Merge pull request #761 from alibaba-damo-academy/dev_add_english_paraformer

nichongjia-2007

2023-07-21 3c7909a9f9744aeedfef15942cdc00abf890cb9f

Merge pull request #761 from alibaba-damo-academy/dev_add_english_paraformer

Dev add english paraformer

 egs_modelscope/asr/paraformer/speech_paraformer_asr-en-16k-vocab4199-pytorch/README.md

New file
@@ -0,0 +1 @@
../../TEMPLATE/README.md

 egs_modelscope/asr/paraformer/speech_paraformer_asr-en-16k-vocab4199-pytorch/demo.py

New file
@@ -0,0 +1,11 @@
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks

inference_pipeline = pipeline(
    task=Tasks.auto_speech_recognition,
    model='damo/damo/speech_paraformer_asr-en-16k-vocab4199-pytorch',
    model_revision="v1.0.1",
)
audio_in='https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_en.wav'
rec_result = inference_pipeline(audio_in=audio_in)
print(rec_result)

 egs_modelscope/asr/paraformer/speech_paraformer_asr-en-16k-vocab4199-pytorch/finetune.py

New file
@@ -0,0 +1,36 @@
import os

from modelscope.metainfo import Trainers
from modelscope.trainers import build_trainer

from funasr.datasets.ms_dataset import MsDataset
from funasr.utils.modelscope_param import modelscope_args


def modelscope_finetune(params):
    if not os.path.exists(params.output_dir):
        os.makedirs(params.output_dir, exist_ok=True)
    # dataset split ["train", "validation"]
    ds_dict = MsDataset.load(params.data_path)
    kwargs = dict(
        model=params.model,
        data_dir=ds_dict,
        dataset_type=params.dataset_type,
        work_dir=params.output_dir,
        batch_bins=params.batch_bins,
        max_epoch=params.max_epoch,
        lr=params.lr)
    trainer = build_trainer(Trainers.speech_asr_trainer, default_args=kwargs)
    trainer.train()


if __name__ == '__main__':
    params = modelscope_args(model="damo/speech_paraformer_asr-en-16k-vocab4199-pytorch", data_path="./data")
    params.output_dir = "./checkpoint"              # m模型保存路径
    params.data_path = "./example_data/"            # 数据路径
    params.dataset_type = "small"                   # 小数据量设置small，若数据量大于1000小时，请使用large
    params.batch_bins = 2000                       # batch size，如果dataset_type="small"，batch_bins单位为fbank特征帧数，如果dataset_type="large"，batch_bins单位为毫秒，
    params.max_epoch = 50                           # 最大训练轮数
    params.lr = 0.00005                             # 设置学习率
    
    modelscope_finetune(params)

 egs_modelscope/asr/paraformer/speech_paraformer_asr-en-16k-vocab4199-pytorch/infer.py

New file
@@ -0,0 +1 @@
../../TEMPLATE/infer.py

 egs_modelscope/asr/paraformer/speech_paraformer_asr-en-16k-vocab4199-pytorch/infer.sh

New file
@@ -0,0 +1,103 @@
#!/usr/bin/env bash

set -e
set -u
set -o pipefail

stage=1
stop_stage=2
model="damo/speech_paraformer_asr-en-16k-vocab4199-pytorch"
data_dir="./data/test"
output_dir="./results"
batch_size=64
gpu_inference=true    # whether to perform gpu decoding
gpuid_list="0,1"    # set gpus, e.g., gpuid_list="0,1"
njob=64    # the number of jobs for CPU decoding, if gpu_inference=false, use CPU decoding, please set njob
checkpoint_dir=
checkpoint_name="valid.cer_ctc.ave.pb"

. utils/parse_options.sh || exit 1;

if ${gpu_inference} == "true"; then
    nj=$(echo $gpuid_list | awk -F "," '{print NF}')
else
    nj=$njob
    batch_size=1
    gpuid_list=""
    for JOB in $(seq ${nj}); do
        gpuid_list=$gpuid_list"-1,"
    done
fi

mkdir -p $output_dir/split
split_scps=""
for JOB in $(seq ${nj}); do
    split_scps="$split_scps $output_dir/split/wav.$JOB.scp"
done
perl utils/split_scp.pl ${data_dir}/wav.scp ${split_scps}

if [ -n "${checkpoint_dir}" ]; then
  python utils/prepare_checkpoint.py ${model} ${checkpoint_dir} ${checkpoint_name}
  model=${checkpoint_dir}/${model}
fi

if [ $stage -le 1 ] && [ $stop_stage -ge 1 ];then
    echo "Decoding ..."
    gpuid_list_array=(${gpuid_list//,/ })
    for JOB in $(seq ${nj}); do
        {
        id=$((JOB-1))
        gpuid=${gpuid_list_array[$id]}
        mkdir -p ${output_dir}/output.$JOB
        python infer.py \
            --model ${model} \
            --audio_in ${output_dir}/split/wav.$JOB.scp \
            --output_dir ${output_dir}/output.$JOB \
            --batch_size ${batch_size} \
            --gpuid ${gpuid}
        }&
    done
    wait

