From 141bc10a98a37c04a3a9d9d2916d2f2961a7c66e Mon Sep 17 00:00:00 2001 From: Dongji Gao Date: Sat, 16 Sep 2023 19:37:44 -0400 Subject: [PATCH] remove unnecessary test files --- .../WSASR/conformer_ctc2/decode.bkp | 1015 -------------- .../WSASR/conformer_ctc2/train_lexicon_new.py | 1177 ----------------- 2 files changed, 2192 deletions(-) delete mode 100755 egs/librispeech/WSASR/conformer_ctc2/decode.bkp delete mode 100755 egs/librispeech/WSASR/conformer_ctc2/train_lexicon_new.py diff --git a/egs/librispeech/WSASR/conformer_ctc2/decode.bkp b/egs/librispeech/WSASR/conformer_ctc2/decode.bkp deleted file mode 100755 index 7c3baf18c..000000000 --- a/egs/librispeech/WSASR/conformer_ctc2/decode.bkp +++ /dev/null @@ -1,1015 +0,0 @@ -#!/usr/bin/env python3 -# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, -# Fangjun Kuang, -# Quandong Wang) -# -# See ../../../../LICENSE for clarification regarding multiple authors -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - - -import argparse -import logging -from collections import defaultdict -from pathlib import Path -from typing import Dict, List, Optional, Tuple - -import k2 -import sentencepiece as spm -import torch -import torch.nn as nn -from asr_datamodule import LibriSpeechAsrDataModule -from conformer import Conformer - -from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler -from icefall.checkpoint import ( - average_checkpoints, - average_checkpoints_with_averaged_model, - find_checkpoints, - load_checkpoint, -) -from icefall.decode import ( - get_lattice, - nbest_decoding, - nbest_oracle, - one_best_decoding, - rescore_with_attention_decoder, - rescore_with_n_best_list, - rescore_with_rnn_lm, - rescore_with_whole_lattice, -) -from icefall.env import get_env_info -from icefall.lexicon import Lexicon -from icefall.rnn_lm.model import RnnLmModel -from icefall.utils import ( - AttributeDict, - get_texts, - load_averaged_model, - setup_logger, - store_transcripts, - str2bool, - write_error_stats, -) - - -def get_parser(): - parser = argparse.ArgumentParser( - formatter_class=argparse.ArgumentDefaultsHelpFormatter - ) - - parser.add_argument( - "--epoch", - type=int, - default=20, - 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=1, - help="Number of checkpoints to average. Automatically select " - "consecutive checkpoints before the checkpoint specified by " - "'--epoch' and '--iter'", - ) - - parser.add_argument( - "--method", - type=str, - default="attention-decoder", - help="""Decoding method. - Supported values are: - - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece - model, i.e., lang_dir/bpe.model, to convert word pieces to words. - It needs neither a lexicon nor an n-gram LM. - - (1) ctc-greedy-search. It only use CTC output and a sentence piece - model for decoding. It produces the same results with ctc-decoding. - - (2) 1best. Extract the best path from the decoding lattice as the - decoding result. - - (3) nbest. Extract n paths from the decoding lattice; the path - with the highest score is the decoding result. - - (4) nbest-rescoring. Extract n paths from the decoding lattice, - rescore them with an n-gram LM (e.g., a 4-gram LM), the path with - the highest score is the decoding result. - - (5) whole-lattice-rescoring. Rescore the decoding lattice with an - n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice - is the decoding result. - - (6) attention-decoder. Extract n paths from the LM rescored - lattice, the path with the highest score is the decoding result. - - (7) rnn-lm. Rescoring with attention-decoder and RNN LM. We assume - you have trained an RNN LM using ./rnn_lm/train.py - - (8) nbest-oracle. Its WER is the lower bound of any n-best - rescoring method can achieve. Useful for debugging n-best - rescoring method. - """, - ) - - parser.add_argument( - "--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( - "--use-best-model", - type=str2bool, - default=False, - ) - - parser.add_argument( - "--num-decoder-layers", - type=int, - default=0, - help="""Number of decoder layer of transformer decoder. - Setting this to 0 will not create the decoder at all (pure CTC model) - """, - ) - - parser.add_argument( - "--num-paths", - type=int, - default=100, - help="""Number of paths for n-best based decoding method. - Used only when "method" is one of the following values: - nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle - """, - ) - - parser.add_argument( - "--nbest-scale", - type=float, - default=0.5, - help="""The scale to be applied to `lattice.scores`. - It's needed if you use any kinds of n-best based rescoring. - Used only when "method" is one of the following values: - nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle - A smaller value results in more unique paths. - """, - ) - - parser.add_argument( - "--exp-dir", - type=str, - default="conformer_ctc2/exp", - help="The experiment dir", - ) - - parser.add_argument( - "--lang-dir", - type=str, - default="data/lang_bpe_500", - help="The lang dir", - ) - - parser.add_argument( - "--lm-dir", - type=str, - default="data/lm", - help="""The n-gram LM dir. - It should contain either G_4_gram.pt or G_4_gram.fst.txt - """, - ) - - parser.add_argument( - "--rnn-lm-exp-dir", - type=str, - default="rnn_lm/exp", - help="""Used only when --method is rnn-lm. - It specifies the path to RNN LM exp dir. - """, - ) - - parser.add_argument( - "--rnn-lm-epoch", - type=int, - default=7, - help="""Used only when --method is rnn-lm. - It specifies the checkpoint to use. - """, - ) - - parser.add_argument( - "--rnn-lm-avg", - type=int, - default=2, - help="""Used only when --method is rnn-lm. - It specifies the number of checkpoints to average. - """, - ) - - parser.add_argument( - "--rnn-lm-embedding-dim", - type=int, - default=2048, - help="Embedding dim of the model", - ) - - parser.add_argument( - "--rnn-lm-hidden-dim", - type=int, - default=2048, - help="Hidden dim of the model", - ) - - parser.add_argument( - "--rnn-lm-num-layers", - type=int, - default=4, - help="Number