onnx_pretrained.py

#!/usr/bin/env python3
# Copyright      2022  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.
"""
This script loads ONNX models and uses them to decode waves.
You can use the following command to get the exported models:

We use the pre-trained model from
https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless5_offline/
as an example to show how to use this file.

1. Download the pre-trained model

cd egs/wenetspeech/ASR

repo_url=https://huggingface.co/luomingshuang/icefall_asr_wenetspeech_pruned_transducer_stateless5_offline/
GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
repo=$(basename $repo_url)

pushd $repo
git lfs pull --include "data/lang_char/Linv.pt"
git lfs pull --include "exp/pretrained_epoch_4_avg_1.pt"
git lfs pull --include "exp/cpu_jit_epoch_4_avg_1_torch.1.7.1.pt"

cd exp
ln -s pretrained_epoch_9_avg_1_torch.1.7.1.pt epoch-99.pt
popd

2. Export the model to ONNX

./pruned_transducer_stateless5/export-onnx.py \
  --lang-dir $repo/data/lang_char \
  --epoch 99 \
  --avg 1 \
  --use-averaged-model 0 \
  --exp-dir $repo/exp \
  --num-encoder-layers 24 \
  --dim-feedforward 1536 \
  --nhead 8 \
  --encoder-dim 384 \
  --decoder-dim 512 \
  --joiner-dim 512

It will generate the following 3 files inside $repo/exp:

  - encoder-epoch-99-avg-1.onnx
  - decoder-epoch-99-avg-1.onnx
  - joiner-epoch-99-avg-1.onnx

3. Run this file

./pruned_transducer_stateless5/onnx_pretrained.py \
  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
  --tokens $repo/data/lang_char/tokens.txt \
  $repo/test_wavs/DEV_T0000000000.wav \
  $repo/test_wavs/DEV_T0000000001.wav \
  $repo/test_wavs/DEV_T0000000002.wav
"""

import argparse
import logging
import math
from typing import List, Tuple

import k2
import kaldifeat
import onnxruntime as ort
import torch
import torchaudio
from torch.nn.utils.rnn import pad_sequence


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

    parser.add_argument(
        "--encoder-model-filename",
        type=str,
        required=True,
        help="Path to the encoder onnx model. ",
    )

    parser.add_argument(
        "--decoder-model-filename",
        type=str,
        required=True,
        help="Path to the decoder onnx model. ",
    )

    parser.add_argument(
        "--joiner-model-filename",
        type=str,
        required=True,
        help="Path to the joiner onnx model. ",
    )

    parser.add_argument(
        "--tokens",
        type=str,
        help="""Path to tokens.txt.""",
    )

    parser.add_argument(
        "sound_files",
        type=str,
        nargs="+",
        help="The input sound file(s) to transcribe. "
        "Supported formats are those supported by torchaudio.load(). "
        "For example, wav and flac are supported. "
        "The sample rate has to be 16kHz.",
    )

    parser.add_argument(
        "--sample-rate",
        type=int,
        default=16000,
        help="The sample rate of the input sound file",
    )

    return parser


class OnnxModel:
    def __init__(
        self,
        encoder_model_filename: str,
        decoder_model_filename: str,
        joiner_model_filename: str,
    ):
        session_opts = ort.SessionOptions()
        session_opts.inter_op_num_threads = 1
        session_opts.intra_op_num_threads = 4

        self.session_opts = session_opts

        self.init_encoder(encoder_model_filename)
        self.init_decoder(decoder_model_filename)
        self.init_joiner(joiner_model_filename)

    def init_encoder(self, encoder_model_filename: str):
        self.encoder = ort.InferenceSession(
            encoder_model_filename,
            sess_options=self.session_opts,
            providers=["CPUExecutionProvider"],
        )

    def init_decoder(self, decoder_model_filename: str):
        self.decoder = ort.InferenceSession(
            decoder_model_filename,
            sess_options=self.session_opts,
            providers=["CPUExecutionProvider"],
        )

        decoder_meta = self.decoder.get_modelmeta().custom_metadata_map
        self.context_size = int(decoder_meta["context_size"])
        self.vocab_size = int(decoder_meta["vocab_size"])

        logging.info(f"context_size: {self.context_size}")
        logging.info(f"vocab_size: {self.vocab_size}")

    def init_joiner(self, joiner_model_filename: str):
        self.joiner = ort.InferenceSession(
            joiner_model_filename,
            sess_options=self.session_opts,
            providers=["CPUExecutionProvider"],
        )

        joiner_meta = self.joiner.get_modelmeta().custom_metadata_map
        self.joiner_dim = int(joiner_meta["joiner_dim"])

        logging.info(f"joiner_dim: {self.joiner_dim}")

    def run_encoder(
        self,
        x: torch.Tensor,
        x_lens: torch.Tensor,
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Args:
          x:
            A 3-D tensor of shape (N, T, C)
          x_lens:
            A 2-D tensor of shape (N,). Its dtype is torch.int64
        Returns:
          Return a tuple containing:
            - encoder_out, its shape is (N, T', joiner_dim)
            - encoder_out_lens, its shape is (N,)
        """
        out = self.encoder.run(
            [
                self.encoder.get_outputs()[0].name,
                self.encoder.get_outputs()[1].name,
            ],
            {
                self.encoder.get_inputs()[0].name: x.numpy(),
                self.encoder.get_inputs()[1].name: x_lens.numpy(),
            },
        )
        return torch.from_numpy(out[0]), torch.from_numpy(out[1])