    mkdir -p ${output_dir}/1best_recog
    for f in token score text; do
        if [ -f "${output_dir}/output.1/1best_recog/${f}" ]; then
          for i in $(seq "${nj}"); do
              cat "${output_dir}/output.${i}/1best_recog/${f}"
          done | sort -k1 >"${output_dir}/1best_recog/${f}"
        fi
    done
fi

if [ $stage -le 2 ] && [ $stop_stage -ge 2 ];then
    echo "Computing WER ..."
    cp ${output_dir}/1best_recog/text ${output_dir}/1best_recog/text.proc
    cp ${data_dir}/text ${output_dir}/1best_recog/text.ref
    python utils/compute_wer.py ${output_dir}/1best_recog/text.ref ${output_dir}/1best_recog/text.proc ${output_dir}/1best_recog/text.cer
    tail -n 3 ${output_dir}/1best_recog/text.cer
fi

if [ $stage -le 3 ] && [ $stop_stage -ge 3 ];then
    echo "SpeechIO TIOBE textnorm"
    echo "$0 --> Normalizing REF text ..."
    ./utils/textnorm_zh.py \
        --has_key --to_upper \
        ${data_dir}/text \
        ${output_dir}/1best_recog/ref.txt

    echo "$0 --> Normalizing HYP text ..."
    ./utils/textnorm_zh.py \
        --has_key --to_upper \
        ${output_dir}/1best_recog/text.proc \
        ${output_dir}/1best_recog/rec.txt
    grep -v $'\t$' ${output_dir}/1best_recog/rec.txt > ${output_dir}/1best_recog/rec_non_empty.txt

    echo "$0 --> computing WER/CER and alignment ..."
    ./utils/error_rate_zh \
        --tokenizer char \
        --ref ${output_dir}/1best_recog/ref.txt \
        --hyp ${output_dir}/1best_recog/rec_non_empty.txt \
        ${output_dir}/1best_recog/DETAILS.txt | tee ${output_dir}/1best_recog/RESULTS.txt
    rm -rf ${output_dir}/1best_recog/rec.txt ${output_dir}/1best_recog/rec_non_empty.txt
fi


 egs_modelscope/asr/paraformer/speech_paraformer_asr-en-16k-vocab4199-pytorch/utils

New file
@@ -0,0 +1 @@
../../../../egs/aishell/transformer/utils

 funasr/export/export_conformer.py

New file
@@ -0,0 +1,151 @@
import json
from typing import Union, Dict
from pathlib import Path

import os
import logging
import torch

from funasr.export.models import get_model
import numpy as np
import random
from funasr.utils.types import str2bool, str2triple_str
# torch_version = float(".".join(torch.__version__.split(".")[:2]))
# assert torch_version > 1.9

class ModelExport:
    def __init__(
        self,
        cache_dir: Union[Path, str] = None,
        onnx: bool = True,
        device: str = "cpu",
        quant: bool = True,
        fallback_num: int = 0,
        audio_in: str = None,
        calib_num: int = 200,
        model_revision: str = None,
    ):
        self.set_all_random_seed(0)

        self.cache_dir = cache_dir
        self.export_config = dict(
            feats_dim=560,
            onnx=False,
        )
        
        self.onnx = onnx
        self.device = device
        self.quant = quant
        self.fallback_num = fallback_num
        self.frontend = None
        self.audio_in = audio_in
        self.calib_num = calib_num
        self.model_revision = model_revision

    def _export(
        self,
        model,
        model_dir: str = None,
        verbose: bool = False,
    ):

        export_dir = model_dir
        os.makedirs(export_dir, exist_ok=True)

        self.export_config["model_name"] = "model"
        model = get_model(
            model,
            self.export_config,
        )
        model.eval()

        if self.onnx:
            self._export_onnx(model, verbose, export_dir)

        print("output dir: {}".format(export_dir))

    def _export_onnx(self, model, verbose, path):
        model._export_onnx(verbose, path)

    def set_all_random_seed(self, seed: int):
        random.seed(seed)
        np.random.seed(seed)
        torch.random.manual_seed(seed)

    def parse_audio_in(self, audio_in):
        
        wav_list, name_list = [], []
        if audio_in.endswith(".scp"):
            f = open(audio_in, 'r')
            lines = f.readlines()[:self.calib_num]
            for line in lines:
                name, path = line.strip().split()
                name_list.append(name)
                wav_list.append(path)
        else:
            wav_list = [audio_in,]
            name_list = ["test",]
        return wav_list, name_list
    
    def load_feats(self, audio_in: str = None):
        import torchaudio

        wav_list, name_list = self.parse_audio_in(audio_in)
        feats = []
        feats_len = []
        for line in wav_list:
            path = line.strip()
            waveform, sampling_rate = torchaudio.load(path)
            if sampling_rate != self.frontend.fs:
                waveform = torchaudio.transforms.Resample(orig_freq=sampling_rate,
                                                          new_freq=self.frontend.fs)(waveform)
            fbank, fbank_len = self.frontend(waveform, [waveform.size(1)])
            feats.append(fbank)
            feats_len.append(fbank_len)
        return feats, feats_len

    def export(self,
               mode: str = None,
               ):

        if mode.startswith('conformer'):
            from funasr.tasks.asr import ASRTask
            config = os.path.join(model_dir, 'config.yaml')
            model_file = os.path.join(model_dir, 'model.pb')
            cmvn_file = os.path.join(model_dir, 'am.mvn')
            model, asr_train_args = ASRTask.build_model_from_file(
                config, model_file, cmvn_file, 'cpu'
            )
            self.frontend = model.frontend
            self.export_config["feats_dim"] = 560

