icefall/egs/tedlium3/ASR/pruned_transducer_stateless/export.py

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

# This script converts several saved checkpoints
# to a single one using model averaging.
"""
Usage:
./pruned_transducer_stateless/export.py \
  --exp-dir ./pruned_transducer_stateless/exp \
  --bpe-model data/lang_bpe_500/bpe.model \
  --epoch 29 \
  --avg 13

It will generate a file exp_dir/pretrained.pt

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

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

    cd /path/to/egs/tedlium3/ASR
    ./pruned_transducer_stateless/decode.py \
        --exp-dir ./pruned_transducer_stateless/exp \
        --epoch 9999 \
        --avg 1 \
        --max-duration 1 \
        --bpe-model data/lang_bpe_500/bpe.model
"""

import argparse
import logging
from pathlib import Path

import sentencepiece as spm
import torch
from train import get_params, get_transducer_model

from icefall.checkpoint import average_checkpoints, load_checkpoint
from icefall.utils import str2bool


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=13,
        help="Number of checkpoints to average. Automatically select "
        "consecutive checkpoints before the checkpoint specified by "
        "'--epoch'. ",
    )

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

    parser.add_argument(
        "--bpe-model",
        type=str,
        default="data/lang_bpe_500/bpe.model",
        help="Path to the BPE model",
    )

    parser.add_argument(
        "--jit",
        type=str2bool,
        default=False,
        help="""True to save a model after applying torch.jit.script.
        """,
    )

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

    return parser


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

    assert args.jit is False, "Support torchscript will be added later"

    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}")

    sp = spm.SentencePieceProcessor()
    sp.load(params.bpe_model)

    # <blk> and <unk> are defined in local/train_bpe_model.py
    params.blank_id = sp.piece_to_id("<blk>")
    params.unk_id = sp.piece_to_id("<unk>")
    params.vocab_size = sp.get_piece_size()

    logging.info(params)

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

    model.to(device)

    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))

    model.eval()

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

    if params.jit:
        logging.info("Using torch.jit.script")
        model = torch.jit.script(model)
        filename = params.exp_dir / "cpu_jit.pt"
        model.save(str(filename))
        logging.info(f"Saved to {filename}")
    else:
        logging.info("Not using torch.jit.script")
        # 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()