsrDataModule 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, - modified_beam_search_lm_rescore, - modified_beam_search_lm_rescore_LODR, - modified_beam_search_lm_shallow_fusion, - modified_beam_search_LODR, ) from train import add_model_arguments, get_model, get_params -from icefall import ContextGraph, LmScorer, NgramLm +from icefall import LmScorer from icefall.checkpoint import ( average_checkpoints, average_checkpoints_with_averaged_model, @@ -133,7 +127,6 @@ from icefall.checkpoint import ( from icefall.lexicon import Lexicon from icefall.utils import ( AttributeDict, - make_pad_mask, setup_logger, store_transcripts, str2bool, @@ -308,68 +301,6 @@ def get_parser(): fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", ) - parser.add_argument( - "--use-shallow-fusion", - type=str2bool, - default=False, - help="""Use neural network LM for shallow fusion. - If you want to use LODR, you will also need to set this to true - """, - ) - - parser.add_argument( - "--lm-type", - type=str, - default="rnn", - help="Type of NN lm", - choices=["rnn", "transformer"], - ) - - parser.add_argument( - "--lm-scale", - type=float, - default=0.3, - help="""The scale of the neural network LM - Used only when `--use-shallow-fusion` is set to True. - """, - ) - - parser.add_argument( - "--tokens-ngram", - type=int, - default=2, - help="""The order of the ngram lm. - """, - ) - - parser.add_argument( - "--backoff-id", - type=int, - default=500, - help="ID of the backoff symbol in the ngram LM", - ) - - parser.add_argument( - "--context-score", - type=float, - default=2, - help=""" - The bonus score of each token for the context biasing words/phrases. - Used only when --decoding-method is modified_beam_search and - modified_beam_search_LODR. - """, - ) - - parser.add_argument( - "--context-file", - type=str, - default="", - help=""" - The path of the context biasing lists, one word/phrase each line - Used only when --decoding-method is modified_beam_search and - modified_beam_search_LODR. - """, - ) add_model_arguments(parser) return parser @@ -380,12 +311,8 @@ def decode_one_batch( model: nn.Module, sp: spm.SentencePieceProcessor, batch: dict, - word_table: Optional[k2.SymbolTable] = None, decoding_graph: Optional[k2.Fsa] = None, - context_graph: Optional[ContextGraph] = None, - LM: Optional[LmScorer] = None, - ngram_lm=None, - ngram_lm_scale: float = 0.0, + word_table: Optional[k2.SymbolTable] = None, ) -> Dict[str, List[List[str]]]: """Decode one batch and return the result in a dict. The dict has the following format: @@ -474,35 +401,6 @@ def decode_one_batch( ) 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, @@ -517,55 +415,9 @@ def decode_one_batch( encoder_out=encoder_out, encoder_out_lens=encoder_out_lens, beam=params.beam_size, - context_graph=context_graph, ) for hyp in sp.decode(hyp_tokens): hyps.append(hyp.split()) - elif params.decoding_method == "modified_beam_search_lm_shallow_fusion": - hyp_tokens = modified_beam_search_lm_shallow_fusion( - model=model, - encoder_out=encoder_out, - encoder_out_lens=encoder_out_lens, - beam=params.beam_size, - LM=LM, - ) - for hyp in sp.decode(hyp_tokens): - hyps.append(hyp.split()) - elif params.decoding_method == "modified_beam_search_LODR": - hyp_tokens = modified_beam_search_LODR( - model=model, - encoder_out=encoder_out, - encoder_out_lens=encoder_out_lens, - beam=params.beam_size, - LODR_lm=ngram_lm, - LODR_lm_scale=ngram_lm_scale, - LM=LM, - context_graph=context_graph, - ) - for hyp in sp.decode(hyp_tokens): - hyps.append(hyp.split()) - elif params.decoding_method == "modified_beam_search_lm_rescore": - lm_scale_list = [0.01 * i for i in range(10, 50)] - ans_dict = modified_beam_search_lm_rescore( - model=model, - encoder_out=encoder_out, - encoder_out_lens=encoder_out_lens, - beam=params.beam_size, - LM=LM, - lm_scale_list=lm_scale_list, - ) - elif params.decoding_method == "modified_beam_search_lm_rescore_LODR": - lm_scale_list = [0.02 * i for i in range(2, 30)] - ans_dict = modified_beam_search_lm_rescore_LODR( - model=model, - encoder_out=encoder_out, - encoder_out_lens=encoder_out_lens, - beam=params.beam_size, - LM=LM, - LODR_lm=ngram_lm, - sp=sp, - lm_scale_list=lm_scale_list, - ) else: batch_size = encoder_out.size(0) @@ -593,6 +445,14 @@ def decode_one_batch( 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}_" @@ -604,22 +464,6 @@ def decode_one_batch( key += f"_ngram_lm_scale_{params.ngram_lm_scale}" return {key: hyps} - elif "modified_beam_search" in params.decoding_method: - prefix = f"beam_size_{params.beam_size}" - if params.decoding_method in ( - "modified_beam_search_lm_rescore", - "modified_beam_search_lm_rescore_LODR", - ): - ans = dict() - assert ans_dict is not None - for key, hyps in ans_dict.items(): - hyps = [sp.decode(hyp).split() for hyp in hyps] - ans[f"{prefix}_{key}"] = hyps - return ans - else: - if params.has_contexts: - prefix += f"-context-score-{params.context_score}" - return {prefix: hyps} else: return {f"beam_size_{params.beam_size}": hyps} @@ -629,12 +473,8 @@ def decode_dataset( params: AttributeDict, model: nn.Module, sp: spm.SentencePieceProcessor, - word_table: Optional[k2.SymbolTable] = None, decoding_graph: Optional[k2.Fsa] = None, - context_graph: Optional[ContextGraph] = None, - LM: Optional[LmScorer] = None, - ngram_lm=None, - ngram_lm_scale: float = 0.0, + word_table: Optional[k2.SymbolTable] = None, ) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: """Decode dataset. @@ -682,12 +522,8 @@ def decode_dataset( model=model, sp=sp, decoding_graph=decoding_graph, - context_graph=context_graph, word_table=word_table, batch=batch, - LM=LM, - ngram_lm=ngram_lm, - ngram_lm_scale=ngram_lm_scale, ) for name, hyps in hyps_dict.items(): @@ -755,7 +591,7 @@ def save_results( @torch.no_grad() def main(): parser = get_parser() - LibriSpeechAsrDataModule.add_arguments(parser) + AmiAsrDataModule.add_arguments(parser) LmScorer.add_arguments(parser) args = parser.parse_args() args.exp_dir = Path(args.exp_dir) @@ -767,37 +603,16 @@ def main(): "greedy_search", "beam_search", "fast_beam_search", - "fast_beam_search_nbest", "fast_beam_search_nbest_LG", - "fast_beam_search_nbest_oracle", "modified_beam_search", - "modified_beam_search_LODR", - "modified_beam_search_lm_shallow_fusion", - "modified_beam_search_lm_rescore", - "modified_beam_search_lm_rescore_LODR", ) params.res_dir = params.exp_dir / params.decoding_method - if os.path.exists(params.context_file): - params.has_contexts = True - else: - params.has_contexts = False - 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.causal: - assert ( - "," not in params.chunk_size - ), "chunk_size