    def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor:
        """
        Args:
          decoder_input:
            A 2-D tensor of shape (N, context_size)
        Returns:
          Return a 2-D tensor of shape (N, joiner_dim)
        """
        out = self.decoder.run(
            [self.decoder.get_outputs()[0].name],
            {self.decoder.get_inputs()[0].name: decoder_input.numpy()},
        )[0]

        return torch.from_numpy(out)

    def run_joiner(
        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
    ) -> torch.Tensor:
        """
        Args:
          encoder_out:
            A 2-D tensor of shape (N, joiner_dim)
          decoder_out:
            A 2-D tensor of shape (N, joiner_dim)
        Returns:
          Return a 2-D tensor of shape (N, vocab_size)
        """
        out = self.joiner.run(
            [self.joiner.get_outputs()[0].name],
            {
                self.joiner.get_inputs()[0].name: encoder_out.numpy(),
                self.joiner.get_inputs()[1].name: decoder_out.numpy(),
            },
        )[0]

        return torch.from_numpy(out)


def read_sound_files(
    filenames: List[str], expected_sample_rate: float
) -> List[torch.Tensor]:
    """Read a list of sound files into a list 1-D float32 torch tensors.
    Args:
      filenames:
        A list of sound filenames.
      expected_sample_rate:
        The expected sample rate of the sound files.
    Returns:
      Return a list of 1-D float32 torch tensors.
    """
    ans = []
    for f in filenames:
        wave, sample_rate = torchaudio.load(f)
        assert (
            sample_rate == expected_sample_rate
        ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
        # We use only the first channel
        ans.append(wave[0])
    return ans


def greedy_search(
    model: OnnxModel,
    encoder_out: torch.Tensor,
    encoder_out_lens: torch.Tensor,
) -> List[List[int]]:
    """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
    Args:
      model:
        The transducer model.
      encoder_out:
        A 3-D tensor of shape (N, T, joiner_dim)
      encoder_out_lens:
        A 1-D tensor of shape (N,).
    Returns:
      Return the decoded results for each utterance.
    """
    assert encoder_out.ndim == 3, encoder_out.shape
    assert encoder_out.size(0) >= 1, encoder_out.size(0)

    packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
        input=encoder_out,
        lengths=encoder_out_lens.cpu(),
        batch_first=True,
        enforce_sorted=False,
    )

    blank_id = 0  # hard-code to 0

    batch_size_list = packed_encoder_out.batch_sizes.tolist()
    N = encoder_out.size(0)

    assert torch.all(encoder_out_lens > 0), encoder_out_lens
    assert N == batch_size_list[0], (N, batch_size_list)

    context_size = model.context_size
    hyps = [[blank_id] * context_size for _ in range(N)]

    decoder_input = torch.tensor(
        hyps,
        dtype=torch.int64,
    )  # (N, context_size)

    decoder_out = model.run_decoder(decoder_input)

    offset = 0
    for batch_size in batch_size_list:
        start = offset
        end = offset + batch_size
        current_encoder_out = packed_encoder_out.data[start:end]
        # current_encoder_out's shape: (batch_size, joiner_dim)
        offset = end

        decoder_out = decoder_out[:batch_size]
        logits = model.run_joiner(current_encoder_out, decoder_out)

        # logits'shape (batch_size, vocab_size)

        assert logits.ndim == 2, logits.shape
        y = logits.argmax(dim=1).tolist()
        emitted = False
        for i, v in enumerate(y):
            if v != blank_id:
                hyps[i].append(v)
                emitted = True
        if emitted:
            # update decoder output
            decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
            decoder_input = torch.tensor(
                decoder_input,
                dtype=torch.int64,
            )
            decoder_out = model.run_decoder(decoder_input)

    sorted_ans = [h[context_size:] for h in hyps]
    ans = []
    unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
    for i in range(N):
        ans.append(sorted_ans[unsorted_indices[i]])

    return ans


@torch.no_grad()
def main():
    parser = get_parser()
    args = parser.parse_args()
    logging.info(vars(args))
    model = OnnxModel(
        encoder_model_filename=args.encoder_model_filename,
        decoder_model_filename=args.decoder_model_filename,
        joiner_model_filename=args.joiner_model_filename,
    )

    logging.info("Constructing Fbank computer")
    opts = kaldifeat.FbankOptions()
    opts.device = "cpu"
    opts.frame_opts.dither = 0
    opts.frame_opts.snip_edges = False
    opts.frame_opts.samp_freq = args.sample_rate
    opts.mel_opts.num_bins = 80
    opts.mel_opts.high_freq = -400

    fbank = kaldifeat.Fbank(opts)

    logging.info(f"Reading sound files: {args.sound_files}")
    waves = read_sound_files(
        filenames=args.sound_files,
        expected_sample_rate=args.sample_rate,
    )

    logging.info("Decoding started")
    features = fbank(waves)
    feature_lengths = [f.size(0) for f in features]

    features = pad_sequence(
        features,
        batch_first=True,
        padding_value=math.log(1e-10),
    )

    feature_lengths = torch.tensor(feature_lengths, dtype=torch.int64)
    encoder_out, encoder_out_lens = model.run_encoder(features, feature_lengths)

    hyps = greedy_search(
        model=model,
        encoder_out=encoder_out,
        encoder_out_lens=encoder_out_lens,
    )
    s = "\n"

    symbol_table = k2.SymbolTable.from_file(args.tokens)

    def token_ids_to_words(token_ids: List[int]) -> str:
        text = ""
        for i in token_ids:
            text += symbol_table[i]
        return text.replace("▁", " ").strip()

    for filename, hyp in zip(args.sound_files, hyps):
        words = token_ids_to_words(hyp)
        s += f"{filename}:\n{words}\n"
    logging.info(s)

    logging.info("Decoding Done")


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

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