        self._export(model, self.cache_dir)

if __name__ == '__main__':
    import argparse
    parser = argparse.ArgumentParser()
    # parser.add_argument('--model-name', type=str, required=True)
    parser.add_argument('--model-name', type=str, action="append", required=True, default=[])
    parser.add_argument('--export-dir', type=str, required=True)
    parser.add_argument('--type', type=str, default='onnx', help='["onnx", "torch"]')
    parser.add_argument('--device', type=str, default='cpu', help='["cpu", "cuda"]')
    parser.add_argument('--quantize', type=str2bool, default=False, help='export quantized model')
    parser.add_argument('--fallback-num', type=int, default=0, help='amp fallback number')
    parser.add_argument('--audio_in', type=str, default=None, help='["wav", "wav.scp"]')
    parser.add_argument('--calib_num', type=int, default=200, help='calib max num')
    parser.add_argument('--model_revision', type=str, default=None, help='model_revision')
    args = parser.parse_args()

    export_model = ModelExport(
        cache_dir=args.export_dir,
        onnx=args.type == 'onnx',
        device=args.device,
        quant=args.quantize,
        fallback_num=args.fallback_num,
        audio_in=args.audio_in,
        calib_num=args.calib_num,
        model_revision=args.model_revision,
    )
    for model_name in args.model_name:
        print("export model: {}".format(model_name))
        export_model.export(model_name)

 funasr/export/models/__init__.py

@@ -1,6 +1,8 @@
from funasr.models.e2e_asr_paraformer import Paraformer, BiCifParaformer
from funasr.export.models.e2e_asr_paraformer import Paraformer as Paraformer_export
from funasr.export.models.e2e_asr_paraformer import BiCifParaformer as BiCifParaformer_export
from funasr.export.models.e2e_asr_conformer import Conformer as Conformer_export

from funasr.models.e2e_vad import E2EVadModel
from funasr.export.models.e2e_vad import E2EVadModel as E2EVadModel_export
from funasr.models.target_delay_transformer import TargetDelayTransformer
@@ -14,6 +16,8 @@
        return BiCifParaformer_export(model, **export_config)
    elif isinstance(model, Paraformer):
        return Paraformer_export(model, **export_config)
    elif isinstance(model, Conformer_export):
        return Conformer_export(model, **export_config)
    elif isinstance(model, E2EVadModel):
        return E2EVadModel_export(model, **export_config)
    elif isinstance(model, PunctuationModel):

 funasr/export/models/decoder/xformer_decoder.py

New file
@@ -0,0 +1,121 @@
import os

import torch
import torch.nn as nn

from funasr.modules.attention import MultiHeadedAttention

from funasr.export.models.modules.decoder_layer import DecoderLayer as OnnxDecoderLayer
from funasr.export.models.language_models.embed import Embedding
from funasr.export.models.modules.multihead_att import \
    OnnxMultiHeadedAttention

from funasr.export.utils.torch_function import MakePadMask, subsequent_mask

class XformerDecoder(nn.Module):
    def __init__(self,
                 model,
                 max_seq_len = 512,
                 model_name = 'decoder',
                 onnx: bool = True,):
        super().__init__()
        self.embed = Embedding(model.embed, max_seq_len)
        self.model = model
        if onnx:
            self.make_pad_mask = MakePadMask(max_seq_len, flip=False)
        else:
            self.make_pad_mask = subsequent_mask(max_seq_len, flip=False)

        if isinstance(self.model.decoders[0].self_attn, MultiHeadedAttention):
            self.num_heads = self.model.decoders[0].self_attn.h
            self.hidden_size = self.model.decoders[0].self_attn.linear_out.out_features

        # replace multi-head attention module into customized module.
        for i, d in enumerate(self.model.decoders):
            # d is DecoderLayer
            if isinstance(d.self_attn, MultiHeadedAttention):
                d.self_attn = OnnxMultiHeadedAttention(d.self_attn)
            if isinstance(d.src_attn, MultiHeadedAttention):
                d.src_attn = OnnxMultiHeadedAttention(d.src_attn)
            self.model.decoders[i] = OnnxDecoderLayer(d)

        self.model_name = model_name

    def prepare_mask(self, mask):
        mask_3d_btd = mask[:, :, None]
        if len(mask.shape) == 2:
            mask_4d_bhlt = 1 - mask[:, None, None, :]
        elif len(mask.shape) == 3:
            mask_4d_bhlt = 1 - mask[:, None, :]

        mask_4d_bhlt = mask_4d_bhlt * -10000.0
        return mask_3d_btd, mask_4d_bhlt

    def forward(self,
                tgt,
                memory,
                cache):

        mask = subsequent_mask(tgt.size(-1)).unsqueeze(0)  # (B, T)

        x = self.embed(tgt)
        mask = self.prepare_mask(mask)
        new_cache = []
        for c, decoder in zip(cache, self.model.decoders):
            x, mask = decoder(x, mask, memory, None, c)
            new_cache.append(x)
            x = x[:, 1:, :]

