diff --git a/egs/aidatatang_200zh/ASR/zipformer/__init__.py b/egs/aidatatang_200zh/ASR/zipformer/__init__.py deleted file mode 100644 index e69de29bb..000000000 diff --git a/egs/aidatatang_200zh/ASR/zipformer/asr_datamodule.py b/egs/aidatatang_200zh/ASR/zipformer/asr_datamodule.py deleted file mode 120000 index a074d6085..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/asr_datamodule.py +++ /dev/null @@ -1 +0,0 @@ -../pruned_transducer_stateless2/asr_datamodule.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/beam_search.py b/egs/aidatatang_200zh/ASR/zipformer/beam_search.py deleted file mode 120000 index 8554e44cc..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/beam_search.py +++ /dev/null @@ -1 +0,0 @@ -../pruned_transducer_stateless2/beam_search.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/decode.py b/egs/aidatatang_200zh/ASR/zipformer/decode.py deleted file mode 100755 index 8a6d2845a..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/decode.py +++ /dev/null @@ -1,814 +0,0 @@ -#!/usr/bin/env python3 -# -# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, -# Zengwei Yao -# Mingshuang Luo) -# -# See ../../../../LICENSE for clarification regarding multiple authors -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -""" -Usage: -(1) greedy search -./zipformer/decode.py \ - --epoch 35 \ - --avg 15 \ - --exp-dir ./zipformer/exp \ - --lang-dir data/lang_char \ - --max-duration 600 \ - --decoding-method greedy_search - -(2) modified beam search -./zipformer/decode.py \ - --epoch 35 \ - --avg 15 \ - --exp-dir ./zipformer/exp \ - --lang-dir data/lang_char \ - --max-duration 600 \ - --decoding-method modified_beam_search \ - --beam-size 4 - -(3) fast beam search (trivial_graph) -./zipformer/decode.py \ - --epoch 35 \ - --avg 15 \ - --exp-dir ./zipformer/exp \ - --lang-dir data/lang_char \ - --max-duration 600 \ - --decoding-method fast_beam_search \ - --beam 20.0 \ - --max-contexts 8 \ - --max-states 64 - -(4) fast beam search (LG) -./zipformer/decode.py \ - --epoch 30 \ - --avg 15 \ - --exp-dir ./zipformer/exp \ - --lang-dir data/lang_char \ - --max-duration 600 \ - --decoding-method fast_beam_search_LG \ - --beam 20.0 \ - --max-contexts 8 \ - --max-states 64 - -(5) fast beam search (nbest oracle WER) -./zipformer/decode.py \ - --epoch 35 \ - --avg 15 \ - --exp-dir ./zipformer/exp \ - --lang-dir data/lang_char \ - --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 -""" - - -import argparse -import logging -import math -from collections import defaultdict -from pathlib import Path -from typing import Dict, List, Optional, Tuple - -import k2 -import torch -import torch.nn as nn -from asr_datamodule import Aidatatang_200zhAsrDataModule -from beam_search import ( - beam_search, - fast_beam_search_nbest, - fast_beam_search_nbest_LG, - fast_beam_search_nbest_oracle, - fast_beam_search_one_best, - greedy_search, - greedy_search_batch, - modified_beam_search, -) -from lhotse.cut import Cut -from train import add_model_arguments, get_model, get_params - -from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler -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( - "--lang-dir", - type=Path, - default="data/lang_char", - 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 - - modified_beam_search - - fast_beam_search - - fast_beam_search_LG - - fast_beam_search_nbest_oracle - If you use fast_beam_search_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, fast_beam_search_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_LG. - It specifies the scale for n-gram LM scores. - """, - ) - - parser.add_argument( - "--ilme-scale", - type=float, - default=0.2, - help=""" - Used only when --decoding_method is fast_beam_search_LG. - It specifies the scale for the internal language model estimation. - """, - ) - - parser.add_argument( - "--max-contexts", - type=int, - default=8, - help="""Used only when --decoding-method is - fast_beam_search, fast_beam_search, fast_beam_search_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, fast_beam_search_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_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 and fast_beam_search_nbest_oracle""", - ) - - parser.add_argument( - "--blank-penalty", - type=float, - default=0.0, - help=""" - The penalty applied on blank symbol during decoding. - Note: It is a positive value that would be applied to logits like - this `logits[:, 0] -= blank_penalty` (suppose logits.shape is - [batch_size, vocab] and blank id is 0). - """, - ) - - add_model_arguments(parser) - - return parser - - -def decode_one_batch( - params: AttributeDict, - model: nn.Module, - lexicon: Lexicon, - graph_compiler: CharCtcTrainingGraphCompiler, - batch: dict, - decoding_graph: Optional[k2.Fsa] = None, -) -> Dict[str, List[List[str]]]: - """Decode one batch and return the result in a dict. The dict has the - following format: - - - key: It indicates the setting used for decoding. For example, - if greedy_search is used, it would be "greedy_search" - If beam search with a beam size of 7 is used, it would be - "beam_7" - - value: It contains the decoding result. `len(value)` equals to - batch size. `value[i]` is the decoding result for the i-th - utterance in the given batch. - Args: - params: - It's the return value of :func:`get_params`. - model: - The neural model. - batch: - It is the return value from iterating - `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation - for the format of the `batch`. - decoding_graph: - The decoding graph. Can be either a `k2.trivial_graph` or LG, Used - only when --decoding_method is fast_beam_search, fast_beam_search_nbest, - fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. - Returns: - Return the decoding result. See above description for the format of - the returned dict. - """ - device = 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, - ) - - x, x_lens = model.encoder_embed(feature, feature_lens) - - 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 = model.encoder(x, x_lens, src_key_padding_mask) - encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) - - 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, - blank_penalty=params.blank_penalty, - ) - for i in range(encoder_out.size(0)): - hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) - elif params.decoding_method == "fast_beam_search_LG": - 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, - blank_penalty=params.blank_penalty, - ilme_scale=params.ilme_scale, - ) - for hyp in hyp_tokens: - sentence = "".join([lexicon.word_table[i] for i in hyp]) - hyps.append(list(sentence)) - 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=graph_compiler.texts_to_ids(supervisions["text"]), - nbest_scale=params.nbest_scale, - blank_penalty=params.blank_penalty, - ) - for i in range(encoder_out.size(0)): - hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) - 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, - blank_penalty=params.blank_penalty, - ) - for i in range(encoder_out.size(0)): - hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) - elif params.decoding_method == "modified_beam_search": - hyp_tokens = modified_beam_search( - model=model, - encoder_out=encoder_out, - encoder_out_lens=encoder_out_lens, - blank_penalty=params.blank_penalty, - beam=params.beam_size, - ) - for i in range(encoder_out.size(0)): - hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) - 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, - blank_penalty=params.blank_penalty, - ) - elif params.decoding_method == "beam_search": - hyp = beam_search( - model=model, - encoder_out=encoder_out_i, - beam=params.beam_size, - blank_penalty=params.blank_penalty, - ) - else: - raise ValueError( - f"Unsupported decoding method: {params.decoding_method}" - ) - hyps.append([lexicon.token_table[idx] for idx in hyp]) - - key = f"blank_penalty_{params.blank_penalty}" - if params.decoding_method == "greedy_search": - return {"greedy_search_" + key: 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"_ilme_scale_{params.ilme_scale}" - key += f"_ngram_lm_scale_{params.ngram_lm_scale}" - - return {key: hyps} - else: - return {f"beam_size_{params.beam_size}_" + key: hyps} - - -def decode_dataset( - dl: torch.utils.data.DataLoader, - params: AttributeDict, - model: nn.Module, - lexicon: Lexicon, - graph_compiler: CharCtcTrainingGraphCompiler, - decoding_graph: Optional[k2.Fsa] = None, -) -> Dict[str, List[Tuple[List[str], List[str]]]]: - """Decode dataset. - - Args: - dl: - PyTorch's dataloader containing the dataset to decode. - params: - It is returned by :func:`get_params`. - model: - The neural model. - decoding_graph: - The decoding graph. Can be either a `k2.trivial_graph` or LG, 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"] - texts = [list("".join(text.split())) for text in texts] - cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] - - hyps_dict = decode_one_batch( - params=params, - model=model, - lexicon=lexicon, - graph_compiler=graph_compiler, - decoding_graph=decoding_graph, - batch=batch, - ) - - 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): - this_batch.append((cut_id, ref_text, hyp_words)) - - results[name].extend(this_batch) - - num_cuts += len(texts) - - if batch_idx % log_interval == 0: - batch_str = f"{batch_idx}/{num_batches}" - - logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") - return results - - -def save_results( - params: AttributeDict, - test_set_name: str, - results_dict: Dict[str, List[Tuple[List[int], List[int]]]], -): - test_set_wers = dict() - for key, results in results_dict.items(): - recog_path = ( - params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" - ) - 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() - Aidatatang_200zhAsrDataModule.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", - "modified_beam_search", - "fast_beam_search", - "fast_beam_search_LG", - "fast_beam_search_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 "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"_ilme_scale_{params.ilme_scale}" - params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" - elif "beam_search" in params.decoding_method: - params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" - else: - params.suffix += f"-context-{params.context_size}" - params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" - params.suffix += f"-blank-penalty-{params.blank_penalty}" - - 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}") - - lexicon = Lexicon(params.lang_dir) - params.blank_id = lexicon.token_table[""] - params.vocab_size = max(lexicon.tokens) + 1 - - graph_compiler = CharCtcTrainingGraphCompiler( - lexicon=lexicon, - device=device, - ) - - 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() - - if "fast_beam_search" in params.decoding_method: - if "LG" in params.decoding_method: - lexicon = Lexicon(params.lang_dir) - 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: - decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) - else: - decoding_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 - aidatatang_200zh = Aidatatang_200zhAsrDataModule(args) - - def remove_short_utt(c: Cut): - T = ((c.num_frames - 7) // 2 + 1) // 2 - if T <= 0: - logging.warning( - f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}." - ) - return T > 0 - - dev_cuts = aidatatang_200zh.valid_cuts() - dev_cuts = dev_cuts.filter(remove_short_utt) - dev_dl = aidatatang_200zh.valid_dataloaders(dev_cuts) - - test_cuts = aidatatang_200zh.test_cuts() - test_cuts = test_cuts.filter(remove_short_utt) - test_dl = aidatatang_200zh.test_dataloaders(test_cuts) - - test_sets = ["valid_cuts", "test_cuts"] - test_dls = [dev_dl, test_dl] - - for test_set, test_dl in zip(test_sets, test_dls): - results_dict = decode_dataset( - dl=test_dl, - params=params, - model=model, - lexicon=lexicon, - graph_compiler=graph_compiler, - 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/aidatatang_200zh/ASR/zipformer/decode_stream.py b/egs/aidatatang_200zh/ASR/zipformer/decode_stream.py deleted file mode 120000 index b8d8ddfc4..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/decode_stream.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/decode_stream.