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

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

import argparse
import logging
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 AishellAsrDataModule
from beam_search import (
    beam_search,
    fast_beam_search,
    greedy_search,
    modified_beam_search,
)
from conformer import Conformer
from decoder import Decoder
from joiner import Joiner
from model import Transducer

from icefall.checkpoint import average_checkpoints, load_checkpoint
from icefall.env import get_env_info
from icefall.lexicon import Lexicon
from icefall.utils import (
    AttributeDict,
    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=30,
        help="It specifies the checkpoint to use for decoding."
        "Note: Epoch counts from 0.",
    )
    parser.add_argument(
        "--avg",
        type=int,
        default=10,
        help="Number of checkpoints to average. Automatically select "
        "consecutive checkpoints before the checkpoint specified by "
        "'--epoch'. ",
    )

    parser.add_argument(
        "--exp-dir",
        type=str,
        default="transducer_stateless/exp",
        help="The experiment dir",
    )

    parser.add_argument(
        "--lang-dir",
        type=str,
        default="data/lang_char",
        help="The lang dir",
    )

    parser.add_argument(
        "--decoding-method",
        type=str,
        default="greedy_search",
        help="""Possible values are:
          - greedy_search
          - beam_search
          - modified_beam_search
          - fast_beam_search
        """,
    )

    parser.add_argument(
        "--beam-size",
        type=int,
        default=4,
        help="""An interger 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=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=8,
        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(
        "--max-sym-per-frame",
        type=int,
        default=3,
        help="Maximum number of symbols per frame",
    )

    parser.add_argument(
        "--export",
        type=str2bool,
        default=False,
        help="""When enabled, the averaged model is saved to
        transducer_stateless/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(
        {
            # parameters for conformer
            "feature_dim": 80,
            "embedding_dim": 512,
            "subsampling_factor": 4,
            "attention_dim": 512,
            "nhead": 4,
            "dim_feedforward": 2048,
            "num_encoder_layers": 12,
            "vgg_frontend": False,
            "env_info": get_env_info(),
        }
    )
    return params


def get_encoder_model(params: AttributeDict) -> nn.Module:
    # TODO: We can add an option to switch between Conformer and Transformer
    encoder = Conformer(
        num_features=params.feature_dim,
        output_dim=params.vocab_size,
        subsampling_factor=params.subsampling_factor,
        d_model=params.attention_dim,
        nhead=params.nhead,
        dim_feedforward=params.dim_feedforward,
        num_encoder_layers=params.num_encoder_layers,
        vgg_frontend=params.vgg_frontend,
    )
    return encoder


def get_decoder_model(params: AttributeDict) -> nn.Module:
    decoder = Decoder(
        vocab_size=params.vocab_size,
        embedding_dim=params.embedding_dim,
        blank_id=params.blank_id,
        context_size=params.context_size,
    )
    return decoder


def get_joiner_model(params: AttributeDict) -> nn.Module:
    joiner = Joiner(
        input_dim=params.vocab_size,
        inner_dim=params.embedding_dim,
        output_dim=params.vocab_size,
    )
    return joiner


def get_transducer_model(params: AttributeDict) -> nn.Module:
    encoder = get_encoder_model(params)
    decoder = get_decoder_model(params)
    joiner = get_joiner_model(params)

    model = Transducer(
        encoder=encoder,
        decoder=decoder,
        joiner=joiner,
    )
    return model


def decode_one_batch(
    params: AttributeDict,
    model: nn.Module,
    lexicon: Lexicon,
    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.
      lexicon:
        It contains the token symbol table and the word symbol table.
      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 HLG, Used
        only when --decoding_method is fast_beam_search.
    Returns:
      Return the decoding result. See above description for the format of
      the returned dict.
    """
    device = model.device
    feature = batch["inputs"]
    assert feature.ndim == 3

    feature = feature.to(device)
    # at entry, feature is (N, T, C)

    supervisions = batch["supervisions"]
    feature_lens = supervisions["num_frames"].to(device)

    encoder_out, encoder_out_lens = model.encoder(
        x=feature, x_lens=feature_lens
    )
    hyps = []

