icefall/egs/aishell/ASR/tdnn_lstm_ctc/decode.py

#!/usr/bin/env python3
# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import argparse
import logging
from collections import defaultdict
from pathlib import Path
from typing import Dict, List, Tuple

import k2
import torch
import torch.nn as nn
from asr_datamodule import AishellAsrDataModule
from model import TdnnLstm

from icefall.checkpoint import average_checkpoints, load_checkpoint
from icefall.decode import get_lattice, nbest_decoding, one_best_decoding
from icefall.lexicon import Lexicon
from icefall.utils import (
    AttributeDict,
    get_texts,
    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=19,
        help="It specifies the checkpoint to use for decoding."
        "Note: Epoch counts from 0.",
    )
    parser.add_argument(
        "--avg",
        type=int,
        default=5,
        help="Number of checkpoints to average. Automatically select "
        "consecutive checkpoints before the checkpoint specified by "
        "'--epoch'. ",
    )
    parser.add_argument(
        "--method",
        type=str,
        default="1best",
        help="""Decoding method.
        Supported values are:
            - (1) 1best. Extract the best path from the decoding lattice as the
              decoding result.
            - (2) nbest. Extract n paths from the decoding lattice; the path
              with the highest score is the decoding result.
        """,
    )
    parser.add_argument(
        "--num-paths",
        type=int,
        default=30,
        help="""Number of paths for n-best based decoding method.
        Used only when "method" is nbest.
        """,
    )
    parser.add_argument(
        "--export",
        type=str2bool,
        default=False,
        help="""When enabled, the averaged model is saved to
        tdnn/exp/pretrained.pt. Note: only model.state_dict() is saved.
        pretrained.pt contains a dict {"model": model.state_dict()},
        which can be loaded by `icefall.checkpoint.load_checkpoint()`.
        """,
    )
    return parser


def get_params() -> AttributeDict:
    params = AttributeDict(
        {
            "exp_dir": Path("tdnn_lstm_ctc/exp/"),
            "lang_dir": Path("data/lang_phone"),
            "lm_dir": Path("data/lm"),
            # parameters for tdnn_lstm_ctc
            "subsampling_factor": 3,
            "feature_dim": 80,
            # parameters for decoding
            "search_beam": 20,
            "output_beam": 7,
            "min_active_states": 30,
            "max_active_states": 10000,
            "use_double_scores": True,
        }
    )
    return params


def decode_one_batch(
    params: AttributeDict,
    model: nn.Module,
    HLG: k2.Fsa,
    batch: dict,
    lexicon: Lexicon,
) -> Dict[str, List[List[int]]]:
    """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 the decoding method is 1best, the key is the string
               `no_rescore`. If the decoding method is nbest, the key is the
               string `no_rescore-xxx`, xxx is the num_paths.

        - 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.

      model:
        The neural model.
      HLG:
        The decoding graph.
      batch:
        It is the return value from iterating
        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
        for the format of the `batch`.
      lexicon:
        It contains word symbol table.
    Returns:
      Return the decoding result. See above description for the format of
      the returned dict.
    """
    device = HLG.device
    feature = batch["inputs"]
    assert feature.ndim == 3
    feature = feature.to(device)
    # at entry, feature is [N, T, C]

    feature = feature.permute(0, 2, 1)  # now feature is [N, C, T]

    nnet_output = model(feature)
    # nnet_output is [N, T, C]

    supervisions = batch["supervisions"]

    supervision_segments = torch.stack(
        (
            supervisions["sequence_idx"],
            supervisions["start_frame"] // params.subsampling_factor,
            supervisions["num_frames"] // params.subsampling_factor,
        ),
        1,
    ).to(torch.int32)

    lattice = get_lattice(
        nnet_output=nnet_output,
        decoding_graph=HLG,
        supervision_segments=supervision_segments,
        search_beam=params.search_beam,
        output_beam=params.output_beam,
        min_active_states=params.min_active_states,
        max_active_states=params.max_active_states,
    )

    assert 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,
        )
        key = f"no_rescore-{params.num_paths}"
    hyps = get_texts(best_path)
    hyps = [[lexicon.word_table[i] for i in ids] for ids in hyps]
    return {key: hyps}


def decode_dataset(
    dl: torch.utils.data.DataLoader,
    params: AttributeDict,
    model: nn.Module,
    HLG: k2.Fsa,
    lexicon: Lexicon,
) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
    """Decode dataset.