of RNN layers the model", - ) - parser.add_argument( - "--rnn-lm-tie-weights", - type=str2bool, - default=False, - help="""True to share the weights between the input embedding layer and the - last output linear layer - """, - ) - - parser.add_argument( - "--blank-bias", - type=float, - default=0, - help="""blank bias""", - ) - - return parser - - -def get_params() -> AttributeDict: - params = AttributeDict( - { - # parameters for conformer - "subsampling_factor": 2, - "feature_dim": 768, - "nhead": 8, - "dim_feedforward": 2048, - "encoder_dim": 512, - "num_encoder_layers": 12, - # parameters for decoding - "search_beam": 20, - "output_beam": 50, - "min_active_states": 300, - "max_active_states": 10000, - "use_double_scores": True, - "env_info": get_env_info(), - } - ) - return params - - -def ctc_greedy_search( - nnet_output: torch.Tensor, - memory: torch.Tensor, - memory_key_padding_mask: torch.Tensor, -) -> List[List[int]]: - """Apply CTC greedy search - - Args: - speech (torch.Tensor): (batch, max_len, feat_dim) - speech_length (torch.Tensor): (batch, ) - Returns: - List[List[int]]: best path result - """ - batch_size = memory.shape[1] - # Let's assume B = batch_size - encoder_out = memory - encoder_mask = memory_key_padding_mask - maxlen = encoder_out.size(0) - - ctc_probs = nnet_output # (B, maxlen, vocab_size) - topk_prob, topk_index = ctc_probs.topk(1, dim=2) # (B, maxlen, 1) - topk_index = topk_index.view(batch_size, maxlen) # (B, maxlen) - topk_index = topk_index.masked_fill_(encoder_mask, 0) # (B, maxlen) - hyps = [hyp.tolist() for hyp in topk_index] - scores = topk_prob.max(1) - hyps = [remove_duplicates_and_blank(hyp) for hyp in hyps] - return hyps, scores - - -def remove_duplicates_and_blank(hyp: List[int]) -> List[int]: - # from https://github.com/wenet-e2e/wenet/blob/main/wenet/utils/common.py - new_hyp: List[int] = [] - cur = 0 - while cur < len(hyp): - if hyp[cur] != 0: - new_hyp.append(hyp[cur]) - prev = cur - while cur < len(hyp) and hyp[cur] == hyp[prev]: - cur += 1 - return new_hyp - - -def decode_one_batch( - params: AttributeDict, - model: nn.Module, - rnn_lm_model: Optional[nn.Module], - HLG: Optional[k2.Fsa], - H: Optional[k2.Fsa], - bpe_model: Optional[spm.SentencePieceProcessor], - batch: dict, - word_table: k2.SymbolTable, - sos_id: int, - eos_id: int, - G: Optional[k2.Fsa] = None, -) -> Dict[str, List[List[str]]]: - """Decode one batch and return the result in a dict. The dict has the - following format: - - - key: It indicates the setting used for decoding. For example, - if no rescoring is used, the key is the string `no_rescore`. - If LM rescoring is used, the key is the string `lm_scale_xxx`, - where `xxx` is the value of `lm_scale`. An example key is - `lm_scale_0.7` - - value: It contains the decoding result. `len(value)` equals to - batch size. `value[i]` is the decoding result for the i-th - utterance in the given batch. - Args: - params: - It's the return value of :func:`get_params`. - - - params.method is "1best", it uses 1best decoding without LM rescoring. - - params.method is "nbest", it uses nbest decoding without LM rescoring. - - params.method is "nbest-rescoring", it uses nbest LM rescoring. - - params.method is "whole-lattice-rescoring", it uses whole lattice LM - rescoring. - - model: - The neural model. - rnn_lm_model: - The neural model for RNN LM. - HLG: - The decoding graph. Used only when params.method is NOT ctc-decoding. - H: - The ctc topo. Used only when params.method is ctc-decoding. - bpe_model: - The BPE model. Used only when params.method is ctc-decoding. - batch: - It is the return value from iterating - `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation - for the format of the `batch`. - word_table: - The word symbol table. - sos_id: - The token ID of the SOS. - eos_id: - The token ID of the EOS. - G: - An LM. It is not None when params.method is "nbest-rescoring" - or "whole-lattice-rescoring". In general, the G in HLG - is a 3-gram LM, while this G is a 4-gram LM. - Returns: - Return the decoding result. See above description for the format of - the returned dict. Note: If it decodes to nothing, then return None. - """ - if HLG is not None: - device = HLG.device - else: - device = H.device - feature = batch["inputs"] - assert feature.ndim == 3 - feature = feature.to(device) - # at entry, feature is (N, T, C) - - supervisions = batch["supervisions"] - - nnet_output, memory, memory_key_padding_mask = model(feature, supervisions) - # nnet_output is (N, T, C) - nnet_output[:, :, 0] += params.blank_bias - - #print(f"nnet_output shape: {nnet_output.shape}") - - supervision_segments = torch.stack( - ( - supervisions["sequence_idx"], - torch.div( - supervisions["start_frame"], - params.subsampling_factor, - rounding_mode="trunc", - ), - torch.div( - supervisions["num_frames"], - params.subsampling_factor, - rounding_mode="trunc", - ), - ), - 1, - ).to(torch.int32) - - if H is None: - assert HLG is not None - decoding_graph = HLG - else: - assert HLG is None - assert bpe_model is not None - decoding_graph = H - - lattice = get_lattice( - nnet_output=nnet_output, - decoding_graph=decoding_graph, - supervision_segments=supervision_segments, - search_beam=params.search_beam, - output_beam=params.output_beam, - min_active_states=params.min_active_states, - max_active_states=params.max_active_states, - subsampling_factor=4, - ) - - if params.method == "ctc-decoding": - best_path = one_best_decoding( - lattice=lattice, use_double_scores=params.use_double_scores - ) - # Note: `best_path.aux_labels` contains token IDs, not word IDs - # since we are using H, not HLG here. - # - # token_ids is a lit-of-list of IDs - token_ids = get_texts(best_path) - - # hyps is a list of str, e.g., ['xxx yyy zzz', ...] - hyps = bpe_model.decode(token_ids) - - # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ] - hyps = [s.split() for s in hyps] - key = "ctc-decoding" - return {key: hyps} - - if