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/decoder.py b/egs/aidatatang_200zh/ASR/zipformer/decoder.py deleted file mode 120000 index 5a8018680..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/decoder.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/decoder.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/encoder_interface.py b/egs/aidatatang_200zh/ASR/zipformer/encoder_interface.py deleted file mode 120000 index b9aa0ae08..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/encoder_interface.py +++ /dev/null @@ -1 +0,0 @@ -../pruned_transducer_stateless2/encoder_interface.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/export-onnx-streaming.py b/egs/aidatatang_200zh/ASR/zipformer/export-onnx-streaming.py deleted file mode 120000 index 2962eb784..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/export-onnx-streaming.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/export-onnx-streaming.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/export-onnx.py b/egs/aidatatang_200zh/ASR/zipformer/export-onnx.py deleted file mode 120000 index 70a15683c..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/export-onnx.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/export-onnx.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/export.py b/egs/aidatatang_200zh/ASR/zipformer/export.py deleted file mode 120000 index dfc1bec08..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/export.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/export.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/jit_pretrained.py b/egs/aidatatang_200zh/ASR/zipformer/jit_pretrained.py deleted file mode 120000 index 25108391f..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/jit_pretrained.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/jit_pretrained.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/jit_pretrained_streaming.py b/egs/aidatatang_200zh/ASR/zipformer/jit_pretrained_streaming.py deleted file mode 120000 index 1962351e9..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/jit_pretrained_streaming.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/jit_pretrained_streaming.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/joiner.py b/egs/aidatatang_200zh/ASR/zipformer/joiner.py deleted file mode 120000 index 5b8a36332..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/joiner.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/joiner.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/model.py b/egs/aidatatang_200zh/ASR/zipformer/model.py deleted file mode 120000 index cd7e07d72..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/model.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/model.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/onnx_check.py b/egs/aidatatang_200zh/ASR/zipformer/onnx_check.py deleted file mode 120000 index f3dd42004..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/onnx_check.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/onnx_check.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/onnx_decode.py b/egs/aidatatang_200zh/ASR/zipformer/onnx_decode.py deleted file mode 100755 index 6cd91686e..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/onnx_decode.py +++ /dev/null @@ -1,330 +0,0 @@ -#!/usr/bin/env python3 -# -# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, -# Zengwei Yao, -# Xiaoyu Yang, -# 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 loads ONNX exported models and uses them to decode the test sets. - -We use the pre-trained model from -https://huggingface.co/pkufool/icefall-asr-zipformer-wenetspeech-20230615 -as an example to show how to use this file. - -1. Download the pre-trained model - -cd egs/wenetspeech/ASR - -repo_url=https://huggingface.co/pkufool/icefall-asr-zipformer-wenetspeech-20230615 -GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url -repo=$(basename $repo_url) - -pushd $repo -git lfs pull --include "data/lang_char/tokens.txt" -git lfs pull --include "exp/pretrained.pt" - -cd exp -ln -s pretrained.pt epoch-9999.pt -popd - -2. Export the model to ONNX - -./zipformer/export-onnx.py \ - --tokens $repo/data/lang_char/tokens.txt \ - --epoch 9999 \ - --avg 1 \ - --exp-dir $repo/exp/ - -It will generate the following 3 files inside $repo/exp: - - - encoder-epoch-9999-avg-1.onnx - - decoder-epoch-9999-avg-1.onnx - - joiner-epoch-9999-avg-1.onnx - -2. Run this file - -./zipformer/onnx_decode.py \ - --exp-dir ./zipformer/exp \ - --max-duration 600 \ - --encoder-model-filename $repo/exp/encoder-epoch-9999-avg-1.onnx \ - --decoder-model-filename $repo/exp/decoder-epoch-9999-avg-1.onnx \ - --joiner-model-filename $repo/exp/joiner-epoch-9999-avg-1.onnx \ -""" - - -import argparse -import logging -import time -from pathlib import Path -from typing import List, Tuple - -import k2 -import torch -import torch.nn as nn -from asr_datamodule import Aidatatang_200zhAsrDataModule -from lhotse.cut import Cut -from onnx_pretrained import OnnxModel, greedy_search - -from icefall.utils import setup_logger, store_transcripts, write_error_stats - - -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, - default="data/lang_char/tokens.txt", - help="Path to the 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: k2.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: - Mapping ids to tokens. - 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 - ) - - hyps = [[token_table[h] for h in hyp] for hyp in hyps] - return hyps - - -def decode_dataset( - dl: torch.utils.data.DataLoader, - model: nn.Module, - token_table: k2.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: - Mapping ids to tokens. - - 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 = list(ref_text) - 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() - Aidatatang_200zhAsrDataModule.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 = k2.SymbolTable.from_file(args.tokens) - assert token_table[0] == "" - - 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 - - aidatatang_200zh = Aidatatang_200zhAsrDataModule(args) - - def remove_short_utt(c: Cut): - T = ((c.num_frames - 7) // 2 + 1) // 2 - if T <= 0: - logging.warning( - f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}." - ) - return T > 0 - - dev_cuts = aidatatang_200zh.valid_cuts() - dev_cuts = dev_cuts.filter(remove_short_utt) - dev_dl = aidatatang_200zh.valid_dataloaders(dev_cuts) - - test_cuts = aidatatang_200zh.test_cuts() - test_cuts = test_cuts.filter(remove_short_utt) - test_dl = aidatatang_200zh.test_dataloaders(test_cuts) - - test_sets = ["dev", "test"] - test_dl = [dev_dl, test_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/aidatatang_200zh/ASR/zipformer/onnx_pretrained-streaming.py b/egs/aidatatang_200zh/ASR/zipformer/onnx_pretrained-streaming.py deleted file mode 120000 index cfea104c2..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/onnx_pretrained-streaming.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/onnx_pretrained-streaming.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/onnx_pretrained.py b/egs/aidatatang_200zh/ASR/zipformer/onnx_pretrained.py deleted file mode 120000 index 8f32f4ee7..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/onnx_pretrained.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/onnx_pretrained.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/optim.py b/egs/aidatatang_200zh/ASR/zipformer/optim.py deleted file mode 120000 index 5eaa3cffd..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/optim.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/optim.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/pretrained.py b/egs/aidatatang_200zh/ASR/zipformer/pretrained.py deleted file mode 120000 index 0bd71dde4..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/pretrained.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/pretrained.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/scaling.py b/egs/aidatatang_200zh/ASR/zipformer/scaling.py deleted file mode 120000 index 6f398f431..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/scaling.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/scaling.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/scaling_converter.py b/egs/aidatatang_200zh/ASR/zipformer/scaling_converter.py deleted file mode 120000 index b0ecee05e..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/scaling_converter.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/scaling_converter.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/streaming_beam_search.py b/egs/aidatatang_200zh/ASR/zipformer/streaming_beam_search.py deleted file mode 120000 index b1ed54557..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/streaming_beam_search.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/streaming_beam_search.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/streaming_decode.py b/egs/aidatatang_200zh/ASR/zipformer/streaming_decode.py deleted file mode 100755 index 23c11d9b9..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/streaming_decode.py +++ /dev/null @@ -1,880 +0,0 @@ -#!/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 16 \ - --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 torch -from asr_datamodule import Aidatatang_200zhAsrDataModule -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_model, get_params - -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=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( - "--lang-dir", - type=str, - default="data/lang_char", - help="Path to the lang dir(containing lexicon, tokens, etc.)", - ) - - 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( - "--blank-penalty", - type=float, - default=0.0, - help=""" - The penalty applied on blank symbol during decoding. - Note: It is a positive value that would be applied to logits like - this `logits[:, 0] -= blank_penalty` (suppose logits.shape is - [batch_size, vocab] and blank id is 0). - """, - ) - - 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, - blank_penalty=params.blank_penalty, - ) - 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, - blank_penalty=params.blank_penalty, - ) - 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, - blank_penalty=params.blank_penalty, - ) - 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, - lexicon: Lexicon, - 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. - lexicon: - The Lexicon. - 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 - if audio.max() > 1: - logging.warning( - f"The audio should be normalized to [-1, 1], audio.max : {audio.max()}." - f"Clipping to [-1, 1]." - ) - audio = np.clip(audio, -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, - list(decode_streams[i].ground_truth.strip()), - [ - lexicon.token_table[idx] - for idx in decode_streams[i].decoding_result() - ], - ) - ) - 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(), - [ - lexicon.token_table[idx] - for idx in decode_streams[i].decoding_result() - ], - ) - ) - del decode_streams[i] - - key = f"blank_penalty_{params.blank_penalty}" - if params.decoding_method == "greedy_search": - key = f"greedy_search_{key}" - elif params.decoding_method == "fast_beam_search": - key = ( - f"beam_{params.beam}_" - f"max_contexts_{params.max_contexts}_" - f"max_states_{params.max_states}_{key}" - ) - elif params.decoding_method == "modified_beam_search": - key = f"num_active_paths_{params.num_active_paths}_{key}" - 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() - Aidatatang_200zhAsrDataModule.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}" - params.suffix += f"-blank-penalty-{params.blank_penalty}" - - # 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}") - - lexicon = Lexicon(params.lang_dir) - params.blank_id = lexicon.token_table[""] - params.vocab_size = max(lexicon.tokens) + 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.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}") - - aidatatang_200zh = Aidatatang_200zhAsrDataModule(args) - - dev_cuts = aidatatang_200zh.valid_cuts() - test_cuts = aidatatang_200zh.test_cuts() - - test_sets = ["valid_cuts", "test_cuts"] - test_cuts = [dev_cuts, test_cuts] - - for test_set, test_cut in zip(test_sets, test_cuts): - results_dict = decode_dataset( - cuts=test_cut, - params=params, - model=model, - lexicon=lexicon, - 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/aidatatang_200zh/ASR/zipformer/subsampling.py b/egs/aidatatang_200zh/ASR/zipformer/subsampling.py deleted file mode 120000 index 01ae9002c..