should be one value in decoding." - assert ( - "," not in params.left_context_frames - ), "left_context_frames should be one value in decoding." - params.suffix += f"-chunk-{params.chunk_size}" - params.suffix += f"-left-context-{params.left_context_frames}" - if "fast_beam_search" in params.decoding_method: params.suffix += f"-beam-{params.beam}" params.suffix += f"-max-contexts-{params.max_contexts}" @@ -809,27 +624,10 @@ def main(): 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}" - if params.decoding_method in ( - "modified_beam_search", - "modified_beam_search_LODR", - ): - if params.has_contexts: - params.suffix += f"-context-score-{params.context_score}" else: params.suffix += f"-context-{params.context_size}" params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" - if params.use_shallow_fusion: - params.suffix += f"-{params.lm_type}-lm-scale-{params.lm_scale}" - - if "LODR" in params.decoding_method: - params.suffix += ( - f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}" - ) - - if params.use_averaged_model: - params.suffix += "-use-averaged-model" - setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") logging.info("Decoding started") @@ -932,54 +730,6 @@ def main(): model.to(device) model.eval() - # only load the neural network LM if required - if params.use_shallow_fusion or params.decoding_method in ( - "modified_beam_search_lm_rescore", - "modified_beam_search_lm_rescore_LODR", - "modified_beam_search_lm_shallow_fusion", - "modified_beam_search_LODR", - ): - LM = LmScorer( - lm_type=params.lm_type, - params=params, - device=device, - lm_scale=params.lm_scale, - ) - LM.to(device) - LM.eval() - else: - LM = None - - # only load N-gram LM when needed - if params.decoding_method == "modified_beam_search_lm_rescore_LODR": - try: - import kenlm - except ImportError: - print("Please install kenlm first. You can use") - print(" pip install https://github.com/kpu/kenlm/archive/master.zip") - print("to install it") - import sys - - sys.exit(-1) - ngram_file_name = str(params.lang_dir / f"{params.tokens_ngram}gram.arpa") - logging.info(f"lm filename: {ngram_file_name}") - ngram_lm = kenlm.Model(ngram_file_name) - ngram_lm_scale = None # use a list to search - - elif params.decoding_method == "modified_beam_search_LODR": - lm_filename = f"{params.tokens_ngram}gram.fst.txt" - logging.info(f"Loading token level lm: {lm_filename}") - ngram_lm = NgramLm( - str(params.lang_dir / lm_filename), - backoff_id=params.backoff_id, - is_binary=False, - ) - logging.info(f"num states: {ngram_lm.lm.num_states}") - ngram_lm_scale = params.ngram_lm_scale - else: - ngram_lm = None - ngram_lm_scale = None - if "fast_beam_search" in params.decoding_method: if params.decoding_method == "fast_beam_search_nbest_LG": lexicon = Lexicon(params.lang_dir) @@ -997,46 +747,51 @@ def main(): decoding_graph = None word_table = None - if "modified_beam_search" in params.decoding_method: - if os.path.exists(params.context_file): - contexts = [] - for line in open(params.context_file).readlines(): - contexts.append(line.strip()) - context_graph = ContextGraph(params.context_score) - context_graph.build(sp.encode(contexts)) - else: - context_graph = None - else: - context_graph = None - num_param = sum([p.numel() for p in model.parameters()]) logging.info(f"Number of model