        if self.model.normalize_before:
            y = self.model.after_norm(x[:, -1])
        else:
            y = x[:, -1]

        if self.model.output_layer is not None:
            y = torch.log_softmax(self.model.output_layer(y), dim=-1)
        return y, new_cache

    def get_dummy_inputs(self, enc_size):
        tgt = torch.LongTensor([0]).unsqueeze(0)
        memory = torch.randn(1, 100, enc_size)
        cache_num = len(self.model.decoders)
        cache = [
            torch.zeros((1, 1, self.model.decoders[0].size))
            for _ in range(cache_num)
        ]
        return (tgt, memory, cache)

    def is_optimizable(self):
        return True

    def get_input_names(self):
        cache_num = len(self.model.decoders)
        return ["tgt", "memory"] + [
            "cache_%d" % i for i in range(cache_num)
        ]

    def get_output_names(self):
        cache_num = len(self.model.decoders)
        return ["y"] + ["out_cache_%d" % i for i in range(cache_num)]

    def get_dynamic_axes(self):
        ret = {
            "tgt": {0: "tgt_batch", 1: "tgt_length"},
            "memory": {0: "memory_batch", 1: "memory_length"},
        }
        cache_num = len(self.model.decoders)
        ret.update(
            {
                "cache_%d" % d: {0: "cache_%d_batch" % d, 2: "cache_%d_length" % d}
                for d in range(cache_num)
            }
        )
        return ret

    def get_model_config(self, path):
        return {
            "dec_type": "XformerDecoder",
            "model_path": os.path.join(path, f"{self.model_name}.onnx"),
            "n_layers": len(self.model.decoders),
            "odim": self.model.decoders[0].size,
        }

 funasr/export/models/e2e_asr_conformer.py

New file
@@ -0,0 +1,69 @@
import os
import logging
import torch
import torch.nn as nn

from funasr.export.utils.torch_function import MakePadMask
from funasr.export.utils.torch_function import sequence_mask
from funasr.models.encoder.conformer_encoder import ConformerEncoder
from funasr.models.decoder.transformer_decoder import TransformerDecoder
from funasr.export.models.encoder.conformer_encoder import ConformerEncoder as ConformerEncoder_export
from funasr.export.models.decoder.xformer_decoder import XformerDecoder as TransformerDecoder_export

class Conformer(nn.Module):
    """
    export conformer into onnx format
    """

    def __init__(
            self,
            model,
            max_seq_len=512,
            feats_dim=560,
            model_name='model',
            **kwargs,
    ):
        super().__init__()
        onnx = False
        if "onnx" in kwargs:
            onnx = kwargs["onnx"]
        if isinstance(model.encoder, ConformerEncoder):
            self.encoder = ConformerEncoder_export(model.encoder, onnx=onnx)
        elif isinstance(model.decoder, TransformerDecoder):
            self.decoder = TransformerDecoder_export(model.decoder, onnx=onnx)
        
        self.feats_dim = feats_dim
        self.model_name = model_name

        if onnx:
            self.make_pad_mask = MakePadMask(max_seq_len, flip=False)
        else:
            self.make_pad_mask = sequence_mask(max_seq_len, flip=False)

    def _export_model(self, model, verbose, path):
        dummy_input = model.get_dummy_inputs()
        model_script = model
        model_path = os.path.join(path, f'{model.model_name}.onnx')
        if not os.path.exists(model_path):
            torch.onnx.export(
                model_script,
                dummy_input,
                model_path,
                verbose=verbose,
                opset_version=14,
                input_names=model.get_input_names(),
                output_names=model.get_output_names(),
                dynamic_axes=model.get_dynamic_axes()
            )

    def _export_encoder_onnx(self, verbose, path):
        model_encoder = self.encoder
        self._export_model(model_encoder, verbose, path)

    def _export_decoder_onnx(self, verbose, path):
        model_decoder = self.decoder
        self._export_model(model_decoder, verbose, path)

    def _export_onnx(self, verbose, path):
        self._export_encoder_onnx(verbose, path)
        self._export_decoder_onnx(verbose, path)

 funasr/export/models/language_models/__init__.py

 funasr/export/models/language_models/embed.py

New file
@@ -0,0 +1,403 @@
"""Positional Encoding Module."""

import math

import torch
import torch.nn as nn
from funasr.modules.embedding import (
    LegacyRelPositionalEncoding, PositionalEncoding, RelPositionalEncoding,
    ScaledPositionalEncoding, StreamPositionalEncoding)
from funasr.modules.subsampling import (
    Conv2dSubsampling, Conv2dSubsampling2, Conv2dSubsampling6,
    Conv2dSubsampling8)
from funasr.modules.subsampling_without_posenc import \
    Conv2dSubsamplingWOPosEnc

from funasr.export.models.language_models.subsampling import (
    OnnxConv2dSubsampling, OnnxConv2dSubsampling2, OnnxConv2dSubsampling6,
    OnnxConv2dSubsampling8)