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/subsampling.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/subsampling.py \ No newline at end of file diff --git a/egs/aidatatang_200zh/ASR/zipformer/train.py b/egs/aidatatang_200zh/ASR/zipformer/train.py deleted file mode 100755 index 78987ecc8..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/train.py +++ /dev/null @@ -1,1350 +0,0 @@ -#!/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="0,1,2,3,4,5,6,7" - -./zipformer/train.py \ - --world-size 8 \ - --num-epochs 12 \ - --start-epoch 1 \ - --exp-dir zipformer/exp \ - --training-subset L - --lr-epochs 1.5 \ - --max-duration 350 - -# For mix precision training: - -./zipformer/train.py \ - --world-size 8 \ - --num-epochs 12 \ - --start-epoch 1 \ - --use-fp16 1 \ - --exp-dir zipformer/exp \ - --training-subset L \ - --lr-epochs 1.5 \ - --max-duration 750 - -""" - - -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 torch -import torch.multiprocessing as mp -import torch.nn as nn -from asr_datamodule import Aidatatang_200zhAsrDataModule -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.char_graph_compiler import CharCtcTrainingGraphCompiler -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.lexicon import Lexicon -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.""", - ) - - -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( - "--lang-dir", - type=str, - default="data/lang_char", - help="""The lang dir - It contains language related input files such as - "lexicon.txt" - """, - ) - - 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( - "--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 0. - """, - ) - - 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, - # 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: - encoder_embed = get_encoder_embed(params) - encoder = get_encoder_model(params) - decoder = get_decoder_model(params) - joiner = get_joiner_model(params) - - model = AsrModel( - encoder_embed=encoder_embed, - encoder=encoder, - decoder=decoder, - joiner=joiner, - encoder_dim=int(max(params.encoder_dim.split(","))), - decoder_dim=params.decoder_dim, - vocab_size=params.vocab_size, - ) - 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"] - - if "cur_batch_idx" in saved_params: - params["cur_batch_idx"] = saved_params["cur_batch_idx"] - - return saved_params - - -def save_checkpoint( - params: AttributeDict, - model: 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], - graph_compiler: CharCtcTrainingGraphCompiler, - batch: dict, - is_training: bool, -) -> Tuple[Tensor, MetricsTracker]: - """ - Compute CTC 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 = graph_compiler.texts_to_ids(texts) - y = k2.RaggedTensor(y).to(device) - - with torch.set_grad_enabled(is_training): - simple_loss, pruned_loss, _ = model( - x=feature, - x_lens=feature_lens, - y=y, - prune_range=params.prune_range, - am_scale=params.am_scale, - lm_scale=params.lm_scale, - ) - - 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 - - assert loss.requires_grad == is_training - - info = MetricsTracker() - with warnings.catch_warnings(): - warnings.simplefilter("ignore") - info["frames"] = (feature_lens // params.subsampling_factor).sum().item() - - # Note: We use reduction=sum while computing the loss. - info["loss"] = loss.detach().cpu().item() - info["simple_loss"] = simple_loss.detach().cpu().item() - info["pruned_loss"] = pruned_loss.detach().cpu().item() - - return loss, info - - -def compute_validation_loss( - params: AttributeDict, - model: Union[nn.Module, DDP], - graph_compiler: CharCtcTrainingGraphCompiler, - 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, - graph_compiler=graph_compiler, - 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, - graph_compiler: CharCtcTrainingGraphCompiler, - 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() - - cur_batch_idx = params.get("cur_batch_idx", 0) - - 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)) - if batch_idx < cur_batch_idx: - continue - cur_batch_idx = batch_idx - - params.batch_idx_train += 1 - batch_size = len(batch["supervisions"]["text"]) - - try: - with torch.cuda.amp.autocast(enabled=params.use_fp16): - loss, loss_info = compute_loss( - params=params, - model=model, - graph_compiler=graph_compiler, - 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, graph_compiler=graph_compiler) - 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 - ): - params.cur_batch_idx = batch_idx - save_checkpoint_with_global_batch_idx( - out_dir=params.exp_dir, - global_batch_idx=params.batch_idx_train, - model=model, - model_avg=model_avg, - params=params, - optimizer=optimizer, - scheduler=scheduler, - sampler=train_dl.sampler, - scaler=scaler, - rank=rank, - ) - del params.cur_batch_idx - remove_checkpoints( - out_dir=params.exp_dir, - topk=params.keep_last_k, - rank=rank, - ) - - if batch_idx % 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, - graph_compiler=graph_compiler, - 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}") - - lexicon = Lexicon(params.lang_dir) - graph_compiler = CharCtcTrainingGraphCompiler( - lexicon=lexicon, - device=device, - ) - - params.blank_id = lexicon.token_table[""] - params.vocab_size = max(lexicon.tokens) + 1 - - 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( - 2**22 - ) # allow 4 megabytes per sub-module - diagnostic = diagnostics.attach_diagnostics(model, opts) - - if params.inf_check: - register_inf_check_hooks(model) - - aidatatang_200zh = Aidatatang_200zhAsrDataModule(args) - - train_cuts = aidatatang_200zh.train_cuts() - valid_cuts = aidatatang_200zh.valid_cuts() - - def remove_short_and_long_utt(c: Cut): - # Keep only utterances with duration between 1 second and 15 seconds - # - # Caution: There is a reason to select 15.0 here. Please see - # ../local/display_manifest_statistics.py - # - # You should use ../local/display_manifest_statistics.py to get - # an utterance duration distribution for your dataset to select - # the threshold - if c.duration < 1.0 or c.duration > 10.0: - # logging.warning( - # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" - # ) - return False - - # In pruned RNN-T, we require that T >= S - # where T is the number of feature frames after subsampling - # and S is the number of tokens in the utterance - - # In ./zipformer.py, the conv module uses the following expression - # for subsampling - T = ((c.num_frames - 7) // 2 + 1) // 2 - tokens = graph_compiler.texts_to_ids([c.supervisions[0].text])[0] - - if T < len(tokens): - logging.warning( - f"Exclude cut with ID {c.id} from training. " - f"Number of frames (before subsampling): {c.num_frames}. " - f"Number of frames (after subsampling): {T}. " - f"Text: {c.supervisions[0].text}. " - f"Tokens: {tokens}. " - f"Number of tokens: {len(tokens)}" - ) - return False - - return True - - train_cuts = train_cuts.filter(remove_short_and_long_utt) - - if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: - # We only load the sampler's state dict when it loads a checkpoint - # saved in the middle of an epoch - sampler_state_dict = checkpoints["sampler"] - else: - sampler_state_dict = None - - train_dl = aidatatang_200zh.train_dataloaders( - train_cuts, sampler_state_dict=sampler_state_dict - ) - - valid_dl = aidatatang_200zh.valid_dataloaders(valid_cuts) - - if False and not params.print_diagnostics: - scan_pessimistic_batches_for_oom( - model=model, - train_dl=train_dl, - optimizer=optimizer, - graph_compiler=graph_compiler, - 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, - graph_compiler=graph_compiler, - 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, - graph_compiler: CharCtcTrainingGraphCompiler, -) -> 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`. - graph_compiler: - The compiler to encode texts to ids. - """ - 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}") - - texts = supervisions["text"] - y = graph_compiler.texts_to_ids(texts) - 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, - graph_compiler: CharCtcTrainingGraphCompiler, - 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, - graph_compiler=graph_compiler, - 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, graph_compiler=graph_compiler) - raise - logging.info( - f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" - ) - - -def main(): - parser = get_parser() - Aidatatang_200zhAsrDataModule.add_arguments(parser) - args = parser.parse_args() - args.lang_dir = Path(args.lang_dir) - 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/aidatatang_200zh/ASR/zipformer/zipformer.py b/egs/aidatatang_200zh/ASR/zipformer/zipformer.py deleted file mode 120000 index 23011dda7..000000000 --- a/egs/aidatatang_200zh/ASR/zipformer/zipformer.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/zipformer.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/__init__.py b/egs/aishell2/ASR/zipformer/__init__.py deleted file mode 100644 index e69de29bb..000000000 diff --git a/egs/aishell2/ASR/zipformer/asr_datamodule.py b/egs/aishell2/ASR/zipformer/asr_datamodule.py deleted file mode 120000 index c473a600a..000000000 --- a/egs/aishell2/ASR/zipformer/asr_datamodule.py +++ /dev/null @@ -1 +0,0 @@ -../pruned_transducer_stateless5/asr_datamodule.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/beam_search.py b/egs/aishell2/ASR/zipformer/beam_search.py deleted file mode 120000 index 4eef3d295..000000000 --- a/egs/aishell2/ASR/zipformer/beam_search.py +++ /dev/null @@ -1 +0,0 @@ -../pruned_transducer_stateless5/beam_search.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/decode.py b/egs/aishell2/ASR/zipformer/decode.py deleted file mode 100755 index 2f0edfe57..000000000 --- a/egs/aishell2/ASR/zipformer/decode.py +++ /dev/null @@ -1,814 +0,0 @@ -#!/usr/bin/env python3 -# -# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, -# Zengwei Yao -# Mingshuang Luo) -# -# See ../../../../LICENSE for clarification regarding multiple authors -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -""" -Usage: -(1) greedy search -./zipformer/decode.py \ - --epoch 35 \ - --avg 15 \ - --exp-dir ./zipformer/exp \ - --lang-dir data/lang_char \ - --max-duration 600 \ - --decoding-method greedy_search - -(2) modified beam search -./zipformer/decode.py \ - --epoch 35 \ - --avg 15 \ - --exp-dir ./zipformer/exp \ - --lang-dir data/lang_char \ - --max-duration 600 \ - --decoding-method modified_beam_search \ - --beam-size 4 - -(3) fast beam search (trivial_graph) -./zipformer/decode.py \ - --epoch 35 \ - --avg 15 \ - --exp-dir ./zipformer/exp \ - --lang-dir data/lang_char \ - --max-duration 600 \ - --decoding-method fast_beam_search \ - --beam 20.0 \ - --max-contexts 8 \ - --max-states 64 - -(4) fast beam search (LG) -./zipformer/decode.py \ - --epoch 30 \ - --avg 15 \ - --exp-dir ./zipformer/exp \ - --lang-dir data/lang_char \ - --max-duration 600 \ - --decoding-method fast_beam_search_LG \ - --beam 20.0 \ - --max-contexts 8 \ - --max-states 64 - -(5) fast beam search (nbest oracle WER) -./zipformer/decode.py \ - --epoch 35 \ - --avg 15 \ - --exp-dir ./zipformer/exp \ - --lang-dir data/lang_char \ - --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 -""" - - -import argparse -import logging -import math -from collections import defaultdict -from pathlib import Path -from typing import Dict, List, Optional, Tuple - -import k2 -import torch -import torch.nn as nn -from asr_datamodule import AiShell2AsrDataModule -from beam_search import ( - beam_search, - fast_beam_search_nbest, - fast_beam_search_nbest_LG, - fast_beam_search_nbest_oracle, - fast_beam_search_one_best, - greedy_search, - greedy_search_batch, - modified_beam_search, -) -from lhotse.cut import Cut -from train import add_model_arguments, get_model, get_params - -from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler -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( - "--lang-dir", - type=Path, - default="data/lang_char", - 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 - - modified_beam_search - - fast_beam_search - - fast_beam_search_LG - - fast_beam_search_nbest_oracle - If you use fast_beam_search_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, fast_beam_search_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_LG. - It specifies the scale for n-gram LM scores. - """, - ) - - parser.add_argument( - "--ilme-scale", - type=float, - default=0.2, - help=""" - Used only when --decoding_method is fast_beam_search_LG. - It specifies the scale for the internal language model estimation. - """, - ) - - parser.add_argument( - "--max-contexts", - type=int, - default=8, - help="""Used only when --decoding-method is - fast_beam_search, fast_beam_search, fast_beam_search_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, fast_beam_search_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_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 and fast_beam_search_nbest_oracle""", - ) - - parser.add_argument( - "--blank-penalty", - type=float, - default=0.0, - help=""" - The penalty applied on blank symbol during decoding. - Note: It is a positive value that would be applied to logits like - this `logits[:, 0] -= blank_penalty` (suppose logits.shape is - [batch_size, vocab] and blank id is 0). - """, - ) - - add_model_arguments(parser) - - return parser - - -def decode_one_batch( - params: AttributeDict, - model: nn.Module, - lexicon: Lexicon, - graph_compiler: CharCtcTrainingGraphCompiler, - batch: dict, - decoding_graph: Optional[k2.Fsa] = None, -) -> Dict[str, List[List[str]]]: - """Decode one batch and return the result in a dict. The dict has the - following format: - - - key: It indicates the setting used for decoding. For example, - if greedy_search is used, it would be "greedy_search" - If beam search with a beam size of 7 is used, it would be - "beam_7" - - value: It contains the decoding result. `len(value)` equals to - batch size. `value[i]` is the decoding result for the i-th - utterance in the given batch. - Args: - params: - It's the return value of :func:`get_params`. - model: - The neural model. - batch: - It is the return value from iterating - `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation - for the format of the `batch`. - decoding_graph: - The decoding graph. Can be either a `k2.trivial_graph` or LG, Used - only when --decoding_method is fast_beam_search, fast_beam_search_nbest, - fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. - Returns: - Return the decoding result. See above description for the format of - the returned dict. - """ - device = 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, - ) - - x, x_lens = model.encoder_embed(feature, feature_lens) - - 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 = model.encoder(x, x_lens, src_key_padding_mask) - encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) - - 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, - blank_penalty=params.blank_penalty, - ) - for i in range(encoder_out.size(0)): - hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) - elif params.decoding_method == "fast_beam_search_LG": - 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, - blank_penalty=params.blank_penalty, - ilme_scale=params.ilme_scale, - ) - for hyp in hyp_tokens: - sentence = "".join([lexicon.word_table[i] for i in hyp]) - hyps.append(list(sentence)) - 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=graph_compiler.texts_to_ids(supervisions["text"]), - nbest_scale=params.nbest_scale, - blank_penalty=params.blank_penalty, - ) - for i in range(encoder_out.size(0)): - hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) - 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, - blank_penalty=params.blank_penalty, - ) - for i in range(encoder_out.size(0)): - hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) - elif params.decoding_method == "modified_beam_search": - hyp_tokens = modified_beam_search( - model=model, - encoder_out=encoder_out, - encoder_out_lens=encoder_out_lens, - blank_penalty=params.blank_penalty, - beam=params.beam_size, - ) - for i in range(encoder_out.size(0)): - hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) - 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, - blank_penalty=params.blank_penalty, - ) - elif params.decoding_method == "beam_search": - hyp = beam_search( - model=model, - encoder_out=encoder_out_i, - beam=params.beam_size, - blank_penalty=params.blank_penalty, - ) - else: - raise ValueError( - f"Unsupported decoding method: {params.decoding_method}" - ) - hyps.append([lexicon.token_table[idx] for idx in hyp]) - - key = f"blank_penalty_{params.blank_penalty}" - if params.decoding_method == "greedy_search": - return {"greedy_search_" + key: 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"_ilme_scale_{params.ilme_scale}" - key += f"_ngram_lm_scale_{params.ngram_lm_scale}" - - return {key: hyps} - else: - return {f"beam_size_{params.beam_size}_" + key: hyps} - - -def decode_dataset( - dl: torch.utils.data.DataLoader, - params: AttributeDict, - model: nn.Module, - lexicon: Lexicon, - graph_compiler: CharCtcTrainingGraphCompiler, - decoding_graph: Optional[k2.Fsa] = None, -) -> Dict[str, List[Tuple[List[str], List[str]]]]: - """Decode dataset. - - Args: - dl: - PyTorch's dataloader containing the dataset to decode. - params: - It is returned by :func:`get_params`. - model: - The neural model. - decoding_graph: - The decoding graph. Can be either a `k2.trivial_graph` or LG, 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"] - texts = [list("".join(text.split())) for text in texts] - cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] - - hyps_dict = decode_one_batch( - params=params, - model=model, - lexicon=lexicon, - graph_compiler=graph_compiler, - decoding_graph=decoding_graph, - batch=batch, - ) - - 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): - this_batch.append((cut_id, ref_text, hyp_words)) - - results[name].extend(this_batch) - - num_cuts += len(texts) - - if batch_idx % log_interval == 0: - batch_str = f"{batch_idx}/{num_batches}" - - logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") - return results - - -def save_results( - params: AttributeDict, - test_set_name: str, - results_dict: Dict[str, List[Tuple[List[int], List[int]]]], -): - test_set_wers = dict() - for key, results in results_dict.items(): - recog_path = ( - params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" - ) - 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() - AiShell2AsrDataModule.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", - "modified_beam_search", - "fast_beam_search", - "fast_beam_search_LG", - "fast_beam_search_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 "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"_ilme_scale_{params.ilme_scale}" - params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" - elif "beam_search" in params.decoding_method: - params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" - else: - params.suffix += f"-context-{params.context_size}" - params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" - params.suffix += f"-blank-penalty-{params.blank_penalty}" - - 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}") - - lexicon = Lexicon(params.lang_dir) - params.blank_id = lexicon.token_table[""] - params.vocab_size = max(lexicon.tokens) + 1 - - graph_compiler = CharCtcTrainingGraphCompiler( - lexicon=lexicon, - device=device, - ) - - 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() - - if "fast_beam_search" in params.decoding_method: - if "LG" in params.decoding_method: - lexicon = Lexicon(params.lang_dir) - 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: - decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) - else: - decoding_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 - aishell2 = AiShell2AsrDataModule(args) - - def remove_short_utt(c: Cut): - T = ((c.num_frames - 7) // 2 + 1) // 2 - if T <= 0: - logging.warning( - f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}." - ) - return T > 0 - - dev_cuts = aishell2.valid_cuts() - dev_cuts = dev_cuts.filter(remove_short_utt) - dev_dl = aishell2.valid_dataloaders(dev_cuts) - - test_cuts = aishell2.test_cuts() - test_cuts = test_cuts.filter(remove_short_utt) - test_dl = aishell2.test_dataloaders(test_cuts) - - test_sets = ["dev", "test"] - test_dls = [dev_dl, test_dl] - - for test_set, test_dl in zip(test_sets, test_dls): - results_dict = decode_dataset( - dl=test_dl, - params=params, - model=model, - lexicon=lexicon, - graph_compiler=graph_compiler, - 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/aishell2/ASR/zipformer/decode_stream.py b/egs/aishell2/ASR/zipformer/decode_stream.py deleted file mode 120000 index b8d8ddfc4..000000000 --- a/egs/aishell2/ASR/zipformer/decode_stream.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/decode_stream.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/decoder.py b/egs/aishell2/ASR/zipformer/decoder.py deleted file mode 120000 index 5a8018680..000000000 --- a/egs/aishell2/ASR/zipformer/decoder.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/decoder.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/encoder_interface.py b/egs/aishell2/ASR/zipformer/encoder_interface.py deleted file mode 120000 index 083f693ef..000000000 --- a/egs/aishell2/ASR/zipformer/encoder_interface.py +++ /dev/null @@ -1 +0,0 @@ -../pruned_transducer_stateless5/encoder_interface.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/export-onnx-streaming.py b/egs/aishell2/ASR/zipformer/export-onnx-streaming.py deleted file mode 120000 index 2962eb784..000000000 --- a/egs/aishell2/ASR/zipformer/export-onnx-streaming.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/export-onnx-streaming.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/export-onnx.py b/egs/aishell2/ASR/zipformer/export-onnx.py deleted file mode 120000 index 70a15683c..000000000 --- a/egs/aishell2/ASR/zipformer/export-onnx.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/export-onnx.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/export.py b/egs/aishell2/ASR/zipformer/export.py deleted file mode 120000 index dfc1bec08..000000000 --- a/egs/aishell2/ASR/zipformer/export.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/export.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/jit_pretrained.py b/egs/aishell2/ASR/zipformer/jit_pretrained.py deleted file mode 120000 index 25108391f..000000000 --- a/egs/aishell2/ASR/zipformer/jit_pretrained.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/jit_pretrained.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/jit_pretrained_streaming.py b/egs/aishell2/ASR/zipformer/jit_pretrained_streaming.py deleted file mode 120000 index 1962351e9..000000000 --- a/egs/aishell2/ASR/zipformer/jit_pretrained_streaming.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/jit_pretrained_streaming.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/joiner.py b/egs/aishell2/ASR/zipformer/joiner.py deleted file mode 120000 index 5b8a36332..000000000 --- a/egs/aishell2/ASR/zipformer/joiner.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/joiner.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/model.py b/egs/aishell2/ASR/zipformer/model.py deleted file mode 120000 index cd7e07d72..000000000 --- a/egs/aishell2/ASR/zipformer/model.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/model.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/onnx_check.py b/egs/aishell2/ASR/zipformer/onnx_check.py deleted file mode 120000 index f3dd42004..000000000 --- a/egs/aishell2/ASR/zipformer/onnx_check.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/onnx_check.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/onnx_decode.py b/egs/aishell2/ASR/zipformer/onnx_decode.py deleted file mode 100755 index d518f4e86..000000000 --- a/egs/aishell2/ASR/zipformer/onnx_decode.py +++ /dev/null @@ -1,286 +0,0 @@ -#!