parameters: {num_param}") # we need cut ids to display recognition results. args.return_cuts = True - librispeech = LibriSpeechAsrDataModule(args) - test_clean_cuts = librispeech.test_clean_cuts() - test_other_cuts = librispeech.test_other_cuts() + ami = AmiAsrDataModule(args) - test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) - test_other_dl = librispeech.test_dataloaders(test_other_cuts) + dev_ihm_cuts = ami.dev_ihm_cuts() + test_ihm_cuts = ami.test_ihm_cuts() + dev_sdm_cuts = ami.dev_sdm_cuts() + test_sdm_cuts = ami.test_sdm_cuts() + dev_gss_cuts = ami.dev_gss_cuts() + test_gss_cuts = ami.test_gss_cuts() - test_sets = ["test-clean", "test-other"] - test_dl = [test_clean_dl, test_other_dl] + dev_ihm_dl = ami.test_dataloaders(dev_ihm_cuts) + test_ihm_dl = ami.test_dataloaders(test_ihm_cuts) + dev_sdm_dl = ami.test_dataloaders(dev_sdm_cuts) + test_sdm_dl = ami.test_dataloaders(test_sdm_cuts) + if dev_gss_cuts is not None: + dev_gss_dl = ami.test_dataloaders(dev_gss_cuts) + if test_gss_cuts is not None: + test_gss_dl = ami.test_dataloaders(test_gss_cuts) - for test_set, test_dl in zip(test_sets, test_dl): + test_sets = { + "dev_ihm": (dev_ihm_dl, dev_ihm_cuts), + "test_ihm": (test_ihm_dl, test_ihm_cuts), + "dev_sdm": (dev_sdm_dl, dev_sdm_cuts), + "test_sdm": (test_sdm_dl, test_sdm_cuts), + } + if dev_gss_cuts is not None: + test_sets["dev_gss"] = (dev_gss_dl, dev_gss_cuts) + if test_gss_cuts is not None: + test_sets["test_gss"] = (test_gss_dl, test_gss_cuts) + + for test_set in test_sets: + logging.info(f"Decoding {test_set}") + dl, cuts = test_sets[test_set] results_dict = decode_dataset( - dl=test_dl, + dl=dl, params=params, model=model, sp=sp, word_table=word_table, decoding_graph=decoding_graph, - context_graph=context_graph, - LM=LM, - ngram_lm=ngram_lm, - ngram_lm_scale=ngram_lm_scale, ) save_results( From 014a59b597de6dccf07ffd62d62bc1eb8e560009 Mon Sep 17 00:00:00 2001 From: jinzr <60612200+JinZr@users.noreply.github.com> Date: Tue, 17 Oct 2023 21:58:48 +0800 Subject: [PATCH 3/5] Update prepare_ami_gss.sh --- egs/ami/ASR/local/prepare_ami_gss.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/egs/ami/ASR/local/prepare_ami_gss.sh b/egs/ami/ASR/local/prepare_ami_gss.sh index 414c22b12..7f0fd905f 100755 --- a/egs/ami/ASR/local/prepare_ami_gss.sh +++ b/egs/ami/ASR/local/prepare_ami_gss.sh @@ -47,7 +47,7 @@ fi if [ $stage -le 3 ]; then log "Stage 3: Split manifests for multi-GPU processing (optional)" - for part in train; do + for part in train dev test; do gss utils split $nj $EXP_DIR/cuts_per_segment_${part}.jsonl.gz \ $EXP_DIR/cuts_per_segment_${part}_split$nj done From f3f0dfc52dca4fe5e822a0f9e3ca471aa9f0d2c4 Mon Sep 17 00:00:00 2001 From: jinzr <60612200+JinZr@users.noreply.github.com> Date: Wed, 18 Oct 2023 15:07:12 +0800 Subject: [PATCH 4/5] minor updates --- egs/ami/ASR/local/compute_fbank_ami.py | 6 ++++++ .../pruned_transducer_stateless7/asr_datamodule.py | 13 ++++++++++++- egs/ami/ASR/pruned_transducer_stateless7/decode.py | 6 ++++++ egs/ami/ASR/zipformer/decode.py | 6 ++++++ 4 files changed, 30 insertions(+), 1 deletion(-) diff --git a/egs/ami/ASR/local/compute_fbank_ami.py b/egs/ami/ASR/local/compute_fbank_ami.py index 4892b40e3..c4a2541ea 100755 --- a/egs/ami/ASR/local/compute_fbank_ami.py +++ b/egs/ami/ASR/local/compute_fbank_ami.py @@ -77,6 +77,12 @@ def compute_fbank_ami(): prefix="ami-sdm", suffix="jsonl.gz", ) + manifests_sdm = read_manifests_if_cached( + dataset_parts=["dev", "test"], + output_dir=src_dir, + prefix="ami-mdm", + suffix="jsonl.gz", + ) # For GSS we already have cuts so we read them directly. manifests_gss = read_manifests_if_cached( dataset_parts=["train", "dev", "test"], diff --git a/egs/ami/ASR/pruned_transducer_stateless7/asr_datamodule.py b/egs/ami/ASR/pruned_transducer_stateless7/asr_datamodule.py index 79474f1d8..592c1aaf9 100644 --- a/egs/ami/ASR/pruned_transducer_stateless7/asr_datamodule.py +++ b/egs/ami/ASR/pruned_transducer_stateless7/asr_datamodule.py @@ -294,7 +294,6 @@ class AmiAsrDataModule: return train_dl def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader: - transforms = [] if self.args.concatenate_cuts: transforms = [ @@ -399,6 +398,12 @@ class AmiAsrDataModule: cs = load_manifest_lazy(self.args.manifest_dir / "cuts_dev_sdm.jsonl.gz") return cs.filter(self.remove_short_cuts) + @lru_cache() + def dev_mdm_cuts(self) -> CutSet: + logging.info("About to get AMI MDM dev cuts") + cs = load_manifest_lazy(self.args.manifest_dir / "cuts_dev_mdm.jsonl.gz") + return cs.filter(self.remove_short_cuts) + @lru_cache() def dev_gss_cuts(self) -> CutSet: if not (self.args.manifest_dir / "cuts_dev_gss.jsonl.gz").exists(): @@ -420,6 +425,12 @@ class AmiAsrDataModule: cs = load_manifest_lazy(self.args.manifest_dir / "cuts_test_sdm.jsonl.gz") return cs.filter(self.remove_short_cuts) + @lru_cache() + def test_mdm_cuts(self) -> CutSet: + logging.info("About to get AMI MDM test cuts") + cs = load_manifest_lazy(self.args.manifest_dir / "cuts_test_mdm.jsonl.gz") + return cs.filter(self.remove_short_cuts) + @lru_cache() def test_gss_cuts(self) -> CutSet: if not (self.args.manifest_dir / "cuts_test_gss.jsonl.gz").exists(): diff --git a/egs/ami/ASR/pruned_transducer_stateless7/decode.py b/egs/ami/ASR/pruned_transducer_stateless7/decode.py index 9999894d1..e205004d6 100755 --- a/egs/ami/ASR/pruned_transducer_stateless7/decode.py +++ b/egs/ami/ASR/pruned_transducer_stateless7/decode.py @@ -691,6 +691,8 @@ def main(): test_ihm_cuts = ami.test_ihm_cuts() dev_sdm_cuts = ami.dev_sdm_cuts() test_sdm_cuts = ami.test_sdm_cuts() + dev_mdm_cuts = ami.dev_mdm_cuts() + test_mdm_cuts = ami.test_mdm_cuts() dev_gss_cuts = ami.dev_gss_cuts() test_gss_cuts = ami.test_gss_cuts() @@ -698,6 +700,8 @@ def main(): test_ihm_dl = ami.test_dataloaders(test_ihm_cuts) dev_sdm_dl = ami.test_dataloaders(dev_sdm_cuts) test_sdm_dl = ami.test_dataloaders(test_sdm_cuts) + dev_mdm_dl = ami.test_dataloaders(dev_mdm_cuts) + test_mdm_dl = ami.test_dataloaders(test_mdm_cuts) if dev_gss_cuts is not None: dev_gss_dl = ami.test_dataloaders(dev_gss_cuts) if test_gss_cuts is not None: @@ -708,6 +712,8 @@ def main(): "test_ihm": (test_ihm_dl, test_ihm_cuts), "dev_sdm": (dev_sdm_dl, dev_sdm_cuts), "test_sdm": (test_sdm_dl, test_sdm_cuts), + "dev_mdm": (dev_mdm_dl, dev_mdm_cuts), + "test_mdm": (test_mdm_dl, test_mdm_cuts), } if dev_gss_cuts is not None: test_sets["dev_gss"] = (dev_gss_dl, dev_gss_cuts) diff --git a/egs/ami/ASR/zipformer/decode.py b/egs/ami/ASR/zipformer/decode.py index be36d3609..5a5bf16eb 100755 --- a/egs/ami/ASR/zipformer/decode.py +++ b/egs/ami/ASR/zipformer/decode.py @@ -759,6 +759,8 @@ def main(): test_ihm_cuts = ami.test_ihm_cuts() dev_sdm_cuts = ami.dev_sdm_cuts() test_sdm_cuts = ami.test_sdm_cuts() + dev_mdm_cuts = ami.dev_mdm_cuts() + test_mdm_cuts = ami.test_mdm_cuts() dev_gss_cuts = ami.dev_gss_cuts() test_gss_cuts = ami.test_gss_cuts() @@ -766,6 +768,8 @@ def main(): test_ihm_dl = ami.test_dataloaders(test_ihm_cuts) dev_sdm_dl = ami.test_dataloaders(dev_sdm_cuts) test_sdm_dl = ami.test_dataloaders(test_sdm_cuts) + dev_mdm_dl = ami.test_dataloaders(dev_mdm_cuts) + test_mdm_dl = ami.test_dataloaders(test_mdm_cuts) if dev_gss_cuts is not None: dev_gss_dl = ami.test_dataloaders(dev_gss_cuts) if test_gss_cuts is not None: @@ -776,6 +780,8 @@ def main(): "test_ihm": (test_ihm_dl, test_ihm_cuts), "dev_sdm": (dev_sdm_dl, dev_sdm_cuts), "test_sdm": (test_sdm_dl, test_sdm_cuts), + "dev_mdm": (dev_mdm_dl, dev_mdm_cuts), + "test_mdm": (test_mdm_dl, test_mdm_cuts), } if dev_gss_cuts is not None: test_sets["dev_gss"] = (dev_gss_dl, dev_gss_cuts) From f3b918452ac6b210c25651f974bc8e8141aff101 Mon Sep 17 00:00:00 2001 From: jinzr <60612200+JinZr@users.noreply.github.com> Date: Thu, 19 Oct 2023 11:09:20 +0800 Subject: [PATCH 5/5] arch align --- egs/ami/ASR/zipformer/train.py | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) diff --git a/egs/ami/ASR/zipformer/train.py b/egs/ami/ASR/zipformer/train.py index 7d42e87b4..6efa085c2 100755 --- a/egs/ami/ASR/zipformer/train.py +++ b/egs/ami/ASR/zipformer/train.py @@ -127,35 +127,35 @@ def add_model_arguments(parser: argparse.ArgumentParser): parser.add_argument( "--num-encoder-layers", type=str, - default="2,2,3,4,3,2", + default="2,4,3,2,4", help="Number of zipformer encoder layers per stack, comma separated.", ) parser.add_argument( "--downsampling-factor", type=str, - default="1,2,4,8,4,2", + default="1,2,4,8,2", help="Downsampling factor for each stack of encoder layers.", ) parser.add_argument( "--feedforward-dim", type=str, - default="512,768,1024,1536,1024,768", + default="1024,1024,2048,2048,1024", help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.", ) parser.add_argument( "--num-heads", type=str, - default="4,4,4,8,4,4", + default="8,8,8,8,8", help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", ) parser.add_argument( "--encoder-dim", type=str, - default="192,256,384,512,384,256", + default="384,384,384,384,384", help="Embedding dimension in encoder stacks: a single int or comma-separated list.", ) @@ -190,7 +190,7 @@ def add_model_arguments(parser: argparse.ArgumentParser): parser.add_argument( "--encoder-unmasked-dim", type=str, - default="192,192,256,256,256,192", + default="256,256,256,256,256", help="Unmasked dimensions in the encoders, relates to augmentation during training. " "A single int or comma-separated list. Must be <= each corresponding encoder_dim.", ) @@ -198,7 +198,7 @@ def add_model_arguments(parser: argparse.ArgumentParser): parser.add_argument( "--cnn-module-kernel", type=str, - default="31,31,15,15,15,31", + default="31,31,31,31,31", help="Sizes of convolutional kernels in convolution modules in each encoder stack: " "a single int or comma-separated list.", ) @@ -329,7 +329,7 @@ def get_parser(): ) parser.add_argument( - "--base-lr", type=float, default=0.045, help="The base learning rate." + "--base-lr", type=float, default=0.05, help="The base learning rate." ) parser.add_argument(