def get_pos_emb(pos_emb, max_seq_len=512, use_cache=True):
    if isinstance(pos_emb, LegacyRelPositionalEncoding):
        return OnnxLegacyRelPositionalEncoding(pos_emb, max_seq_len, use_cache)
    elif isinstance(pos_emb, ScaledPositionalEncoding):
        return OnnxScaledPositionalEncoding(pos_emb, max_seq_len, use_cache)
    elif isinstance(pos_emb, RelPositionalEncoding):
        return OnnxRelPositionalEncoding(pos_emb, max_seq_len, use_cache)
    elif isinstance(pos_emb, PositionalEncoding):
        return OnnxPositionalEncoding(pos_emb, max_seq_len, use_cache)
    elif isinstance(pos_emb, StreamPositionalEncoding):
        return OnnxStreamPositionalEncoding(pos_emb, max_seq_len, use_cache)
    elif (isinstance(pos_emb, nn.Sequential) and len(pos_emb) == 0) or (
        isinstance(pos_emb, Conv2dSubsamplingWOPosEnc)
    ):
        return pos_emb
    else:
        raise ValueError("Embedding model is not supported.")


class Embedding(nn.Module):
    def __init__(self, model, max_seq_len=512, use_cache=True):
        super().__init__()
        self.model = model
        if not isinstance(model, nn.Embedding):
            if isinstance(model, Conv2dSubsampling):
                self.model = OnnxConv2dSubsampling(model)
                self.model.out[-1] = get_pos_emb(model.out[-1], max_seq_len)
            elif isinstance(model, Conv2dSubsampling2):
                self.model = OnnxConv2dSubsampling2(model)
                self.model.out[-1] = get_pos_emb(model.out[-1], max_seq_len)
            elif isinstance(model, Conv2dSubsampling6):
                self.model = OnnxConv2dSubsampling6(model)
                self.model.out[-1] = get_pos_emb(model.out[-1], max_seq_len)
            elif isinstance(model, Conv2dSubsampling8):
                self.model = OnnxConv2dSubsampling8(model)
                self.model.out[-1] = get_pos_emb(model.out[-1], max_seq_len)
            else:
                self.model[-1] = get_pos_emb(model[-1], max_seq_len)

    def forward(self, x, mask=None):
        if mask is None:
            return self.model(x)
        else:
            return self.model(x, mask)


def _pre_hook(
    state_dict,
    prefix,
    local_metadata,
    strict,
    missing_keys,
    unexpected_keys,
    error_msgs,
):
    """Perform pre-hook in load_state_dict for backward compatibility.

    Note:
        We saved self.pe until v.0.5.2 but we have omitted it later.
        Therefore, we remove the item "pe" from `state_dict` for backward compatibility.

    """
    k = prefix + "pe"
    if k in state_dict:
        state_dict.pop(k)


class OnnxPositionalEncoding(torch.nn.Module):
    """Positional encoding.

    Args:
        d_model (int): Embedding dimension.
        dropout_rate (float): Dropout rate.
        max_seq_len (int): Maximum input length.
        reverse (bool): Whether to reverse the input position. Only for
        the class LegacyRelPositionalEncoding. We remove it in the current
        class RelPositionalEncoding.
    """

    def __init__(self, model, max_seq_len=512, reverse=False, use_cache=True):
        """Construct an PositionalEncoding object."""
        super(OnnxPositionalEncoding, self).__init__()
        self.d_model = model.d_model
        self.reverse = reverse
        self.max_seq_len = max_seq_len
        self.xscale = math.sqrt(self.d_model)
        self._register_load_state_dict_pre_hook(_pre_hook)
        self.pe = model.pe
        self.use_cache = use_cache
        self.model = model
        if self.use_cache:
            self.extend_pe()
        else:
            self.div_term = torch.exp(
                torch.arange(0, self.d_model, 2, dtype=torch.float32)
                * -(math.log(10000.0) / self.d_model)
            )

    def extend_pe(self):
        """Reset the positional encodings."""
        pe_length = len(self.pe[0])
        if self.max_seq_len < pe_length:
            self.pe = self.pe[:, : self.max_seq_len]
        else:
            self.model.extend_pe(torch.tensor(0.0).expand(1, self.max_seq_len))
            self.pe = self.model.pe

    def _add_pe(self, x):
        """Computes positional encoding"""
        if self.reverse:
            position = torch.arange(
                x.size(1) - 1, -1, -1.0, dtype=torch.float32
            ).unsqueeze(1)
        else:
            position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)

        x = x * self.xscale
        x[:, :, 0::2] += torch.sin(position * self.div_term)
        x[:, :, 1::2] += torch.cos(position * self.div_term)
        return x

    def forward(self, x: torch.Tensor):
        """Add positional encoding.

        Args:
            x (torch.Tensor): Input tensor (batch, time, `*`).

        Returns:
            torch.Tensor: Encoded tensor (batch, time, `*`).
        """
        if self.use_cache:
            x = x * self.xscale + self.pe[:, : x.size(1)]
        else:
            x = self._add_pe(x)
        return x


class OnnxScaledPositionalEncoding(OnnxPositionalEncoding):
    """Scaled positional encoding module.

    See Sec. 3.2  https://arxiv.org/abs/1809.08895

    Args:
        d_model (int): Embedding dimension.
        dropout_rate (float): Dropout rate.
        max_seq_len (int): Maximum input length.