/usr/bin/env python3 -# -# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, -# Zengwei Yao, -# Xiaoyu Yang, -# 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 loads ONNX exported models and uses them to decode the test sets. -""" - -import argparse -import logging -import time -from pathlib import Path -from typing import List, Tuple - -import k2 -import torch -import torch.nn as nn -from asr_datamodule import AiShell2AsrDataModule -from lhotse.cut import Cut -from onnx_pretrained import OnnxModel, greedy_search - -from icefall.utils import setup_logger, store_transcripts, write_error_stats - - -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="pruned_transducer_stateless7/exp", - help="The experiment dir", - ) - - parser.add_argument( - "--tokens", - type=str, - default="data/lang_char/tokens.txt", - help="Path to the 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: k2.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: - Mapping ids to tokens. - 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 - ) - - hyps = [[token_table[h] for h in hyp] for hyp in hyps] - return hyps - - -def decode_dataset( - dl: torch.utils.data.DataLoader, - model: nn.Module, - token_table: k2.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: - Mapping ids to tokens. - - 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 = list(ref_text) - 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() - AiShell2AsrDataModule.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 = k2.SymbolTable.from_file(args.tokens) - assert token_table[0] == "" - - 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 - - aishell2 = AiShell2AsrDataModule(args) - - def remove_short_utt(c: Cut): - T = ((c.num_frames - 7) // 2 + 1) // 2 - if T <= 0: - logging.warning( - f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}." - ) - return T > 0 - - dev_cuts = aishell2.valid_cuts() - dev_cuts = dev_cuts.filter(remove_short_utt) - dev_dl = aishell2.valid_dataloaders(dev_cuts) - - test_cuts = aishell2.test_net_cuts() - test_cuts = test_cuts.filter(remove_short_utt) - test_dl = aishell2.test_dataloaders(test_cuts) - - test_sets = ["dev", "test"] - test_dl = [dev_dl, test_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/aishell2/ASR/zipformer/onnx_pretrained-streaming.py b/egs/aishell2/ASR/zipformer/onnx_pretrained-streaming.py deleted file mode 120000 index cfea104c2..000000000 --- a/egs/aishell2/ASR/zipformer/onnx_pretrained-streaming.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/onnx_pretrained-streaming.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/onnx_pretrained.py b/egs/aishell2/ASR/zipformer/onnx_pretrained.py deleted file mode 120000 index 8f32f4ee7..000000000 --- a/egs/aishell2/ASR/zipformer/onnx_pretrained.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/onnx_pretrained.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/optim.py b/egs/aishell2/ASR/zipformer/optim.py deleted file mode 120000 index 5eaa3cffd..000000000 --- a/egs/aishell2/ASR/zipformer/optim.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/optim.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/pretrained.py b/egs/aishell2/ASR/zipformer/pretrained.py deleted file mode 120000 index 0bd71dde4..000000000 --- a/egs/aishell2/ASR/zipformer/pretrained.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/pretrained.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/scaling.py b/egs/aishell2/ASR/zipformer/scaling.py deleted file mode 120000 index 6f398f431..000000000 --- a/egs/aishell2/ASR/zipformer/scaling.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/scaling.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/scaling_converter.py b/egs/aishell2/ASR/zipformer/scaling_converter.py deleted file mode 120000 index b0ecee05e..000000000 --- a/egs/aishell2/ASR/zipformer/scaling_converter.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/scaling_converter.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/streaming_beam_search.py b/egs/aishell2/ASR/zipformer/streaming_beam_search.py deleted file mode 120000 index b1ed54557..000000000 --- a/egs/aishell2/ASR/zipformer/streaming_beam_search.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/streaming_beam_search.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/streaming_decode.py b/egs/aishell2/ASR/zipformer/streaming_decode.py deleted file mode 100755 index 1b43a9366..000000000 --- a/egs/aishell2/ASR/zipformer/streaming_decode.py +++ /dev/null @@ -1,880 +0,0 @@ -#!/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 16 \ - --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 torch -from asr_datamodule import AiShell2AsrDataModule -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_model, get_params - -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=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( - "--lang-dir", - type=str, - default="data/lang_char", - help="Path to the lang dir(containing lexicon, tokens, etc.)", - ) - - 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( - "--blank-penalty", - type=float, - default=0.0, - help=""" - The penalty applied on blank symbol during decoding. - Note: It is a positive value that would be applied to logits like - this `logits[:, 0] -= blank_penalty` (suppose logits.shape is - [batch_size, vocab] and blank id is 0). - """, - ) - - 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, - blank_penalty=params.blank_penalty, - ) - 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, - blank_penalty=params.blank_penalty, - ) - 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, - blank_penalty=params.blank_penalty, - ) - 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, - lexicon: Lexicon, - 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. - lexicon: - The Lexicon. - 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 - if audio.max() > 1: - logging.warning( - f"The audio should be normalized to [-1, 1], audio.max : {audio.max()}." - f"Clipping to [-1, 1]." - ) - audio = np.clip(audio, -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, - list(decode_streams[i].ground_truth.strip()), - [ - lexicon.token_table[idx] - for idx in decode_streams[i].decoding_result() - ], - ) - ) - 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(), - [ - lexicon.token_table[idx] - for idx in decode_streams[i].decoding_result() - ], - ) - ) - del decode_streams[i] - - key = f"blank_penalty_{params.blank_penalty}" - if params.decoding_method == "greedy_search": - key = f"greedy_search_{key}" - elif params.decoding_method == "fast_beam_search": - key = ( - f"beam_{params.beam}_" - f"max_contexts_{params.max_contexts}_" - f"max_states_{params.max_states}_{key}" - ) - elif params.decoding_method == "modified_beam_search": - key = f"num_active_paths_{params.num_active_paths}_{key}" - 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() - AiShell2AsrDataModule.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}" - params.suffix += f"-blank-penalty-{params.blank_penalty}" - - # 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}") - - lexicon = Lexicon(params.lang_dir) - params.blank_id = lexicon.token_table[""] - params.vocab_size = max(lexicon.tokens) + 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.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}") - - aishell2 = AiShell2AsrDataModule(args) - - dev_cuts = aishell2.valid_cuts() - test_cuts = aishell2.test_cuts() - - test_sets = ["dev", "test"] - test_cuts = [dev_cuts, test_cuts] - - for test_set, test_cut in zip(test_sets, test_cuts): - results_dict = decode_dataset( - cuts=test_cut, - params=params, - model=model, - lexicon=lexicon, - 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/aishell2/ASR/zipformer/subsampling.py b/egs/aishell2/ASR/zipformer/subsampling.py deleted file mode 120000 index 01ae9002c..000000000 --- a/egs/aishell2/ASR/zipformer/subsampling.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/subsampling.py \ No newline at end of file diff --git a/egs/aishell2/ASR/zipformer/train.py b/egs/aishell2/ASR/zipformer/train.py deleted file mode 100755 index 1c5869ef0..000000000 --- a/egs/aishell2/ASR/zipformer/train.py +++ /dev/null @@ -1,1350 +0,0 @@ -#!/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="0,1,2,3,4,5,6,7" - -./zipformer/train.py \ - --world-size 8 \ - --num-epochs 12 \ - --start-epoch 1 \ - --exp-dir zipformer/exp \ - --training-subset L - --lr-epochs 1.5 \ - --max-duration 350 - -# For mix precision training: - -./zipformer/train.py \ - --world-size 8 \ - --num-epochs 12 \ - --start-epoch 1 \ - --use-fp16 1 \ - --exp-dir zipformer/exp \ - --training-subset L \ - --lr-epochs 1.5 \ - --max-duration 750 - -""" - - -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 torch -import torch.multiprocessing as mp -import torch.nn as nn -from asr_datamodule import AiShell2AsrDataModule -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.char_graph_compiler import CharCtcTrainingGraphCompiler -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.lexicon import Lexicon -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.""", - ) - - -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( - "--lang-dir", - type=str, - default="data/lang_char", - help="""The lang dir - It contains language related input files such as - "lexicon.txt" - """, - ) - - 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( - "--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 0. - """, - ) - - 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, - # 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: - encoder_embed = get_encoder_embed(params) - encoder = get_encoder_model(params) - decoder = get_decoder_model(params) - joiner = get_joiner_model(params) - - model = AsrModel( - encoder_embed=encoder_embed, - encoder=encoder, - decoder=decoder, - joiner=joiner, - encoder_dim=int(max(params.encoder_dim.split(","))), - decoder_dim=params.decoder_dim, - vocab_size=params.vocab_size, - ) - 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"] - - if "cur_batch_idx" in saved_params: - params["cur_batch_idx"] = saved_params["cur_batch_idx"] - - return saved_params - - -def save_checkpoint( - params: AttributeDict, - model: 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], - graph_compiler: CharCtcTrainingGraphCompiler, - batch: dict, - is_training: bool, -) -> Tuple[Tensor, MetricsTracker]: - """ - Compute CTC 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 = graph_compiler.texts_to_ids(texts) - y = k2.RaggedTensor(y).to(device) - - with torch.set_grad_enabled(is_training): - simple_loss, pruned_loss, _ = model( - x=feature, - x_lens=feature_lens, - y=y, - prune_range=params.prune_range, - am_scale=params.am_scale, - lm_scale=params.lm_scale, - ) - - 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 - - assert loss.requires_grad == is_training - - info = MetricsTracker() - with warnings.catch_warnings(): - warnings.simplefilter("ignore") - info["frames"] = (feature_lens // params.subsampling_factor).sum().item() - - # Note: We use reduction=sum while computing the loss. - info["loss"] = loss.detach().cpu().item() - info["simple_loss"] = simple_loss.detach().cpu().item() - info["pruned_loss"] = pruned_loss.detach().cpu().item() - - return loss, info - - -def compute_validation_loss( - params: AttributeDict, - model: Union[nn.Module, DDP], - graph_compiler: CharCtcTrainingGraphCompiler, - 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, - graph_compiler=graph_compiler, - 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, - graph_compiler: CharCtcTrainingGraphCompiler, - 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() - - cur_batch_idx = params.get("cur_batch_idx", 0) - - 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)) - if batch_idx < cur_batch_idx: - continue - cur_batch_idx = batch_idx - - params.batch_idx_train += 1 - batch_size = len(batch["supervisions"]["text"]) - - try: - with torch.cuda.amp.autocast(enabled=params.use_fp16): - loss, loss_info = compute_loss( - params=params, - model=model, - graph_compiler=graph_compiler, - 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, graph_compiler=graph_compiler) - 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 - ): - params.cur_batch_idx = batch_idx - save_checkpoint_with_global_batch_idx( - out_dir=params.exp_dir, - global_batch_idx=params.batch_idx_train, - model=model, - model_avg=model_avg, - params=params, - optimizer=optimizer, - scheduler=scheduler, - sampler=train_dl.sampler, - scaler=scaler, - rank=rank, - ) - del params.cur_batch_idx - remove_checkpoints( - out_dir=params.exp_dir, - topk=params.keep_last_k, - rank=rank, - ) - - if batch_idx % 