    Args:
      dl:
        PyTorch's dataloader containing the dataset to decode.
      params:
        It is returned by :func:`get_params`.
      model:
        The neural model.
      HLG:
        The decoding graph.
      lexicon:
        It contains word symbol table.
    Returns:
      Return a dict, whose key may be "no-rescore" if decoding method is 1best,
      or it may be "no-rescoer-100" if decoding method is nbest.
      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.
    """
    results = []

    num_cuts = 0

    try:
        num_batches = len(dl)
    except TypeError:
        num_batches = "?"

    results = defaultdict(list)
    for batch_idx, batch in enumerate(dl):
        texts = batch["supervisions"]["text"]
        cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]

        hyps_dict = decode_one_batch(
            params=params,
            model=model,
            HLG=HLG,
            batch=batch,
            lexicon=lexicon,
        )

        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)

        num_cuts += len(batch["supervisions"]["text"])

        if batch_idx % 100 == 0:
            batch_str = f"{batch_idx}/{num_batches}"

            logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
    return results


def save_results(
    params: AttributeDict,
    test_set_name: str,
    results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
):
    test_set_wers = dict()
    for key, results in results_dict.items():
        recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt"
        results = sorted(results)
        store_transcripts(filename=recog_path, texts=results, char_level=True)
        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"
        # We compute CER for aishell dataset.
        results_char = []
        for res in results:
            results_char.append((res[0], list("".join(res[1])), list("".join(res[2]))))
        with open(errs_filename, "w") as f:
            wer = write_error_stats(
                f,
                f"{test_set_name}-{key}",
                results_char,
                enable_log=True,
                compute_CER=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.exp_dir / f"cer-summary-{test_set_name}.txt"
    with open(errs_info, "w") as f:
        print("settings\tCER", file=f)
        for key, val in test_set_wers:
            print("{}\t{}".format(key, val), file=f)

    s = "\nFor {}, CER 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()
    AishellAsrDataModule.add_arguments(parser)
    args = parser.parse_args()

    params = get_params()
    params.update(vars(args))

    setup_logger(f"{params.exp_dir}/log/log-decode")
    logging.info("Decoding started")
    logging.info(params)

    lexicon = Lexicon(params.lang_dir)
    max_phone_id = max(lexicon.tokens)

    device = torch.device("cpu")
    if torch.cuda.is_available():
        device = torch.device("cuda", 0)

    logging.info(f"device: {device}")

    HLG = k2.Fsa.from_dict(torch.load(f"{params.lang_dir}/HLG.pt", map_location="cpu", weights_only=False))
    HLG = HLG.to(device)
    assert HLG.requires_grad is False

    if not hasattr(HLG, "lm_scores"):
        HLG.lm_scores = HLG.scores.clone()

    model = TdnnLstm(
        num_features=params.feature_dim,
        num_classes=max_phone_id + 1,  # +1 for the blank symbol
        subsampling_factor=params.subsampling_factor,
    )
    if 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.load_state_dict(average_checkpoints(filenames))

    if params.export:
        logging.info(f"Export averaged model to {params.exp_dir}/pretrained.pt")
        torch.save({"model": model.state_dict()}, f"{params.exp_dir}/pretrained.pt")

    model.to(device)
    model.eval()

    # we need cut ids to display recognition results.
    args.return_cuts = True
    aishell = AishellAsrDataModule(args)
    test_cuts = aishell.test_cuts()
    test_dl = aishell.test_dataloaders(test_cuts)
    # CAUTION: `test_sets` is for displaying only.
    # If you want to skip test-clean, you have to skip
    # it inside the for loop. That is, use
    #
    #   if test_set == 'test-clean': continue
    #
    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,
            HLG=HLG,
            lexicon=lexicon,
        )

        save_results(params=params, test_set_name=test_set, results_dict=results_dict)

    logging.info("Done!")


if __name__ == "__main__":
    main()