icefall/egs/librispeech/ASR/lstm_transducer_stateless2/export.py

#!/usr/bin/env python3
# flake8: noqa
#
# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, Zengwei Yao)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# This script converts several saved checkpoints
# to a single one using model averaging.
"""

Usage:

(1) Export to torchscript model using torch.jit.trace()

./lstm_transducer_stateless2/export.py \
  --exp-dir ./lstm_transducer_stateless2/exp \
  --tokens ./data/lang_bpe_500/tokens.txt \
  --epoch 35 \
  --avg 10 \
  --jit-trace 1

It will generate 3 files: `encoder_jit_trace.pt`,
`decoder_jit_trace.pt`, and `joiner_jit_trace.pt`.

(2) Export `model.state_dict()`

./lstm_transducer_stateless2/export.py \
  --exp-dir ./lstm_transducer_stateless2/exp \
  --tokens ./data/lang_bpe_500/tokens.txt \
  --epoch 35 \
  --avg 10

It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
load it by `icefall.checkpoint.load_checkpoint()`.

To use the generated file with `lstm_transducer_stateless2/decode.py`,
you can do:

    cd /path/to/exp_dir
    ln -s pretrained.pt epoch-9999.pt

    cd /path/to/egs/librispeech/ASR
    ./lstm_transducer_stateless2/decode.py \
        --exp-dir ./lstm_transducer_stateless2/exp \
        --epoch 9999 \
        --avg 1 \
        --max-duration 600 \
        --decoding-method greedy_search \
        --bpe-model data/lang_bpe_500/bpe.model

Check ./pretrained.py for its usage.

Note: If you don't want to train a model from scratch, we have
provided one for you. You can get it at

https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03

with the following commands:

    sudo apt-get install git-lfs
    git lfs install
    git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
    # You will find the pre-trained models in icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp
"""

import argparse
import logging
from pathlib import Path

import k2
import torch
import torch.nn as nn
from scaling_converter import convert_scaled_to_non_scaled
from train import add_model_arguments, get_params, get_transducer_model

from icefall.checkpoint import (
    average_checkpoints,
    average_checkpoints_with_averaged_model,
    find_checkpoints,
    load_checkpoint,
)
from icefall.utils import num_tokens, str2bool


def get_parser():
    parser = argparse.ArgumentParser(
        formatter_class=argparse.ArgumentDefaultsHelpFormatter
    )

    parser.add_argument(
        "--epoch",
        type=int,
        default=28,
        help="""It specifies the checkpoint to use for averaging.
        Note: Epoch counts from 0.
        You can specify --avg to use more checkpoints for model averaging.""",
    )

    parser.add_argument(
        "--iter",
        type=int,
        default=0,
        help="""If positive, --epoch is ignored and it
        will use the checkpoint exp_dir/checkpoint-iter.pt.
        You can specify --avg to use more checkpoints for model averaging.
        """,
    )

    parser.add_argument(
        "--avg",
        type=int,
        default=15,
        help="Number of checkpoints to average. Automatically select "
        "consecutive checkpoints before the checkpoint specified by "
        "'--epoch' and '--iter'",
    )

    parser.add_argument(
        "--use-averaged-model",
        type=str2bool,
        default=True,
        help="Whether to load averaged model. Currently it only supports "
        "using --epoch. If True, it would decode with the averaged model "
        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
        "Actually only the models with epoch number of `epoch-avg` and "
        "`epoch` are loaded for averaging. ",
    )

    parser.add_argument(
        "--exp-dir",
        type=str,
        default="pruned_transducer_stateless3/exp",
        help="""It specifies the directory where all training related
        files, e.g., checkpoints, log, etc, are saved
        """,
    )

    parser.add_argument(
        "--tokens",
        type=str,
        default="data/lang_bpe_500/tokens.txt",
        help="Path to the tokens.txt.",
    )

    parser.add_argument(
        "--jit-trace",
        type=str2bool,
        default=False,
        help="""True to save a model after applying torch.jit.trace.
        It will generate 3 files:
         - encoder_jit_trace.pt
         - decoder_jit_trace.pt
         - joiner_jit_trace.pt

        Check ./jit_pretrained.py for how to use them.
        """,
    )

    parser.add_argument(
        "--context-size",
        type=int,
        default=2,
        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
    )

    add_model_arguments(parser)

    return parser


def export_encoder_model_jit_trace(
    encoder_model: nn.Module,
    encoder_filename: str,
) -> None:
    """Export the given encoder model with torch.jit.trace()

    Note: The warmup argument is fixed to 1.