    """

    def __init__(self, model, max_seq_len=512, use_cache=True):
        """Initialize class."""
        super().__init__(model, max_seq_len, use_cache=use_cache)
        self.alpha = torch.nn.Parameter(torch.tensor(1.0))

    def reset_parameters(self):
        """Reset parameters."""
        self.alpha.data = torch.tensor(1.0)

    def _add_pe(self, x):
        """Computes positional encoding"""
        if self.reverse:
            position = torch.arange(
                x.size(1) - 1, -1, -1.0, dtype=torch.float32
            ).unsqueeze(1)
        else:
            position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)

        x = x * self.alpha
        x[:, :, 0::2] += torch.sin(position * self.div_term)
        x[:, :, 1::2] += torch.cos(position * self.div_term)
        return x

    def forward(self, x):
        """Add positional encoding.

        Args:
            x (torch.Tensor): Input tensor (batch, time, `*`).

        Returns:
            torch.Tensor: Encoded tensor (batch, time, `*`).

        """
        if self.use_cache:
            x = x + self.alpha * self.pe[:, : x.size(1)]
        else:
            x = self._add_pe(x)
        return x


class OnnxLegacyRelPositionalEncoding(OnnxPositionalEncoding):
    """Relative positional encoding module (old version).

    Details can be found in https://github.com/espnet/espnet/pull/2816.

    See : Appendix B in https://arxiv.org/abs/1901.02860

    Args:
        d_model (int): Embedding dimension.
        dropout_rate (float): Dropout rate.
        max_seq_len (int): Maximum input length.

    """

    def __init__(self, model, max_seq_len=512, use_cache=True):
        """Initialize class."""
        super().__init__(model, max_seq_len, reverse=True, use_cache=use_cache)

    def _get_pe(self, x):
        """Computes positional encoding"""
        if self.reverse:
            position = torch.arange(
                x.size(1) - 1, -1, -1.0, dtype=torch.float32
            ).unsqueeze(1)
        else:
            position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)

        pe = torch.zeros(x.shape)
        pe[:, :, 0::2] += torch.sin(position * self.div_term)
        pe[:, :, 1::2] += torch.cos(position * self.div_term)
        return pe

    def forward(self, x):
        """Compute positional encoding.

        Args:
            x (torch.Tensor): Input tensor (batch, time, `*`).

        Returns:
            torch.Tensor: Encoded tensor (batch, time, `*`).
            torch.Tensor: Positional embedding tensor (1, time, `*`).

        """
        x = x * self.xscale
        if self.use_cache:
            pos_emb = self.pe[:, : x.size(1)]
        else:
            pos_emb = self._get_pe(x)
        return x, pos_emb


class OnnxRelPositionalEncoding(torch.nn.Module):
    """Relative positional encoding module (new implementation).
    Details can be found in https://github.com/espnet/espnet/pull/2816.
    See : Appendix B in https://arxiv.org/abs/1901.02860
    Args:
        d_model (int): Embedding dimension.
        dropout_rate (float): Dropout rate.
        max_seq_len (int): Maximum input length.
    """

    def __init__(self, model, max_seq_len=512, use_cache=True):
        """Construct an PositionalEncoding object."""
        super(OnnxRelPositionalEncoding, self).__init__()
        self.d_model = model.d_model
        self.xscale = math.sqrt(self.d_model)
        self.pe = None
        self.use_cache = use_cache
        if self.use_cache:
            self.extend_pe(torch.tensor(0.0).expand(1, max_seq_len))
        else:
            self.div_term = torch.exp(
                torch.arange(0, self.d_model, 2, dtype=torch.float32)
                * -(math.log(10000.0) / self.d_model)
            )

    def extend_pe(self, x):
        """Reset the positional encodings."""
        if self.pe is not None and self.pe.size(1) >= x.size(1) * 2 - 1:
            # self.pe contains both positive and negative parts
            # the length of self.pe is 2 * input_len - 1
            if self.pe.dtype != x.dtype or self.pe.device != x.device:
                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
            return
        # Suppose `i` means to the position of query vecotr and `j` means the
        # position of key vector. We use position relative positions when keys
        # are to the left (i>j) and negative relative positions otherwise (i<j).
        pe_positive = torch.zeros(x.size(1), self.d_model)
        pe_negative = torch.zeros(x.size(1), self.d_model)
        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
        div_term = torch.exp(
            torch.arange(0, self.d_model, 2, dtype=torch.float32)
            * -(math.log(10000.0) / self.d_model)
        )
        pe_positive[:, 0::2] = torch.sin(position * div_term)
        pe_positive[:, 1::2] = torch.cos(position * div_term)
        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)

        # Reserve the order of positive indices and concat both positive and
        # negative indices. This is used to support the shifting trick
        # as in https://arxiv.org/abs/1901.02860
        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
        pe_negative = pe_negative[1:].unsqueeze(0)
        pe = torch.cat([pe_positive, pe_negative], dim=1)
        self.pe = pe.to(device=x.device, dtype=x.dtype)

    def _get_pe(self, x):
        pe_positive = torch.zeros(x.size(1), self.d_model)
        pe_negative = torch.zeros(x.size(1), self.d_model)
        theta = (
            torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1) * self.div_term
        )
        pe_positive[:, 0::2] = torch.sin(theta)
        pe_positive[:, 1::2] = torch.cos(theta)
        pe_negative[:, 0::2] = -1 * torch.sin(theta)
        pe_negative[:, 1::2] = torch.cos(theta)

