enginex-mr_series-sherpa-onnx/scripts/wenet/test-onnx-streaming.py

#!/usr/bin/env python3
# Copyright      2023  Xiaomi Corp.        (authors: Fangjun Kuang)

import kaldi_native_fbank as knf
import onnxruntime as ort
import torch
import torchaudio
from torch.nn.utils.rnn import pad_sequence


class OnnxModel:
    def __init__(
        self,
        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.model = ort.InferenceSession(
            filename,
            sess_options=self.session_opts,
            providers=["CPUExecutionProvider"],
        )

        meta = self.model.get_modelmeta().custom_metadata_map
        self.left_chunks = int(meta["left_chunks"])
        self.num_blocks = int(meta["num_blocks"])
        self.chunk_size = int(meta["chunk_size"])
        self.head = int(meta["head"])
        self.output_size = int(meta["output_size"])
        self.cnn_module_kernel = int(meta["cnn_module_kernel"])
        self.right_context = int(meta["right_context"])
        self.subsampling_factor = int(meta["subsampling_factor"])

        self._init_cache()

    def _init_cache(self):
        required_cache_size = self.chunk_size * self.left_chunks

        self.attn_cache = torch.zeros(
            self.num_blocks,
            self.head,
            required_cache_size,
            self.output_size // self.head * 2,
            dtype=torch.float32,
        ).numpy()

        self.conv_cache = torch.zeros(
            self.num_blocks,
            1,
            self.output_size,
            self.cnn_module_kernel - 1,
            dtype=torch.float32,
        ).numpy()

        self.offset = torch.tensor([required_cache_size], dtype=torch.int64).numpy()

        self.required_cache_size = torch.tensor(
            [self.chunk_size * self.left_chunks], dtype=torch.int64
        ).numpy()

    def __call__(self, x: torch.Tensor) -> torch.Tensor:
        """
        Args:
          x:
            A 2-D tensor of shape (T, C)
        Returns:
          Return a 2-D tensor of shape (T, C) containing log_probs.
        """
        attn_mask = torch.ones(
            1, 1, int(self.required_cache_size + self.chunk_size), dtype=torch.bool
        )
        chunk_idx = self.offset // self.chunk_size - self.left_chunks
        if chunk_idx < self.left_chunks:
            attn_mask[
                :, :, : int(self.required_cache_size - chunk_idx * self.chunk_size)
            ] = False

        log_probs, new_attn_cache, new_conv_cache = self.model.run(
            [
                self.model.get_outputs()[0].name,
                self.model.get_outputs()[1].name,
                self.model.get_outputs()[2].name,
            ],
            {
                self.model.get_inputs()[0].name: x.unsqueeze(0).numpy(),
                self.model.get_inputs()[1].name: self.offset,
                self.model.get_inputs()[2].name: self.required_cache_size,
                self.model.get_inputs()[3].name: self.attn_cache,
                self.model.get_inputs()[4].name: self.conv_cache,
                self.model.get_inputs()[5].name: attn_mask.numpy(),
            },
        )

        self.attn_cache = new_attn_cache
        self.conv_cache = new_conv_cache

        log_probs = torch.from_numpy(log_probs)

        self.offset += log_probs.shape[1]

        return log_probs.squeeze(0)


def get_features(test_wav_filename):
    wave, sample_rate = torchaudio.load(test_wav_filename)
    audio = wave[0].contiguous()  # only use the first channel
    if sample_rate != 16000:
        audio = torchaudio.functional.resample(
            audio, orig_freq=sample_rate, new_freq=16000
        )
    audio *= 372768

    opts = knf.FbankOptions()
    opts.frame_opts.dither = 0
    opts.mel_opts.num_bins = 80
    opts.frame_opts.snip_edges = False
    opts.mel_opts.debug_mel = False

    fbank = knf.OnlineFbank(opts)
    fbank.accept_waveform(16000, audio.numpy())
    frames = []
    for i in range(fbank.num_frames_ready):
        frames.append(torch.from_numpy(fbank.get_frame(i)))
    frames = torch.stack(frames)
    return frames


def main():
    model_filename = "./model-streaming.onnx"
    model = OnnxModel(model_filename)

    filename = "./0.wav"
    x = get_features(filename)

    padding = torch.zeros(int(16000 * 0.5), 80)
    x = torch.cat([x, padding], dim=0)

    chunk_length = (
        (model.chunk_size - 1) * model.subsampling_factor + model.right_context + 1
    )
    chunk_length = int(chunk_length)
    chunk_shift = int(model.required_cache_size)
    print(chunk_length, chunk_shift)

    num_frames = x.shape[0]
    n = (num_frames - chunk_length) // chunk_shift + 1
    tokens = []
    for i in range(n):
        start = i * chunk_shift
        end = start + chunk_length
        frames = x[start:end, :]
        log_probs = model(frames)

        indexes = log_probs.argmax(dim=1)
        indexes = torch.unique_consecutive(indexes)
        indexes = indexes[indexes != 0].tolist()
        if indexes:
            tokens.extend(indexes)

    id2word = dict()
    with open("./units.txt", encoding="utf-8") as f:
        for line in f:
            word, idx = line.strip().split()
            id2word[int(idx)] = word
    text = "".join([id2word[i] for i in tokens])
    print(text)


if __name__ == "__main__":
    main()