    Args:
      encoder_model:
        The input encoder model
      encoder_filename:
        The filename to save the exported model.
    """
    x = torch.zeros(1, 100, 80, dtype=torch.float32)
    x_lens = torch.tensor([100], dtype=torch.int64)
    states = encoder_model.get_init_states()

    traced_model = torch.jit.trace(encoder_model, (x, x_lens, states))
    traced_model.save(encoder_filename)
    logging.info(f"Saved to {encoder_filename}")


def export_decoder_model_jit_trace(
    decoder_model: nn.Module,
    decoder_filename: str,
) -> None:
    """Export the given decoder model with torch.jit.trace()

    Note: The argument need_pad is fixed to False.

    Args:
      decoder_model:
        The input decoder model
      decoder_filename:
        The filename to save the exported model.
    """
    # TODO(fangjun): Change the function name since we are actually using
    # torch.jit.script instead of torch.jit.trace
    traced_model = torch.jit.script(decoder_model)
    traced_model.save(decoder_filename)
    logging.info(f"Saved to {decoder_filename}")


def export_joiner_model_jit_trace(
    joiner_model: nn.Module,
    joiner_filename: str,
) -> None:
    """Export the given joiner model with torch.jit.trace()

    Note: The argument project_input is fixed to True. A user should not
    project the encoder_out/decoder_out by himself/herself. The exported joiner
    will do that for the user.

    Args:
      joiner_model:
        The input joiner model
      joiner_filename:
        The filename to save the exported model.

    """
    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)

    traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out))
    traced_model.save(joiner_filename)
    logging.info(f"Saved to {joiner_filename}")


@torch.no_grad()
def main():
    args = get_parser().parse_args()
    args.exp_dir = Path(args.exp_dir)

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

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

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

    # Load tokens.txt here
    token_table = k2.SymbolTable.from_file(params.tokens)

    # Load id of the <blk> token and the vocab size, <blk> is
    # defined in local/train_bpe_model.py
    params.blank_id = token_table["<blk>"]
    params.vocab_size = num_tokens(token_table) + 1  # +1 for <blk>

    logging.info(params)

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

    num_param = sum([p.numel() for p in model.parameters()])
    logging.info(f"Number of model parameters: {num_param}")

    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),
                strict=False,
            )
        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),
                strict=False,
            )
    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,
                ),
                strict=False,
            )
        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,
                ),
                strict=False,
            )

    model.to("cpu")
    model.eval()

    if params.jit_trace is True:
        convert_scaled_to_non_scaled(model, inplace=True)
        logging.info("Using torch.jit.trace()")
        encoder_filename = params.exp_dir / "encoder_jit_trace.pt"
        export_encoder_model_jit_trace(model.encoder, encoder_filename)

        decoder_filename = params.exp_dir / "decoder_jit_trace.pt"
        export_decoder_model_jit_trace(model.decoder, decoder_filename)

        joiner_filename = params.exp_dir / "joiner_jit_trace.pt"
        export_joiner_model_jit_trace(model.joiner, joiner_filename)
    else:
        logging.info("Not using torchscript")
        # Save it using a format so that it can be loaded
        # by :func:`load_checkpoint`
        filename = params.exp_dir / "pretrained.pt"
        torch.save({"model": model.state_dict()}, str(filename))
        logging.info(f"Saved to {filename}")


if __name__ == "__main__":
    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"

    logging.basicConfig(format=formatter, level=logging.INFO)
    main()