        # Reserve the order of positive indices and concat both positive and
        # negative indices. This is used to support the shifting trick
        # as in https://arxiv.org/abs/1901.02860
        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
        pe_negative = pe_negative[1:].unsqueeze(0)
        return torch.cat([pe_positive, pe_negative], dim=1)

    def forward(self, x: torch.Tensor, use_cache=True):
        """Add positional encoding.
        Args:
            x (torch.Tensor): Input tensor (batch, time, `*`).
        Returns:
            torch.Tensor: Encoded tensor (batch, time, `*`).
        """
        x = x * self.xscale
        if self.use_cache:
            pos_emb = self.pe[
                :,
                self.pe.size(1) // 2 - x.size(1) + 1 : self.pe.size(1) // 2 + x.size(1),
            ]
        else:
            pos_emb = self._get_pe(x)
        return x, pos_emb


class OnnxStreamPositionalEncoding(torch.nn.Module):
    """Streaming Positional encoding."""

    def __init__(self, model, max_seq_len=5000, use_cache=True):
        """Construct an PositionalEncoding object."""
        super(StreamPositionalEncoding, self).__init__()
        self.use_cache = use_cache
        self.d_model = model.d_model
        self.xscale = model.xscale
        self.pe = model.pe
        self.use_cache = use_cache
        self.max_seq_len = max_seq_len
        if self.use_cache:
            self.extend_pe()
        else:
            self.div_term = torch.exp(
                torch.arange(0, self.d_model, 2, dtype=torch.float32)
                * -(math.log(10000.0) / self.d_model)
            )
        self._register_load_state_dict_pre_hook(_pre_hook)

    def extend_pe(self):
        """Reset the positional encodings."""
        pe_length = len(self.pe[0])
        if self.max_seq_len < pe_length:
            self.pe = self.pe[:, : self.max_seq_len]
        else:
            self.model.extend_pe(self.max_seq_len)
            self.pe = self.model.pe

    def _add_pe(self, x, start_idx):
        position = torch.arange(start_idx, x.size(1), dtype=torch.float32).unsqueeze(1)
        x = x * self.xscale
        x[:, :, 0::2] += torch.sin(position * self.div_term)
        x[:, :, 1::2] += torch.cos(position * self.div_term)
        return x

    def forward(self, x: torch.Tensor, start_idx: int = 0):
        """Add positional encoding.

        Args:
            x (torch.Tensor): Input tensor (batch, time, `*`).

        Returns:
            torch.Tensor: Encoded tensor (batch, time, `*`).

        """
        if self.use_cache:
            return x * self.xscale + self.pe[:, start_idx : start_idx + x.size(1)]
        else:
            return self._add_pe(x, start_idx)

 funasr/export/models/language_models/seq_rnn.py

New file
@@ -0,0 +1,84 @@
import os

import torch
import torch.nn as nn

class SequentialRNNLM(nn.Module):
    def __init__(self, model, **kwargs):
        super().__init__()
        self.encoder = model.encoder
        self.rnn = model.rnn
        self.rnn_type = model.rnn_type
        self.decoder = model.decoder
        self.nlayers = model.nlayers
        self.nhid = model.nhid
        self.model_name = "seq_rnnlm"

    def forward(self, y, hidden1, hidden2=None):
        # batch_score function.
        emb = self.encoder(y)
        if self.rnn_type == "LSTM":
            output, (hidden1, hidden2) = self.rnn(emb, (hidden1, hidden2))
        else:
            output, hidden1 = self.rnn(emb, hidden1)

        decoded = self.decoder(
            output.contiguous().view(output.size(0) * output.size(1), output.size(2))
        )
        if self.rnn_type == "LSTM":
            return (
                decoded.view(output.size(0), output.size(1), decoded.size(1)),
                hidden1,
                hidden2,
            )
        else:
            return (
                decoded.view(output.size(0), output.size(1), decoded.size(1)),
                hidden1,
            )

    def get_dummy_inputs(self):
        tgt = torch.LongTensor([0, 1]).unsqueeze(0)
        hidden = torch.randn(self.nlayers, 1, self.nhid)
        if self.rnn_type == "LSTM":
            return (tgt, hidden, hidden)
        else:
            return (tgt, hidden)

    def get_input_names(self):
        if self.rnn_type == "LSTM":
            return ["x", "in_hidden1", "in_hidden2"]
        else:
            return ["x", "in_hidden1"]

    def get_output_names(self):
        if self.rnn_type == "LSTM":
            return ["y", "out_hidden1", "out_hidden2"]
        else:
            return ["y", "out_hidden1"]

    def get_dynamic_axes(self):
        ret = {
            "x": {0: "x_batch", 1: "x_length"},
            "y": {0: "y_batch"},
            "in_hidden1": {1: "hidden1_batch"},
            "out_hidden1": {1: "out_hidden1_batch"},
        }
        if self.rnn_type == "LSTM":
            ret.update(
                {
                    "in_hidden2": {1: "hidden2_batch"},
                    "out_hidden2": {1: "out_hidden2_batch"},
                }
            )
        return ret

    def get_model_config(self, path):
        return {
            "use_lm": True,
            "model_path": os.path.join(path, f"{self.model_name}.onnx"),
            "lm_type": "SequentialRNNLM",
            "rnn_type": self.rnn_type,
            "nhid": self.nhid,
            "nlayers": self.nlayers,
        }

 funasr/export/models/language_models/subsampling.py

New file
@@ -0,0 +1,185 @@
"""Subsampling layer definition."""

import torch


class OnnxConv2dSubsampling(torch.nn.Module):
    """Convolutional 2D subsampling (to 1/4 length).