params.method == "ctc-greedy-search": - hyps, _ = ctc_greedy_search( - nnet_output, - memory, - memory_key_padding_mask, - ) - - # hyps is a list of str, e.g., ['xxx yyy zzz', ...] - hyps = bpe_model.decode(hyps) - - # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ] - hyps = [s.split() for s in hyps] - key = "ctc-greedy-search" - return {key: hyps} - - if params.method == "nbest-oracle": - # Note: You can also pass rescored lattices to it. - # We choose the HLG decoded lattice for speed reasons - # as HLG decoding is faster and the oracle WER - # is only slightly worse than that of rescored lattices. - best_path = nbest_oracle( - lattice=lattice, - num_paths=params.num_paths, - ref_texts=supervisions["text"], - word_table=word_table, - nbest_scale=params.nbest_scale, - oov="", - ) - hyps = get_texts(best_path) - hyps = [[word_table[i] for i in ids] for ids in hyps] - key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}" # noqa - return {key: hyps} - - if params.method in ["1best", "nbest"]: - if params.method == "1best": - best_path = one_best_decoding( - lattice=lattice, use_double_scores=params.use_double_scores - ) - key = "no_rescore" - else: - best_path = nbest_decoding( - lattice=lattice, - num_paths=params.num_paths, - use_double_scores=params.use_double_scores, - nbest_scale=params.nbest_scale, - ) - key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}" # noqa - - hyps = get_texts(best_path) - hyps = [[word_table[i] for i in ids] for ids in hyps] - return {key: hyps} - - assert params.method in [ - "nbest-rescoring", - "whole-lattice-rescoring", - "attention-decoder", - "rnn-lm", - ] - - lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7] - lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3] - lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0] - - if params.method == "nbest-rescoring": - best_path_dict = rescore_with_n_best_list( - lattice=lattice, - G=G, - num_paths=params.num_paths, - lm_scale_list=lm_scale_list, - nbest_scale=params.nbest_scale, - ) - elif params.method == "whole-lattice-rescoring": - best_path_dict = rescore_with_whole_lattice( - lattice=lattice, - G_with_epsilon_loops=G, - lm_scale_list=lm_scale_list, - ) - elif params.method == "attention-decoder": - # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM. - rescored_lattice = rescore_with_whole_lattice( - lattice=lattice, - G_with_epsilon_loops=G, - lm_scale_list=None, - ) - # TODO: pass `lattice` instead of `rescored_lattice` to - # `rescore_with_attention_decoder` - - best_path_dict = rescore_with_attention_decoder( - lattice=rescored_lattice, - num_paths=params.num_paths, - model=model, - memory=memory, - memory_key_padding_mask=memory_key_padding_mask, - sos_id=sos_id, - eos_id=eos_id, - nbest_scale=params.nbest_scale, - ) - elif params.method == "rnn-lm": - # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM. - rescored_lattice = rescore_with_whole_lattice( - lattice=lattice, - G_with_epsilon_loops=G, - lm_scale_list=None, - ) - - best_path_dict = rescore_with_rnn_lm( - lattice=rescored_lattice, - num_paths=params.num_paths, - rnn_lm_model=rnn_lm_model, - model=model, - memory=memory, - memory_key_padding_mask=memory_key_padding_mask, - sos_id=sos_id, - eos_id=eos_id, - blank_id=0, - nbest_scale=params.nbest_scale, - ) - else: - assert False, f"Unsupported decoding method: {params.method}" - - ans = dict() - if best_path_dict is not None: - for lm_scale_str, best_path in best_path_dict.items(): - hyps = get_texts(best_path) - hyps = [[word_table[i] for i in ids] for ids in hyps] - ans[lm_scale_str] = hyps - else: - ans = None - return ans - - -def decode_dataset( - dl: torch.utils.data.DataLoader, - params: AttributeDict, - model: nn.Module, - rnn_lm_model: Optional[nn.Module], - HLG: Optional[k2.Fsa], - H: Optional[k2.Fsa], - bpe_model: Optional[spm.SentencePieceProcessor], - word_table: k2.SymbolTable, - sos_id: int, - eos_id: int, - G: Optional[k2.Fsa] = None, -) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: - """Decode dataset. - - Args: - dl: - PyTorch's dataloader containing the dataset to decode. - params: - It is returned by :func:`get_params`. - model: - The neural model. - rnn_lm_model: - The neural model for RNN LM. - HLG: - The decoding graph. Used only when params.method is NOT ctc-decoding. - H: - The ctc topo. Used only when params.method is ctc-decoding. - bpe_model: - The BPE model. Used only when params.method is ctc-decoding. - word_table: - It is the word symbol table. - sos_id: - The token ID for SOS. - eos_id: - The token ID for EOS. - G: - An LM. It is not None when params.method is "nbest-rescoring" - or "whole-lattice-rescoring". In general, the G in HLG - is a 3-gram LM, while this G is a 4-gram LM. - Returns: - Return a dict, whose key may be "no-rescore" if no LM rescoring - is used, or it may be "lm_scale_0.7" if LM rescoring is used. - Its value is a list of tuples. Each tuple contains two elements: - The first is the reference transcript, and the second is the - predicted result. - """ - num_cuts = 0 - - try: - num_batches = len(dl) - except TypeError: - num_batches = "?" - - results = defaultdict(list) - for batch_idx, batch in enumerate(dl): - texts = batch["supervisions"]["text"] - cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] - - hyps_dict = decode_one_batch( - params=params, - model=model, - rnn_lm_model=rnn_lm_model, - HLG=HLG, - H=H, - bpe_model=bpe_model, - batch=batch, - word_table=word_table, - G=G, - sos_id=sos_id, - eos_id=eos_id, - ) - - if hyps_dict is not None: - for lm_scale, hyps in hyps_dict.items(): - this_batch = [] - assert len(hyps) == len(texts) - for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): - ref_words = ref_text.split() - this_batch.append((cut_id, ref_words, hyp_words)) - - results[lm_scale].extend(this_batch) - else: - assert len(results) > 0, "It should not decode to empty in the first batch!" - this_batch = [] - hyp_words = [] - for ref_text in texts: - ref_words = ref_text.split() - this_batch.append((ref_words, hyp_words)) - - for lm_scale in results.keys(): - results[lm_scale].extend(this_batch) - - num_cuts += len(texts) - - if batch_idx % 100 == 0: - batch_str = f"{batch_idx}/{num_batches}" - - logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") - return results - - -def save_results( - params: AttributeDict, - test_set_name: str, - results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], -): - if params.method in ("attention-decoder", "rnn-lm"): - # Set it to False since there are too many logs. - enable_log = False - else: - enable_log = True - test_set_wers = dict() - for key, results in results_dict.items(): - recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt" - results = sorted(results) - store_transcripts(filename=recog_path, texts=results) - if enable_log: - 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.exp_dir / f"errs-{test_set_name}-{key}.txt" - with open(errs_filename, "w") as f: - wer = write_error_stats( - f, f"{test_set_name}-{key}", results, enable_log=enable_log - ) - test_set_wers[key] = wer - - if enable_log: - 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.exp_dir / f"wer-summary-{test_set_name}.txt" - with open(errs_info, "w") as f: - print("settings\tWER", file=f) - for key, val in test_set_wers: - print("{}\t{}".format(key, val), file=f) - - s = "\nFor {}, WER of different settings are:\n".format(test_set_name) - note = "\tbest for {}".format(test_set_name) - for key, val in test_set_wers: - s += "{}\t{}{}\n".format(key, val, note) - note = "" - logging.info(s) - - -@torch.no_grad() -def main(): - parser = get_parser() - LibriSpeechAsrDataModule.add_arguments(parser) - args = parser.parse_args() - args.exp_dir = Path(args.exp_dir) - args.lang_dir = Path(args.lang_dir) - args.lm_dir = Path(args.lm_dir) - - params = get_params() - params.update(vars(args)) - - setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode") - logging.info("Decoding started") - logging.info(params) - - lexicon = Lexicon(params.lang_dir) - max_token_id = max(lexicon.tokens) - 1 - num_classes = max_token_id + 1 # +1 for the blank - - device = torch.device("cpu") - if torch.cuda.is_available(): - device = torch.device("cuda", 0) - - logging.info(f"device: {device}") - - graph_compiler = BpeCtcTrainingGraphCompiler( - params.lang_dir, - device=device, - sos_token="", - eos_token="", - ) - sos_id = graph_compiler.sos_id - eos_id = graph_compiler.eos_id - - params.num_classes = num_classes - params.sos_id = sos_id - params.eos_id = eos_id - - if params.method == "ctc-decoding" or params.method == "ctc-greedy-search": - HLG = None - H = k2.ctc_topo( - max_token=max_token_id, - modified=False, - device=device, - ) - bpe_model = spm.SentencePieceProcessor() - bpe_model.load(str(params.lang_dir / "bpe.model")) - else: - H = None - bpe_model = None - HLG = k2.Fsa.from_dict( - torch.load(f"{params.lang_dir}/HLG.pt", map_location=device) - ) - assert HLG.requires_grad is False - - if not hasattr(HLG, "lm_scores"): - HLG.lm_scores = HLG.scores.clone() - - if params.method in ( - "nbest-rescoring", - "whole-lattice-rescoring", - "attention-decoder", - "rnn-lm", - ): - if not (params.lm_dir / "G_4_gram.pt").is_file(): - logging.info("Loading G_4_gram.fst.txt") - logging.warning("It may take 8 minutes.") - with open(params.lm_dir / "G_4_gram.fst.txt") as f: - first_word_disambig_id = lexicon.word_table["#0"] - - G = k2.Fsa.from_openfst(f.read(), acceptor=False) - # G.aux_labels is not needed in later computations, so - # remove it here. - del G.aux_labels - # CAUTION: The following line is crucial. - # Arcs entering the back-off state have label equal to #0. - # We have to change it to 0 here. - G.labels[G.labels >= first_word_disambig_id] = 0 - # See https://github.com/k2-fsa/k2/issues/874 - # for why we need to set G.properties to None - G.__dict__["_properties"] = None - G = k2.Fsa.from_fsas([G]).to(device) - G = k2.arc_sort(G) - # Save a dummy value so that it can be loaded in C++. - # See https://github.com/pytorch/pytorch/issues/67902 - # for why we need to do this. - G.dummy = 1 - - torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt") - else: - logging.info("Loading pre-compiled G_4_gram.pt") - d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device) - G = k2.Fsa.from_dict(d) - - if params.method in [ - "whole-lattice-rescoring", - "attention-decoder", - "rnn-lm", - ]: - # Add epsilon self-loops to G as we will compose - # it with the whole lattice later - G = k2.add_epsilon_self_loops(G) - G = k2.arc_sort(G) - G = G.to(device) - - # G.lm_scores is used to replace HLG.lm_scores during - # LM rescoring. - G.lm_scores = G.scores.clone() - else: - G = None - - model = Conformer( - num_features=params.feature_dim, - nhead=params.nhead, - d_model=params.encoder_dim, - num_classes=num_classes, - subsampling_factor=params.subsampling_factor, - num_encoder_layers=params.num_encoder_layers, - num_decoder_layers=params.num_decoder_layers, - ) - - if params.use_best_model: - logging.info("Loading best-valid-loss.py") - load_checkpoint(f"{params.exp_dir}/best-valid-loss.pt", model) - elif not params.use_averaged_model: - if params.iter > 0: - filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ - : params.avg - ] - if len(filenames) == 0: - raise ValueError( - f"No checkpoints found for" - f" --iter {params.iter}, --avg {params.avg}" - ) - elif len(filenames) < params.avg: - raise ValueError( - f"Not enough checkpoints ({len(filenames)}) found for" - f" --iter {params.iter}, --avg {params.avg}" - ) - logging.info(f"averaging {filenames}") - model.to(device) - model.load_state_dict(average_checkpoints(filenames, device=device)) - elif params.avg == 1: - load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) - else: - start = params.epoch - params.avg + 1 - filenames = [] - for i in range(start, params.epoch + 1): - if i >= 1: - filenames.append(f"{params.exp_dir}/epoch-{i}.pt") - logging.info(f"averaging {filenames}") - model.to(device) - model.load_state_dict(average_checkpoints(filenames, device=device)) - else: - if params.iter > 0: - filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ - : params.avg + 1 - ] - if len(filenames) == 0: - raise ValueError( - f"No checkpoints found for" - f" --iter {params.iter}, --avg {params.avg}" - ) - elif len(filenames) < params.avg + 1: - raise ValueError( - f"Not enough checkpoints ({len(filenames)}) found for" - f" --iter {params.iter}, --avg {params.avg}" - ) - filename_start = filenames[-1] - filename_end = filenames[0] - logging.info( - "Calculating the averaged model over iteration checkpoints" - f" from {filename_start} (excluded) to {filename_end}" - ) - model.to(device) - model.load_state_dict( - average_checkpoints_with_averaged_model( - filename_start=filename_start, - filename_end=filename_end, - device=device, - ) - ) - else: - assert params.avg > 0, params.avg - start = params.epoch - params.avg - assert start >= 1, start - filename_start = f"{params.exp_dir}/epoch-{start}.pt" - filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" - logging.info( - f"Calculating the averaged model over epoch range from " - f"{start} (excluded) to {params.epoch}" - ) - model.to(device) - model.load_state_dict( - average_checkpoints_with_averaged_model( - filename_start=filename_start, - filename_end=filename_end, - device=device, - ) - ) - - model.to(device) - model.eval() - num_param = sum([p.numel() for p in model.parameters()]) - logging.info(f"Number of model parameters: {num_param}") - - rnn_lm_model = None - if params.method == "rnn-lm": - rnn_lm_model = RnnLmModel( - vocab_size=params.num_classes, - embedding_dim=params.rnn_lm_embedding_dim, - hidden_dim=params.rnn_lm_hidden_dim, - num_layers=params.rnn_lm_num_layers, - tie_weights=params.rnn_lm_tie_weights, - ) - if params.rnn_lm_avg == 1: - load_checkpoint( - f"{params.rnn_lm_exp_dir}/epoch-{params.rnn_lm_epoch}.pt", - rnn_lm_model, - ) - rnn_lm_model.to(device) - else: - rnn_lm_model = load_averaged_model( - params.rnn_lm_exp_dir, - rnn_lm_model, - params.rnn_lm_epoch, - params.rnn_lm_avg, - device, - ) - rnn_lm_model.eval() - - # we need cut ids to display recognition results. - args.return_cuts = True - librispeech = LibriSpeechAsrDataModule(args) - - dev_clean_cuts = librispeech.dev_clean_cuts() - dev_other_cuts = librispeech.dev_other_cuts() - test_clean_cuts = librispeech.test_clean_cuts() - test_other_cuts = librispeech.test_other_cuts() - - dev_clean_dl = librispeech.test_dataloaders(dev_clean_cuts) - dev_other_dl = librispeech.test_dataloaders(dev_other_cuts) - test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) - test_other_dl = librispeech.test_dataloaders(test_other_cuts) - - test_sets = ["test-clean", "test-other"] - #test_sets = ["test-clean"] - test_dl = [test_clean_dl , test_other_dl] - #test_dl = [test_clean_dl] - - for test_set, test_dl in zip(test_sets, test_dl): - results_dict = decode_dataset( - dl=test_dl, - params=params, - model=model, - rnn_lm_model=rnn_lm_model, - HLG=HLG, - H=H, - bpe_model=bpe_model, - word_table=lexicon.word_table, - G=G, - sos_id=sos_id, - eos_id=eos_id, - ) - - save_results(params=params, test_set_name=test_set, results_dict=results_dict) - - logging.info("Done!") - - -torch.set_num_threads(1) -torch.set_num_interop_threads(1) - -if __name__ == "__main__": - main() diff --git a/egs/librispeech/WSASR/conformer_ctc2/train_lexicon_new.py b/egs/librispeech/WSASR/conformer_ctc2/train_lexicon_new.py deleted file mode 100755 index bc472d142..000000000 --- a/egs/librispeech/WSASR/conformer_ctc2/train_lexicon_new.py +++ /dev/null @@ -1,1177 +0,0 @@ -#!/usr/bin/env python3 -# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, -# Wei Kang, -# Mingshuang Luo, -# Zengwei Yao, -# Quandong Wang) -# -# 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" - -./conformer_ctc2/train.py \ - --world-size 4 \ - --num-epochs 30 \ - --start-epoch 1 \ - --exp-dir conformer_ctc2/exp \ - --full-libri 1 \ - --max-duration 300 - -# For mix precision training: - -./conformer_ctc2/train.py \ - --world-size 4 \ - --num-epochs 30 \ - --start-epoch 1 \ - --use-fp16 1 \ - --exp-dir conformer_ctc2/exp \ - --full-libri 1 \ - --max-duration 550 - -""" -import sys - -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 LibriSpeechAsrDataModule -from conformer import Conformer -from lhotse.cut import Cut -from lhotse.dataset.sampling.base import CutSampler -from lhotse.utils import fix_random_seed -from optim import Eden, Eve -from torch import Tensor -from torch.cuda.amp import GradScaler -from torch.nn.parallel import DistributedDataParallel as DDP -from torch.utils.tensorboard import SummaryWriter - -from icefall import diagnostics -from icefall.checkpoint import load_checkpoint, remove_checkpoints -from icefall.checkpoint import save_checkpoint as save_checkpoint_impl -from icefall.checkpoint import ( - save_checkpoint_with_global_batch_idx, - update_averaged_model, -) -from icefall.dist import cleanup_dist, setup_dist -from icefall.env import get_env_info -#from icefall.otc_lexicon_graph_compiler import OtcTrainingGraphCompiler -from icefall.otc_graph_compiler import OtcTrainingGraphCompiler -from icefall.lexicon import Lexicon -from icefall.utils import ( - AttributeDict, - MetricsTracker, - get_texts, - encode_supervisions_otc, - setup_logger, - str2bool, -) -from icefall.decode import one_best_decoding - -LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] - - -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="conformer_ctc2/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_bpe_500", - help="""The lang dir - It contains language related input files such as - "lexicon.txt" - """, - ) - - parser.add_argument( - "--initial-lr", - type=float, - default=0.003, - help="""The initial learning rate. This value should not need to be - changed.""", - ) - - parser.add_argument( - "--lr-batches", - type=float, - default=5000, - help="""Number of steps