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, - graph_compiler=graph_compiler, - 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}") - - lexicon = Lexicon(params.lang_dir) - graph_compiler = CharCtcTrainingGraphCompiler( - lexicon=lexicon, - device=device, - ) - - params.blank_id = lexicon.token_table[""] - params.vocab_size = max(lexicon.tokens) + 1 - - 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( - 2**22 - ) # allow 4 megabytes per sub-module - diagnostic = diagnostics.attach_diagnostics(model, opts) - - if params.inf_check: - register_inf_check_hooks(model) - - aishell2 = AiShell2AsrDataModule(args) - - train_cuts = aishell2.train_cuts() - valid_cuts = aishell2.valid_cuts() - - def remove_short_and_long_utt(c: Cut): - # Keep only utterances with duration between 1 second and 15 seconds - # - # Caution: There is a reason to select 15.0 here. Please see - # ../local/display_manifest_statistics.py - # - # You should use ../local/display_manifest_statistics.py to get - # an utterance duration distribution for your dataset to select - # the threshold - if c.duration < 1.0 or c.duration > 8.0: - # logging.warning( - # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" - # ) - return False - - # In pruned RNN-T, we require that T >= S - # where T is the number of feature frames after subsampling - # and S is the number of tokens in the utterance - - # In ./zipformer.py, the conv module uses the following expression - # for subsampling - T = ((c.num_frames - 7) // 2 + 1) // 2 - tokens = graph_compiler.texts_to_ids([c.supervisions[0].text])[0] - - if T < len(tokens): - logging.warning( - f"Exclude cut with ID {c.id} from training. " - f"Number of frames (before subsampling): {c.num_frames}. " - f"Number of frames (after subsampling): {T}. " - f"Text: {c.supervisions[0].text}. " - f"Tokens: {tokens}. " - f"Number of tokens: {len(tokens)}" - ) - return False - - return True - - train_cuts = train_cuts.filter(remove_short_and_long_utt) - - if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: - # We only load the sampler's state dict when it loads a checkpoint - # saved in the middle of an epoch - sampler_state_dict = checkpoints["sampler"] - else: - sampler_state_dict = None - - train_dl = aishell2.train_dataloaders( - train_cuts, sampler_state_dict=sampler_state_dict - ) - - valid_dl = aishell2.valid_dataloaders(valid_cuts) - - if False and not params.print_diagnostics: - scan_pessimistic_batches_for_oom( - model=model, - train_dl=train_dl, - optimizer=optimizer, - graph_compiler=graph_compiler, - 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, - graph_compiler=graph_compiler, - 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, - graph_compiler: CharCtcTrainingGraphCompiler, -) -> 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`. - graph_compiler: - The compiler to encode texts to ids. - """ - 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}") - - texts = supervisions["text"] - y = graph_compiler.texts_to_ids(texts) - 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, - graph_compiler: CharCtcTrainingGraphCompiler, - 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, - graph_compiler=graph_compiler, - 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, graph_compiler=graph_compiler) - raise - logging.info( - f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" - ) - - -def main(): - parser = get_parser() - AiShell2AsrDataModule.add_arguments(parser) - args = parser.parse_args() - args.lang_dir = Path(args.lang_dir) - 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/aishell2/ASR/zipformer/zipformer.py b/egs/aishell2/ASR/zipformer/zipformer.py deleted file mode 120000 index 23011dda7..000000000 --- a/egs/aishell2/ASR/zipformer/zipformer.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/zipformer.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/__init__.py b/egs/aishell4/ASR/zipformer/__init__.py deleted file mode 100644 index e69de29bb..000000000 diff --git a/egs/aishell4/ASR/zipformer/asr_datamodule.py b/egs/aishell4/ASR/zipformer/asr_datamodule.py deleted file mode 120000 index c473a600a..000000000 --- a/egs/aishell4/ASR/zipformer/asr_datamodule.py +++ /dev/null @@ -1 +0,0 @@ -../pruned_transducer_stateless5/asr_datamodule.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/beam_search.py b/egs/aishell4/ASR/zipformer/beam_search.py deleted file mode 120000 index 4eef3d295..000000000 --- a/egs/aishell4/ASR/zipformer/beam_search.py +++ /dev/null @@ -1 +0,0 @@ -../pruned_transducer_stateless5/beam_search.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/decode.py b/egs/aishell4/ASR/zipformer/decode.py deleted file mode 100755 index 82d630696..000000000 --- a/egs/aishell4/ASR/zipformer/decode.py +++ /dev/null @@ -1,819 +0,0 @@ -#!/usr/bin/env python3 -# -# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, -# Zengwei Yao -# Mingshuang Luo) -# -# See ../../../../LICENSE for clarification regarding multiple authors -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -""" -Usage: -(1) greedy search -./zipformer/decode.py \ - --epoch 35 \ - --avg 15 \ - --exp-dir ./zipformer/exp \ - --lang-dir data/lang_char \ - --max-duration 600 \ - --decoding-method greedy_search - -(2) modified beam search -./zipformer/decode.py \ - --epoch 35 \ - --avg 15 \ - --exp-dir ./zipformer/exp \ - --lang-dir data/lang_char \ - --max-duration 600 \ - --decoding-method modified_beam_search \ - --beam-size 4 - -(3) fast beam search (trivial_graph) -./zipformer/decode.py \ - --epoch 35 \ - --avg 15 \ - --exp-dir ./zipformer/exp \ - --lang-dir data/lang_char \ - --max-duration 600 \ - --decoding-method fast_beam_search \ - --beam 20.0 \ - --max-contexts 8 \ - --max-states 64 - -(4) fast beam search (LG) -./zipformer/decode.py \ - --epoch 30 \ - --avg 15 \ - --exp-dir ./zipformer/exp \ - --lang-dir data/lang_char \ - --max-duration 600 \ - --decoding-method fast_beam_search_LG \ - --beam 20.0 \ - --max-contexts 8 \ - --max-states 64 - -(5) fast beam search (nbest oracle WER) -./zipformer/decode.py \ - --epoch 35 \ - --avg 15 \ - --exp-dir ./zipformer/exp \ - --lang-dir data/lang_char \ - --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 -""" - - -import argparse -import logging -import math -from collections import defaultdict -from pathlib import Path -from typing import Dict, List, Optional, Tuple - -import k2 -import torch -import torch.nn as nn -from asr_datamodule import Aishell4AsrDataModule -from beam_search import ( - beam_search, - fast_beam_search_nbest, - fast_beam_search_nbest_LG, - fast_beam_search_nbest_oracle, - fast_beam_search_one_best, - greedy_search, - greedy_search_batch, - modified_beam_search, -) -from lhotse.cut import Cut -from local.text_normalize import text_normalize -from train import add_model_arguments, get_model, get_params - -from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler -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( - "--lang-dir", - type=Path, - default="data/lang_char", - 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 - - modified_beam_search - - fast_beam_search - - fast_beam_search_LG - - fast_beam_search_nbest_oracle - If you use fast_beam_search_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, fast_beam_search_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_LG. - It specifies the scale for n-gram LM scores. - """, - ) - - parser.add_argument( - "--ilme-scale", - type=float, - default=0.2, - help=""" - Used only when --decoding_method is fast_beam_search_LG. - It specifies the scale for the internal language model estimation. - """, - ) - - parser.add_argument( - "--max-contexts", - type=int, - default=8, - help="""Used only when --decoding-method is - fast_beam_search, fast_beam_search, fast_beam_search_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, fast_beam_search_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_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 and fast_beam_search_nbest_oracle""", - ) - - parser.add_argument( - "--blank-penalty", - type=float, - default=0.0, - help=""" - The penalty applied on blank symbol during decoding. - Note: It is a positive value that would be applied to logits like - this `logits[:, 0] -= blank_penalty` (suppose logits.shape is - [batch_size, vocab] and blank id is 0). - """, - ) - - add_model_arguments(parser) - - return parser - - -def decode_one_batch( - params: AttributeDict, - model: nn.Module, - lexicon: Lexicon, - graph_compiler: CharCtcTrainingGraphCompiler, - batch: dict, - decoding_graph: Optional[k2.Fsa] = None, -) -> Dict[str, List[List[str]]]: - """Decode one batch and return the result in a dict. The dict has the - following format: - - - key: It indicates the setting used for decoding. For example, - if greedy_search is used, it would be "greedy_search" - If beam search with a beam size of 7 is used, it would be - "beam_7" - - value: It contains the decoding result. `len(value)` equals to - batch size. `value[i]` is the decoding result for the i-th - utterance in the given batch. - Args: - params: - It's the return value of :func:`get_params`. - model: - The neural model. - batch: - It is the return value from iterating - `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation - for the format of the `batch`. - decoding_graph: - The decoding graph. Can be either a `k2.trivial_graph` or LG, Used - only when --decoding_method is fast_beam_search, fast_beam_search_nbest, - fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. - Returns: - Return the decoding result. See above description for the format of - the returned dict. - """ - device = 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, - ) - - x, x_lens = model.encoder_embed(feature, feature_lens) - - 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 = model.encoder(x, x_lens, src_key_padding_mask) - encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) - - 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, - blank_penalty=params.blank_penalty, - ) - for i in range(encoder_out.size(0)): - hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) - elif params.decoding_method == "fast_beam_search_LG": - 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, - blank_penalty=params.blank_penalty, - ilme_scale=params.ilme_scale, - ) - for hyp in hyp_tokens: - sentence = "".join([lexicon.word_table[i] for i in hyp]) - hyps.append(list(sentence)) - 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=graph_compiler.texts_to_ids(supervisions["text"]), - nbest_scale=params.nbest_scale, - blank_penalty=params.blank_penalty, - ) - for i in range(encoder_out.size(0)): - hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) - 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, - blank_penalty=params.blank_penalty, - ) - for i in range(encoder_out.size(0)): - hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) - elif params.decoding_method == "modified_beam_search": - hyp_tokens = modified_beam_search( - model=model, - encoder_out=encoder_out, - encoder_out_lens=encoder_out_lens, - blank_penalty=params.blank_penalty, - beam=params.beam_size, - ) - for i in range(encoder_out.size(0)): - hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) - 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, - blank_penalty=params.blank_penalty, - ) - elif params.decoding_method == "beam_search": - hyp = beam_search( - model=model, - encoder_out=encoder_out_i, - beam=params.beam_size, - blank_penalty=params.blank_penalty, - ) - else: - raise ValueError( - f"Unsupported decoding method: {params.decoding_method}" - ) - hyps.append([lexicon.token_table[idx] for idx in hyp]) - - key = f"blank_penalty_{params.blank_penalty}" - if params.decoding_method == "greedy_search": - return {"greedy_search_" + key: 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"_ilme_scale_{params.ilme_scale}" - key += f"_ngram_lm_scale_{params.ngram_lm_scale}" - - return {key: hyps} - else: - return {f"beam_size_{params.beam_size}_" + key: hyps} - - -def decode_dataset( - dl: torch.utils.data.DataLoader, - params: AttributeDict, - model: nn.Module, - lexicon: Lexicon, - graph_compiler: CharCtcTrainingGraphCompiler, - decoding_graph: Optional[k2.Fsa] = None, -) -> Dict[str, List[Tuple[List[str], List[str]]]]: - """Decode dataset. - - Args: - dl: - PyTorch's dataloader containing the dataset to decode. - params: - It is returned by :func:`get_params`. - model: - The neural model. - decoding_graph: - The decoding graph. Can be either a `k2.trivial_graph` or LG, 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"] - texts = [list("".join(text.split())) for text in texts] - cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] - - hyps_dict = decode_one_batch( - params=params, - model=model, - lexicon=lexicon, - graph_compiler=graph_compiler, - decoding_graph=decoding_graph, - batch=batch, - ) - - 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): - this_batch.append((cut_id, ref_text, hyp_words)) - - results[name].extend(this_batch) - - num_cuts += len(texts) - - if batch_idx % log_interval == 0: - batch_str = f"{batch_idx}/{num_batches}" - - logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") - return results - - -def save_results( - params: AttributeDict, - test_set_name: str, - results_dict: Dict[str, List[Tuple[List[int], List[int]]]], -): - test_set_wers = dict() - for key, results in results_dict.items(): - recog_path = ( - params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" - ) - 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() - Aishell4AsrDataModule.