    if params.decoding_method == "fast_beam_search":
        hyp_tokens = fast_beam_search(
            model=model,
            decoding_graph=decoding_graph,
            encoder_out=encoder_out,
            encoder_out_lens=encoder_out_lens,
            beam=params.beam,
            max_contexts=params.max_contexts,
            max_states=params.max_states,
        )
        for hyp in hyp_tokens:
            hyps.append([lexicon.token_table[i] for i in hyp])
    else:
        batch_size = encoder_out.size(0)

        for i in range(batch_size):
            # fmt: off
            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
            # fmt: on
            if params.decoding_method == "greedy_search":
                hyp = greedy_search(
                    model=model,
                    encoder_out=encoder_out_i,
                    max_sym_per_frame=params.max_sym_per_frame,
                )
            elif params.decoding_method == "beam_search":
                hyp = beam_search(
                    model=model,
                    encoder_out=encoder_out_i,
                    beam=params.beam_size,
                )
            elif params.decoding_method == "modified_beam_search":
                hyp = modified_beam_search(
                    model=model,
                    encoder_out=encoder_out_i,
                    beam=params.beam_size,
                )
            else:
                raise ValueError(
                    f"Unsupported decoding method: {params.decoding_method}"
                )
            hyps.append([lexicon.token_table[i] for i in hyp])

    if params.decoding_method == "greedy_search":
        return {"greedy_search": hyps}
    elif params.decoding_method == "fast_beam_search":
        return {
            (
                f"beam_{params.beam}_"
                f"max_contexts_{params.max_contexts}_"
                f"max_states_{params.max_states}"
            ): hyps
        }
    else:
        return {f"beam_size_{params.beam_size}": hyps}


def decode_dataset(
    dl: torch.utils.data.DataLoader,
    params: AttributeDict,
    model: nn.Module,
    lexicon: Lexicon,
    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.
      lexicon:
        It contains the token symbol table and the word symbol table.
      decoding_graph:
        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
        only when --decoding_method is fast_beam_search.
    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 = 100
    else:
        log_interval = 2

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

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

        for name, hyps in hyps_dict.items():
            this_batch = []
            assert len(hyps) == len(texts)
            for hyp_words, ref_text in zip(hyps, texts):
                ref_words = ref_text.split()
                this_batch.append((ref_words, hyp_words))

            results[name].extend(this_batch)

        num_cuts += len(texts)

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

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


def save_results(
    params: AttributeDict,
    test_set_name: str,
    results_dict: Dict[str, List[Tuple[List[int], List[int]]]],
):
    test_set_wers = dict()
    for key, results in results_dict.items():
        recog_path = (
            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
        )
        store_transcripts(filename=recog_path, texts=results)

        # 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"
        )
        # we compute CER for aishell dataset.
        results_char = []
        for res in results:
            results_char.append((list("".join(res[0])), list("".join(res[1]))))
        with open(errs_filename, "w") as f:
            wer = write_error_stats(
                f, f"{test_set_name}-{key}", results_char, 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\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()
    args.exp_dir = Path(args.exp_dir)
    args.lang_dir = Path(args.lang_dir)

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

    assert params.decoding_method in (
        "greedy_search",
        "beam_search",
        "modified_beam_search",
        "fast_beam_search",
    )
    params.res_dir = params.exp_dir / params.decoding_method

    params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"

    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}"
    elif "beam_search" in params.decoding_method:
        params.suffix += f"-beam-{params.beam_size}"
    else:
        params.suffix += f"-context-{params.context_size}"
        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"

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

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

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

    lexicon = Lexicon(params.lang_dir)

    params.blank_id = 0
    params.vocab_size = max(lexicon.tokens) + 1

    logging.info(params)

    logging.info("About to create model")
    model = get_transducer_model(params)

    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.to(device)
        model.load_state_dict(average_checkpoints(filenames, device=device))

    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"
        )
        return

    model.to(device)
    model.eval()
    model.device = device

    if params.decoding_method == "fast_beam_search":
        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}")

    aishell = AishellAsrDataModule(args)
    test_cuts = aishell.test_cuts()
    test_dl = aishell.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,
            decoding_graph=decoding_graph,
        )

        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()