that affects how rapidly the learning rate decreases. - We suggest not to change this.""", - ) - - parser.add_argument( - "--lr-epochs", - type=float, - default=6, - help="""Number of epochs that affects how rapidly the learning rate decreases. - """, - ) - - parser.add_argument( - "--att-rate", - type=float, - default=0.0, - help="""The attention rate. - The total loss is (1 - att_rate) * ctc_loss + att_rate * att_loss - """, - ) - - parser.add_argument( - "--num-decoder-layers", - type=int, - default=0, - help="""Number of decoder layer of transformer decoder. - Setting this to 0 will not create the decoder at all (pure CTC model) - """, - ) - - parser.add_argument( - "--seed", - type=int, - default=42, - help="The seed for random generators intended for reproducibility", - ) - - parser.add_argument( - "--print-diagnostics", - type=str2bool, - default=False, - help="Accumulate stats on activations, print them and exit.", - ) - - parser.add_argument( - "--save-every-n", - type=int, - default=8000, - help="""Save checkpoint after processing this number of batches" - periodically. We save checkpoint to exp-dir/ whenever - params.batch_idx_train % save_every_n == 0. The checkpoint filename - has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' - Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the - end of each epoch where `xxx` is the epoch number counting from 0. - """, - ) - - parser.add_argument( - "--keep-last-k", - type=int, - default=5, - 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=100, - 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.", - ) - - parser.add_argument( - "--wasr-type", type=str, default="star", - ) - - parser.add_argument( - "--wasr-token", type=str, default="", - ) - - parser.add_argument( - "--allow-bypass", type=str2bool, default=True, - ) - - parser.add_argument( - "--allow-self-loop", type=str2bool, default=True, - ) - - parser.add_argument( - "--initial-bypass-penalty", type=float, default=0, - ) - - parser.add_argument( - "--initial-self-loop-penalty", type=float, default=0, - ) - - parser.add_argument( - "--min-penalty", type=float, default=0, - ) - - parser.add_argument( - "--bypass-penalty-decay", type=float, default=1.0, - ) - - parser.add_argument( - "--self-loop-penalty-decay", type=float, default=1.0, - ) - - parser.add_argument( - "--show-alignment", type=str2bool, default=False, - ) - - 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. - - - beam_size: It is used in k2.ctc_loss - - - reduction: It is used in k2.ctc_loss - - - use_double_scores: It is used in k2.ctc_loss - - - warm_step: The warm_step for Noam optimizer. - """ - params = AttributeDict( - { - "best_train_loss": float("inf"), - "best_valid_loss": float("inf"), - "best_train_epoch": -1, - "best_valid_epoch": -1, - "batch_idx_train": 0, - "log_interval": 1, - "reset_interval": 200, - "valid_interval": 800, # For the 100h subset, use 800 - # parameters for conformer - "feature_dim": 768, - "subsampling_factor": 2, - "encoder_dim": 512, - "nhead": 8, - "dim_feedforward": 2048, - "num_encoder_layers": 12, - # parameters for ctc loss - "beam_size": 10, - "reduction": "sum", - "use_double_scores": True, - # parameters for Noam - "model_warm_step": 3000, # arg given to model, not for lrate - "env_info": get_env_info(), - } - ) - - return params - - -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], - batch: dict, - graph_compiler: OtcTrainingGraphCompiler, - is_training: bool, - warmup: float = 2.0, - bypass_penalty: float = 0.0, - self_loop_penalty: float = 0.0, -) -> 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 Conformer in our case. - batch: - A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` - for the content in it. - graph_compiler: - It is used to build a decoding graph from a ctc topo and training - transcript. The training transcript is contained in the given `batch`, - while the ctc topo is built when this compiler is instantiated. - 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) - - with torch.set_grad_enabled(is_training): - nnet_output, encoder_memory, memory_mask = model( - feature, supervisions, warmup=warmup - ) - # Note: we assume is the last symbol in token list (tokens.txt) - _, _, C = nnet_output.shape - star_log_prob = torch.logsumexp( - nnet_output[:, :, 1:], dim=-1, keepdim=True - ) - torch.log(torch.tensor([C - 1])).to(device) - nnet_output = torch.cat([nnet_output, star_log_prob], dim=-1) - - # NOTE: We need `encode_supervisions` to sort sequences with - # different duration in decreasing order, required by - # `k2.intersect_dense` called in `k2.ctc_loss` - supervision_segments, texts, ids, orig_texts = encode_supervisions_otc( - supervisions, subsampling_factor=params.subsampling_factor - ) - decoding_graph = graph_compiler.compile( - texts=texts, - allow_bypass_arc=params.allow_bypass, - allow_self_loop_arc=params.allow_self_loop, - bypass_weight=bypass_penalty, - self_loop_weight=self_loop_penalty, - ) - - dense_fsa_vec = k2.DenseFsaVec(nnet_output, supervision_segments, allow_truncate=3,) - - ctc_loss = k2.ctc_loss( - decoding_graph=decoding_graph, - dense_fsa_vec=dense_fsa_vec, - output_beam=params.beam_size, - reduction=params.reduction, - use_double_scores=params.use_double_scores, - ) - - if params.att_rate != 0.0: - raise ValueError("not supported") - # with torch.set_grad_enabled(is_training): - # mmodel = model.module if hasattr(model, "module") else model - # # Note: We need to generate an unsorted version of token_ids - # # `encode_supervisions()` called above sorts text, but - # # encoder_memory and memory_mask are not sorted, so we - # # use an unsorted version `supervisions["text"]` to regenerate - # # the token_ids - # # - # # See https://github.com/k2-fsa/icefall/issues/97 - # # for more details - # unsorted_token_ids = graph_compiler.texts_to_ids(supervisions["text"]) - # att_loss = mmodel.decoder_forward( - # encoder_memory, - # memory_mask, - # token_ids=unsorted_token_ids, - # sos_id=graph_compiler.sos_id, - # eos_id=graph_compiler.eos_id, - # ) - # loss = (1.0 - params.att_rate) * ctc_loss + params.att_rate * att_loss - else: - loss = ctc_loss - att_loss = torch.tensor([0]) - - assert loss.requires_grad == is_training - - info = MetricsTracker() - with warnings.catch_warnings(): - warnings.simplefilter("ignore") - info["frames"] = (feature_lens // params.subsampling_factor).sum().item() - info["ctc_loss"] = ctc_loss.detach().cpu().item() - if params.att_rate != 0.0: - info["att_loss"] = att_loss.detach().cpu().item() - - # Note: We use reduction=sum while computing the loss. - info["loss"] = loss.detach().cpu().item() - - # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances` # noqa - info["utterances"] = feature.size(0) - # averaged input duration in frames over utterances - info["utt_duration"] = feature_lens.sum().item() - # averaged padding proportion over utterances - info["utt_pad_proportion"] = ( - ((feature.size(1) - feature_lens) / feature.size(1)).sum().item() - ) - - if params.show_alignment: - for index, id in enumerate(ids): - if id.startswith("103-1240"): - ref_text = orig_texts[index] - lattice = k2.intersect_dense( - decoding_graph, dense_fsa_vec, params.beam_size, - ) - best_path = one_best_decoding( - lattice=lattice, use_double_scores=params.use_double_scores, - ) - hyp = get_texts(best_path)[index] - hyp_text_list = [graph_compiler.token_table[i] for i in hyp] - hyp_text = " ".join(hyp_text_list) - #hyp_text = graph_compiler.sp.decode(hyp_text_list) - - logging.info(f"[utt]: {id}") - logging.info(f"[ref]: {ref_text}") - logging.info(f"[ali]: {hyp_text}") - - return loss, info - - -def compute_validation_loss( - params: AttributeDict, - model: Union[nn.Module, DDP], - graph_compiler: OtcTrainingGraphCompiler, - 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, - batch=batch, - graph_compiler=graph_compiler, - 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, - graph_compiler: OtcTrainingGraphCompiler, - scheduler: LRSchedulerType, - 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, - bypass_penalty: float = 0, - self_loop_penalty: float = 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. - graph_compiler: - It is used to convert transcripts to FSAs. - 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) - - for batch_idx, batch in enumerate(train_dl): - if batch_idx < cur_batch_idx: - continue - cur_batch_idx = batch_idx - - params.batch_idx_train += 1 - batch_size = len(batch["supervisions"]["text"]) - # batch_name = batch["supervisions"]["uttid"] - batch_name = "fake" - - with torch.cuda.amp.autocast(enabled=params.use_fp16): - loss, loss_info = compute_loss( - params=params, - model=model, - batch=batch, - graph_compiler=graph_compiler, - is_training=True, - warmup=(params.batch_idx_train / params.model_warm_step), - bypass_penalty=bypass_penalty, - self_loop_penalty=self_loop_penalty, - ) - # 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() - - try: - # loss.backward() - scaler.scale(loss).backward() - except RuntimeError as e: - if "CUDA out of memory" in str(e): - logging.error( - f"failing batch size:{batch_size} " - f"failing batch names {batch_name}" - ) - raise - - scheduler.step_batch(params.batch_idx_train) - scaler.step(optimizer) - scaler.update() - optimizer.zero_grad() - - if params.print_diagnostics and batch_idx == 30: - 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 % params.log_interval == 0: - cur_lr = scheduler.get_last_lr()[0] - logging.info( - f"Epoch {params.cur_epoch}, " - f"batch {batch_idx}, loss[{loss_info}], " - f"tot_loss[{tot_loss}], batch size: {batch_size}, " - f"lr: {cur_lr:.2e}" - ) - if loss_info["ctc_loss"] == float("inf") or loss_info["att_loss"] == float( - "inf" - ): - logging.error( - "Your loss contains inf, something goes wrong" - f"failing batch names {batch_name}" - ) - if tb_writer is not None: - tb_writer.add_scalar( - "train/learning_rate", cur_lr, params.batch_idx_train - ) - - loss_info.write_summary( - tb_writer, "train/current_", params.batch_idx_train - ) - tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) - - if batch_idx > 0 and batch_idx % params.valid_interval == 0: - logging.info("Computing validation loss") - valid_info = compute_validation_loss( - params=params, - model=model, - graph_compiler=graph_compiler, - valid_dl=valid_dl, - world_size=world_size, - ) - model.train() - logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") - if tb_writer is not None: - valid_info.write_summary( - tb_writer, "train/valid_", params.batch_idx_train - ) - - loss_value = tot_loss["loss"] / tot_loss["frames"] - params.train_loss = loss_value - if params.train_loss < params.best_train_loss: - params.best_train_epoch = params.cur_epoch - params.best_train_loss = params.train_loss - - -def run(rank, world_size, args): - """ - Args: - rank: - It is a value between 0 and `world_size-1`, which is - passed automatically by `mp.spawn()` in :func:`main`. - The node with rank 0 is responsible for saving checkpoint. - world_size: - Number of GPUs for DDP training. - args: - The return value of get_parser().parse_args() - """ - params = get_params() - params.update(vars(args)) - if params.full_libri is False: - params.valid_interval = 1600 - - fix_random_seed(params.seed) - if world_size > 1: - setup_dist(rank, world_size, params.master_port) - - setup_logger(f"{params.exp_dir}/log/log-train") - logging.info("Training started") - logging.info(params) - - if args.tensorboard and rank == 0: - tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") - else: - tb_writer = None - - lexicon = Lexicon(params.lang_dir) - # remove which will be assembled later in nnet_output - max_token_id = max(lexicon.tokens) - 1 - # add blank - num_classes = max_token_id + 1 - - device = torch.device("cpu") - if torch.cuda.is_available(): - device = torch.device("cuda", rank) - - # if "lang_bpe" in str(params.lang_dir): - # graph_compiler = BpeCtcTrainingGraphCompiler( - # params.lang_dir, - # device=device, - # sos_token="", - # eos_token="", - # ) - # elif "lang_phone" in str(params.lang_dir): - # assert params.att_rate == 0, ( - # "Attention decoder training does not support phone lang dirs " - # "at this time due to a missing symbol. Set --att-rate=0 " - # "for pure CTC training when using a phone-based lang dir." - # ) - # assert params.num_decoder_layers == 0, ( - # "Attention decoder training does not support phone lang dirs " - # "at this time due to a missing symbol. " - # "Set --num-decoder-layers=0 for pure CTC training when using " - # "a phone-based lang dir." - # ) - # graph_compiler = CtcTrainingGraphCompiler( - # lexicon, - # device=device, - # ) - # # Manually add the sos/eos ID with their default values - # # from the BPE recipe which we're adapting here. - # graph_compiler.sos_id = 1 - # graph_compiler.eos_id = 1 - # else: - # raise ValueError( - # f"Unsupported type of lang dir (we expected it to have " - # f"'lang_bpe' or 'lang_phone' in its name): {params.lang_dir}" - # ) - graph_compiler = OtcTrainingGraphCompiler( - lang_dir=params.lang_dir, - otc_token = "▁", - device=device, - ) - - logging.info("About to create model") - model = Conformer( - num_features=params.feature_dim, - nhead=params.nhead, - d_model=params.encoder_dim, - num_classes=num_classes, - subsampling_factor=params.subsampling_factor, - num_encoder_layers=params.num_encoder_layers, - num_decoder_layers=params.num_decoder_layers, - ) - - print(model) - - 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) - - 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]) - - optimizer = Eve(model.parameters(), lr=params.initial_lr) - - scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) - - if checkpoints and "optimizer" in checkpoints: - logging.info("Loading optimizer state dict") - optimizer.load_state_dict(checkpoints["optimizer"]) - - if ( - checkpoints - and "scheduler" in checkpoints - and checkpoints["scheduler"] is not None - ): - logging.info("Loading scheduler state dict") - scheduler.load_state_dict(checkpoints["scheduler"]) - - if params.print_diagnostics: - diagnostic = diagnostics.attach_diagnostics(model) - - librispeech = LibriSpeechAsrDataModule(args) - - if params.full_libri: - train_cuts = librispeech.train_all_shuf_cuts() - else: - train_cuts = librispeech.train_clean_100_cuts() - - def remove_short_and_long_utt(c: Cut): - # Keep only utterances with duration between 1 second and 20 seconds - # - # Caution: There is a reason to select 20.0 here. Please see - # ../local/display_manifest_statistics.py - # - # You should use ../local/display_manifest_statistics.py to get - # an utterance duration distribution for your dataset to select - # the threshold - return 1.0 <= c.duration <= 20.0 - - def remove_invalid_utt_ctc(c: Cut): - # Caution: We assume the subsampling factor is 4! - # num_tokens = len(sp.encode(c.supervisions[0].text, out_type=int)) - num_tokens = len(graph_compiler.texts_to_ids(c.supervisions[0].text)) - min_output_input_ratio = 0.0005 - max_output_input_ratio = 0.1 - return ( - min_output_input_ratio - < num_tokens / float(c.features.num_frames) - < max_output_input_ratio - ) - - train_cuts = train_cuts.filter(remove_short_and_long_utt) - # train_cuts = train_cuts.filter(remove_invalid_utt_ctc) - - 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 = librispeech.train_dataloaders( - train_cuts, sampler_state_dict=sampler_state_dict - ) - - valid_cuts = librispeech.dev_clean_cuts() - valid_cuts += librispeech.dev_other_cuts() - valid_dl = librispeech.valid_dataloaders(valid_cuts) - - if 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) - if checkpoints and "grad_scaler" in checkpoints: - logging.info("Loading grad scaler state dict") - scaler.load_state_dict(checkpoints["grad_scaler"]) - - bypass_penalty = params.initial_bypass_penalty - self_loop_penalty = params.initial_self_loop_penalty - 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 - - logging.info( - f"bypass penalty: {bypass_penalty}, decay: {params.bypass_penalty_decay}" - ) - logging.info( - f"self loop penalty: {self_loop_penalty}, decay: {params.self_loop_penalty_decay}" - ) - train_one_epoch( - params=params, - model=model, - model_avg=model_avg, - optimizer=optimizer, - graph_compiler=graph_compiler, - scheduler=scheduler, - train_dl=train_dl, - valid_dl=valid_dl, - scaler=scaler, - tb_writer=tb_writer, - world_size=world_size, - rank=rank, - bypass_penalty=bypass_penalty, - self_loop_penalty=self_loop_penalty, - ) - bypass_penalty *= params.bypass_penalty_decay - self_loop_penalty *= params.self_loop_penalty_decay - # params.penalty_decay = params.penalty_decay ** 0.5 - - 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 scan_pessimistic_batches_for_oom( - model: Union[nn.Module, DDP], - train_dl: torch.utils.data.DataLoader, - optimizer: torch.optim.Optimizer, - graph_compiler: OtcTrainingGraphCompiler, - 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: - # warmup = 0.0 is so that the derivs for the pruned loss stay zero - # (i.e. are not remembered by the decaying-average in adam), because - # we want to avoid these params being subject to shrinkage in adam. - with torch.cuda.amp.autocast(enabled=params.use_fp16): - loss, _ = compute_loss( - params=params, - model=model, - batch=batch, - graph_compiler=graph_compiler, - is_training=True, - warmup=0.0, - ) - loss.backward() - optimizer.step() - optimizer.zero_grad() - except RuntimeError 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]}) ..." - ) - raise - - -def main(): - parser = get_parser() - LibriSpeechAsrDataModule.add_arguments(parser) - args = parser.parse_args() - args.exp_dir = Path(args.exp_dir) - - world_size = args.world_size - assert world_size >= 1 - if world_size > 1: - mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) - else: - run(rank=0, world_size=1, args=args) - - -torch.set_num_threads(1) -torch.set_num_interop_threads(1) - -if __name__ == "__main__": - main()