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", - "modified_beam_search", - "fast_beam_search", - "fast_beam_search_LG", - "fast_beam_search_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 "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"_ilme_scale_{params.ilme_scale}" - params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" - elif "beam_search" in params.decoding_method: - params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" - else: - params.suffix += f"-context-{params.context_size}" - params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" - params.suffix += f"-blank-penalty-{params.blank_penalty}" - - 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}") - - lexicon = Lexicon(params.lang_dir) - params.blank_id = lexicon.token_table[""] - params.vocab_size = max(lexicon.tokens) + 1 - - graph_compiler = CharCtcTrainingGraphCompiler( - lexicon=lexicon, - device=device, - ) - - 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() - - if "fast_beam_search" in params.decoding_method: - if "LG" in params.decoding_method: - lexicon = Lexicon(params.lang_dir) - 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: - decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) - else: - decoding_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 - aishell4 = Aishell4AsrDataModule(args) - - def remove_short_utt(c: Cut): - T = ((c.num_frames - 7) // 2 + 1) // 2 - if T <= 0: - logging.warning( - f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}." - ) - return T > 0 - - def text_normalize_for_cut(c: Cut): - # Text normalize for each sample - text = c.supervisions[0].text - text = text.strip("\n").strip("\t") - c.supervisions[0].text = text_normalize(text) - return c - - test_cuts = aishell4.test_cuts() - test_cuts = test_cuts.map(text_normalize_for_cut) - test_cuts = test_cuts.filter(remove_short_utt) - test_dl = aishell4.test_dataloaders(test_cuts) - - test_sets = ["test"] - test_dls = [test_dl] - - for test_set, test_dl in zip(test_sets, test_dls): - results_dict = decode_dataset( - dl=test_dl, - params=params, - model=model, - lexicon=lexicon, - graph_compiler=graph_compiler, - 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/aishell4/ASR/zipformer/decode_stream.py b/egs/aishell4/ASR/zipformer/decode_stream.py deleted file mode 120000 index b8d8ddfc4..000000000 --- a/egs/aishell4/ASR/zipformer/decode_stream.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/decode_stream.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/decoder.py b/egs/aishell4/ASR/zipformer/decoder.py deleted file mode 120000 index 5a8018680..000000000 --- a/egs/aishell4/ASR/zipformer/decoder.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/decoder.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/encoder_interface.py b/egs/aishell4/ASR/zipformer/encoder_interface.py deleted file mode 120000 index 083f693ef..000000000 --- a/egs/aishell4/ASR/zipformer/encoder_interface.py +++ /dev/null @@ -1 +0,0 @@ -../pruned_transducer_stateless5/encoder_interface.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/export-onnx-streaming.py b/egs/aishell4/ASR/zipformer/export-onnx-streaming.py deleted file mode 120000 index 2962eb784..000000000 --- a/egs/aishell4/ASR/zipformer/export-onnx-streaming.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/export-onnx-streaming.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/export-onnx.py b/egs/aishell4/ASR/zipformer/export-onnx.py deleted file mode 120000 index 70a15683c..000000000 --- a/egs/aishell4/ASR/zipformer/export-onnx.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/export-onnx.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/export.py b/egs/aishell4/ASR/zipformer/export.py deleted file mode 120000 index dfc1bec08..000000000 --- a/egs/aishell4/ASR/zipformer/export.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/export.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/jit_pretrained.py b/egs/aishell4/ASR/zipformer/jit_pretrained.py deleted file mode 120000 index 25108391f..000000000 --- a/egs/aishell4/ASR/zipformer/jit_pretrained.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/jit_pretrained.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/jit_pretrained_streaming.py b/egs/aishell4/ASR/zipformer/jit_pretrained_streaming.py deleted file mode 120000 index 1962351e9..000000000 --- a/egs/aishell4/ASR/zipformer/jit_pretrained_streaming.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/jit_pretrained_streaming.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/joiner.py b/egs/aishell4/ASR/zipformer/joiner.py deleted file mode 120000 index 5b8a36332..000000000 --- a/egs/aishell4/ASR/zipformer/joiner.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/joiner.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/local b/egs/aishell4/ASR/zipformer/local deleted file mode 120000 index c820590c5..000000000 --- a/egs/aishell4/ASR/zipformer/local +++ /dev/null @@ -1 +0,0 @@ -../local \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/model.py b/egs/aishell4/ASR/zipformer/model.py deleted file mode 120000 index cd7e07d72..000000000 --- a/egs/aishell4/ASR/zipformer/model.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/model.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/onnx_check.py b/egs/aishell4/ASR/zipformer/onnx_check.py deleted file mode 120000 index f3dd42004..000000000 --- a/egs/aishell4/ASR/zipformer/onnx_check.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/onnx_check.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/onnx_decode.py b/egs/aishell4/ASR/zipformer/onnx_decode.py deleted file mode 100755 index 6541bc1b9..000000000 --- a/egs/aishell4/ASR/zipformer/onnx_decode.py +++ /dev/null @@ -1,291 +0,0 @@ -#!/usr/bin/env python3 -# -# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, -# Zengwei Yao, -# Xiaoyu Yang, -# 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 loads ONNX exported models and uses them to decode the test sets. -""" - -import argparse -import logging -import time -from pathlib import Path -from typing import List, Tuple - -import k2 -import torch -import torch.nn as nn -from asr_datamodule import Aishell4AsrDataModule -from lhotse.cut import Cut -from local.text_normalize import text_normalize -from onnx_pretrained import OnnxModel, greedy_search - -from icefall.utils import setup_logger, store_transcripts, write_error_stats - - -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="pruned_transducer_stateless7/exp", - help="The experiment dir", - ) - - parser.add_argument( - "--tokens", - type=str, - default="data/lang_char/tokens.txt", - help="Path to the 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: k2.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: - Mapping ids to tokens. - 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 - ) - - hyps = [[token_table[h] for h in hyp] for hyp in hyps] - return hyps - - -def decode_dataset( - dl: torch.utils.data.DataLoader, - model: nn.Module, - token_table: k2.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: - Mapping ids to tokens. - - 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 = list(ref_text) - 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() - Aishell4AsrDataModule.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 = k2.SymbolTable.from_file(args.tokens) - assert token_table[0] == "" - - 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 - - aishell4 = Aishell4AsrDataModule(args) - - def remove_short_utt(c: Cut): - T = ((c.num_frames - 7) // 2 + 1) // 2 - if T <= 0: - logging.warning( - f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}." - ) - return T > 0 - - def text_normalize_for_cut(c: Cut): - # Text normalize for each sample - text = c.supervisions[0].text - text = text.strip("\n").strip("\t") - c.supervisions[0].text = text_normalize(text) - return c - - test_cuts = aishell4.test_cuts() - test_cuts = test_cuts.map(text_normalize_for_cut) - test_cuts = test_cuts.filter(remove_short_utt) - test_dl = aishell4.test_dataloaders(test_cuts) - - test_sets = ["test"] - test_dls = [test_dl] - - for test_set, test_dl in zip(test_sets, test_dls): - 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/aishell4/ASR/zipformer/onnx_pretrained-streaming.py b/egs/aishell4/ASR/zipformer/onnx_pretrained-streaming.py deleted file mode 120000 index cfea104c2..000000000 --- a/egs/aishell4/ASR/zipformer/onnx_pretrained-streaming.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/onnx_pretrained-streaming.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/onnx_pretrained.py b/egs/aishell4/ASR/zipformer/onnx_pretrained.py deleted file mode 120000 index 8f32f4ee7..000000000 --- a/egs/aishell4/ASR/zipformer/onnx_pretrained.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/onnx_pretrained.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/optim.py b/egs/aishell4/ASR/zipformer/optim.py deleted file mode 120000 index 5eaa3cffd..000000000 --- a/egs/aishell4/ASR/zipformer/optim.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/optim.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/pretrained.py b/egs/aishell4/ASR/zipformer/pretrained.py deleted file mode 120000 index 0bd71dde4..000000000 --- a/egs/aishell4/ASR/zipformer/pretrained.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/pretrained.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/scaling.py b/egs/aishell4/ASR/zipformer/scaling.py deleted file mode 120000 index 6f398f431..000000000 --- a/egs/aishell4/ASR/zipformer/scaling.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/scaling.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/scaling_converter.py b/egs/aishell4/ASR/zipformer/scaling_converter.py deleted file mode 120000 index b0ecee05e..000000000 --- a/egs/aishell4/ASR/zipformer/scaling_converter.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/scaling_converter.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/streaming_beam_search.py b/egs/aishell4/ASR/zipformer/streaming_beam_search.py deleted file mode 120000 index b1ed54557..000000000 --- a/egs/aishell4/ASR/zipformer/streaming_beam_search.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/streaming_beam_search.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/streaming_decode.py b/egs/aishell4/ASR/zipformer/streaming_decode.py deleted file mode 100755 index f0b2bb26d..000000000 --- a/egs/aishell4/ASR/zipformer/streaming_decode.py +++ /dev/null @@ -1,888 +0,0 @@ -#!/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 16 \ - --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 torch -from asr_datamodule import Aishell4AsrDataModule -from decode_stream import DecodeStream -from kaldifeat import Fbank, FbankOptions -from lhotse import CutSet -from lhotse.cut import Cut -from local.text_normalize import text_normalize -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_model, get_params - -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=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( - "--lang-dir", - type=str, - default="data/lang_char", - help="Path to the lang dir(containing lexicon, tokens, etc.)", - ) - - 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( - "--blank-penalty", - type=float, - default=0.0, - help=""" - The penalty applied on blank symbol during decoding. - Note: It is a positive value that would be applied to logits like - this `logits[:, 0] -= blank_penalty` (suppose logits.shape is - [batch_size, vocab] and blank id is 0). - """, - ) - - 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, - blank_penalty=params.blank_penalty, - ) - 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, - blank_penalty=params.blank_penalty, - ) - 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, - blank_penalty=params.blank_penalty, - ) - 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, - lexicon: Lexicon, - 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. - lexicon: - The Lexicon. - 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 - if audio.max() > 1: - logging.warning( - f"The audio should be normalized to [-1, 1], audio.max : {audio.max()}." - f"Clipping to [-1, 1]." - ) - audio = np.clip(audio, -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, - list(decode_streams[i].ground_truth.strip()), - [ - lexicon.token_table[idx] - for idx in decode_streams[i].decoding_result() - ], - ) - ) - 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(), - [ - lexicon.token_table[idx] - for idx in decode_streams[i].decoding_result() - ], - ) - ) - del decode_streams[i] - - key = f"blank_penalty_{params.blank_penalty}" - if params.decoding_method == "greedy_search": - key = f"greedy_search_{key}" - elif params.decoding_method == "fast_beam_search": - key = ( - f"beam_{params.beam}_" - f"max_contexts_{params.max_contexts}_" - f"max_states_{params.max_states}_{key}" - ) - elif params.decoding_method == "modified_beam_search": - key = f"num_active_paths_{params.num_active_paths}_{key}" - 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() - Aishell4AsrDataModule.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}" - params.suffix += f"-blank-penalty-{params.blank_penalty}" - - # 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}") - - lexicon = Lexicon(params.lang_dir) - params.blank_id = lexicon.token_table[""] - params.vocab_size = max(lexicon.tokens) + 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.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}") - - aishell4 = Aishell4AsrDataModule(args) - - def text_normalize_for_cut(c: Cut): - # Text normalize for each sample - text = c.supervisions[0].text - text = text.strip("\n").strip("\t") - c.supervisions[0].text = text_normalize(text) - return c - - test_cuts = aishell4.test_cuts() - test_cuts = test_cuts.map(text_normalize_for_cut) - test_sets = ["test"] - test_cuts = [test_cuts] - - for test_set, test_cut in zip(test_sets, test_cuts): - results_dict = decode_dataset( - cuts=test_cut, - params=params, - model=model, - lexicon=lexicon, - 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/aishell4/ASR/zipformer/subsampling.py b/egs/aishell4/ASR/zipformer/subsampling.py deleted file mode 120000 index 01ae9002c..000000000 --- a/egs/aishell4/ASR/zipformer/subsampling.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/subsampling.py \ No newline at end of file diff --git a/egs/aishell4/ASR/zipformer/train.py b/egs/aishell4/ASR/zipformer/train.py deleted file mode 100755 index 3d8b9fbc9..000000000 --- a/egs/aishell4/ASR/zipformer/train.py +++ /dev/null @@ -1,1362 +0,0 @@ -#!/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="0,1,2,3,4,5,6,7" - -./zipformer/train.py \ - --world-size 8 \ - --num-epochs 12 \ - --start-epoch 1 \ - --exp-dir zipformer/exp \ - --training-subset L - --lr-epochs 1.5 \ - --max-duration 350 - -# For mix precision training: - -./zipformer/train.py \ - --world-size 8 \ - --num-epochs 12 \ - --start-epoch 1 \ - --use-fp16 1 \ - --exp-dir zipformer/exp \ - --training-subset L \ - --lr-epochs 1.5 \ - --max-duration 750 - -""" - - -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 torch -import torch.multiprocessing as mp -import torch.nn as nn -from asr_datamodule import Aishell4AsrDataModule -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 local.text_normalize import text_normalize -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.char_graph_compiler import CharCtcTrainingGraphCompiler -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.lexicon import Lexicon -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.""", - ) - - -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( - "--lang-dir", - type=str, - default="data/lang_char", - help="""The lang dir - It contains language related input files such as - "lexicon.txt" - """, - ) - - 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( - "--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 0. - """, - ) - - 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, - # 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: - encoder_embed = get_encoder_embed(params) - encoder = get_encoder_model(params) - decoder = get_decoder_model(params) - joiner = get_joiner_model(params) - - model = AsrModel( - encoder_embed=encoder_embed, - encoder=encoder, - decoder=decoder, - joiner=joiner, - encoder_dim=int(max(params.encoder_dim.split(","))), - decoder_dim=params.decoder_dim, - vocab_size=params.vocab_size, - ) - 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"] - - if "cur_batch_idx" in saved_params: - params["cur_batch_idx"] = saved_params["cur_batch_idx"] - - return saved_params - - -def save_checkpoint( - params: AttributeDict, - model: 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], - graph_compiler: CharCtcTrainingGraphCompiler, - batch: dict, - is_training: bool, -) -> Tuple[Tensor, MetricsTracker]: - """ - Compute CTC 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 = graph_compiler.texts_to_ids(texts) - y = k2.RaggedTensor(y).to(device) - - with torch.set_grad_enabled(is_training): - simple_loss, pruned_loss, _ = model( - x=feature, - x_lens=feature_lens, - y=y, - prune_range=params.prune_range, - am_scale=params.am_scale, - lm_scale=params.lm_scale, - ) - - 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 - - assert loss.requires_grad == is_training - - info = MetricsTracker() - with warnings.catch_warnings(): - warnings.simplefilter("ignore") - info["frames"] = (feature_lens // params.subsampling_factor).sum().item() - - # Note: We use reduction=sum while computing the loss. - info["loss"] = loss.detach().cpu().item() - info["simple_loss"] = simple_loss.detach().cpu().item() - info["pruned_loss"] = pruned_loss.detach().cpu().item() - - return loss, info - - -def compute_validation_loss( - params: AttributeDict, - model: Union[nn.Module, DDP], - graph_compiler: CharCtcTrainingGraphCompiler, - 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, - graph_compiler=graph_compiler, - 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, - graph_compiler: CharCtcTrainingGraphCompiler, - 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() - - cur_batch_idx = params.get("cur_batch_idx", 0) - - 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)) - if batch_idx < cur_batch_idx: - continue - cur_batch_idx = batch_idx - - params.batch_idx_train += 1 - batch_size = len(batch["supervisions"]["text"]) - - try: - with torch.cuda.amp.autocast(enabled=params.use_fp16): - loss, loss_info = compute_loss( - params=params, - model=model, - graph_compiler=graph_compiler, - 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, graph_compiler=graph_compiler) - 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 - ): - params.cur_batch_idx = batch_idx - save_checkpoint_with_global_batch_idx( - out_dir=params.exp_dir, - global_batch_idx=params.batch_idx_train, - model=model, - model_avg=model_avg, - params=params, - optimizer=optimizer, - scheduler=scheduler, - sampler=train_dl.sampler, - scaler=scaler, - rank=rank, - ) - del params.cur_batch_idx - remove_checkpoints( - out_dir=params.exp_dir, - topk=params.keep_last_k, - rank=rank, - ) - - if batch_idx % 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, - graph_compiler=graph_compiler, - 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}") - - lexicon = Lexicon(params.lang_dir) - graph_compiler = CharCtcTrainingGraphCompiler( - lexicon=lexicon, - device=device, - ) - - params.blank_id = lexicon.token_table[""] - params.vocab_size = max(lexicon.tokens) + 1 - - 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( - 2**22 - ) # allow 4 megabytes per sub-module - diagnostic = diagnostics.attach_diagnostics(model, opts) - - if params.inf_check: - register_inf_check_hooks(model) - - aishell4 = Aishell4AsrDataModule(args) - # Combine all of the training data - train_cuts = aishell4.train_S_cuts() - train_cuts += aishell4.train_M_cuts() - train_cuts += aishell4.train_L_cuts() - - def remove_short_and_long_utt(c: Cut): - # Keep only utterances with duration between 1 second and 15 seconds - # - # Caution: There is a reason to select 15.0 here. Please see - # ../local/display_manifest_statistics.py - # - # You should use ../local/display_manifest_statistics.py to get - # an utterance duration distribution for your dataset to select - # the threshold - if c.duration < 1.0 or c.duration > 20.0: - # logging.warning( - # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" - # ) - return False - - # In pruned RNN-T, we require that T >= S - # where T is the number of feature frames after subsampling - # and S is the number of tokens in the utterance - - # In ./zipformer.py, the conv module uses the following expression - # for subsampling - T = ((c.num_frames - 7) // 2 + 1) // 2 - tokens = graph_compiler.texts_to_ids([c.supervisions[0].text])[0] - - if T < len(tokens): - logging.warning( - f"Exclude cut with ID {c.id} from training. " - f"Number of frames (before subsampling): {c.num_frames}. " - f"Number of frames (after subsampling): {T}. " - f"Text: {c.supervisions[0].text}. " - f"Tokens: {tokens}. " - f"Number of tokens: {len(tokens)}" - ) - return False - - return True - - def text_normalize_for_cut(c: Cut): - # Text normalize for each sample - text = c.supervisions[0].text - text = text.strip("\n").strip("\t") - c.supervisions[0].text = text_normalize(text) - return c - - train_cuts = train_cuts.filter(remove_short_and_long_utt) - train_cuts = train_cuts.map(text_normalize_for_cut) - - 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 = aishell4.train_dataloaders( - train_cuts, sampler_state_dict=sampler_state_dict - ) - - valid_cuts = aishell4.valid_cuts() - valid_cuts = valid_cuts.map(text_normalize_for_cut) - valid_dl = aishell4.valid_dataloaders(valid_cuts) - - if False and not params.print_diagnostics: - scan_pessimistic_batches_for_oom( - model=model, - train_dl=train_dl, - optimizer=optimizer, - graph_compiler=graph_compiler, - 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, - graph_compiler=graph_compiler, - 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, - graph_compiler: CharCtcTrainingGraphCompiler, -) -> 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`. - graph_compiler: - The compiler to encode texts to ids. - """ - 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}") - - texts = supervisions["text"] - y = graph_compiler.texts_to_ids(texts) - 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, - graph_compiler: CharCtcTrainingGraphCompiler, - 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, - graph_compiler=graph_compiler, - 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, graph_compiler=graph_compiler) - raise - logging.info( - f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" - ) - - -def main(): - parser = get_parser() - Aishell4AsrDataModule.add_arguments(parser) - args = parser.parse_args() - args.lang_dir = Path(args.lang_dir) - 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/aishell4/ASR/zipformer/zipformer.py b/egs/aishell4/ASR/zipformer/zipformer.py deleted file mode 120000 index 23011dda7..000000000 --- a/egs/aishell4/ASR/zipformer/zipformer.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/zipformer/zipformer.py \ No newline at end of file