    Args:
        idim (int): Input dimension.
        odim (int): Output dimension.
        dropout_rate (float): Dropout rate.
        pos_enc (torch.nn.Module): Custom position encoding layer.

    """

    def __init__(self, model):
        """Construct an Conv2dSubsampling object."""
        super().__init__()
        self.conv = model.conv
        self.out = model.out

    def forward(self, x, x_mask):
        """Subsample x.

        Args:
            x (torch.Tensor): Input tensor (#batch, time, idim).
            x_mask (torch.Tensor): Input mask (#batch, 1, time).

        Returns:
            torch.Tensor: Subsampled tensor (#batch, time', odim),
                where time' = time // 4.
            torch.Tensor: Subsampled mask (#batch, 1, time'),
                where time' = time // 4.

        """
        x = x.unsqueeze(1)  # (b, c, t, f)
        x = self.conv(x)
        b, c, t, f = x.size()
        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
        if x_mask is None:
            return x, None
        return x, x_mask[:, :-2:2][:, :-2:2]

    def __getitem__(self, key):
        """Get item.

        When reset_parameters() is called, if use_scaled_pos_enc is used,
            return the positioning encoding.

        """
        if key != -1:
            raise NotImplementedError("Support only `-1` (for `reset_parameters`).")
        return self.out[key]


class OnnxConv2dSubsampling2(torch.nn.Module):
    """Convolutional 2D subsampling (to 1/2 length).

    Args:
        idim (int): Input dimension.
        odim (int): Output dimension.
        dropout_rate (float): Dropout rate.
        pos_enc (torch.nn.Module): Custom position encoding layer.

    """

    def __init__(self, model):
        """Construct an Conv2dSubsampling object."""
        super().__init__()
        self.conv = model.conv
        self.out = model.out

    def forward(self, x, x_mask):
        """Subsample x.

        Args:
            x (torch.Tensor): Input tensor (#batch, time, idim).
            x_mask (torch.Tensor): Input mask (#batch, 1, time).

        Returns:
            torch.Tensor: Subsampled tensor (#batch, time', odim),
                where time' = time // 2.
            torch.Tensor: Subsampled mask (#batch, 1, time'),
                where time' = time // 2.

        """
        x = x.unsqueeze(1)  # (b, c, t, f)
        x = self.conv(x)
        b, c, t, f = x.size()
        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
        if x_mask is None:
            return x, None
        return x, x_mask[:, :-2:2][:, :-2:1]

    def __getitem__(self, key):
        """Get item.

        When reset_parameters() is called, if use_scaled_pos_enc is used,
            return the positioning encoding.

        """
        if key != -1:
            raise NotImplementedError("Support only `-1` (for `reset_parameters`).")
        return self.out[key]


class OnnxConv2dSubsampling6(torch.nn.Module):
    """Convolutional 2D subsampling (to 1/6 length).

    Args:
        idim (int): Input dimension.
        odim (int): Output dimension.
        dropout_rate (float): Dropout rate.
        pos_enc (torch.nn.Module): Custom position encoding layer.

    """

    def __init__(self, model):
        """Construct an Conv2dSubsampling object."""
        super().__init__()
        self.conv = model.conv
        self.out = model.out

    def forward(self, x, x_mask):
        """Subsample x.

        Args:
            x (torch.Tensor): Input tensor (#batch, time, idim).
            x_mask (torch.Tensor): Input mask (#batch, 1, time).

        Returns:
            torch.Tensor: Subsampled tensor (#batch, time', odim),
                where time' = time // 6.
            torch.Tensor: Subsampled mask (#batch, 1, time'),
                where time' = time // 6.

        """
        x = x.unsqueeze(1)  # (b, c, t, f)
        x = self.conv(x)
        b, c, t, f = x.size()
        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
        if x_mask is None:
            return x, None
        return x, x_mask[:, :-2:2][:, :-4:3]


class OnnxConv2dSubsampling8(torch.nn.Module):
    """Convolutional 2D subsampling (to 1/8 length).

    Args:
        idim (int): Input dimension.
        odim (int): Output dimension.
        dropout_rate (float): Dropout rate.
        pos_enc (torch.nn.Module): Custom position encoding layer.

    """

    def __init__(self, model):
        """Construct an Conv2dSubsampling object."""
        super().__init__()
        self.conv = model.conv
        self.out = model.out

    def forward(self, x, x_mask):
        """Subsample x.

        Args:
            x (torch.Tensor): Input tensor (#batch, time, idim).
            x_mask (torch.Tensor): Input mask (#batch, 1, time).

        Returns:
            torch.Tensor: Subsampled tensor (#batch, time', odim),
                where time' = time // 8.
            torch.Tensor: Subsampled mask (#batch, 1, time'),
                where time' = time // 8.

        """
        x = x.unsqueeze(1)  # (b, c, t, f)
        x = self.conv(x)
        b, c, t, f = x.size()
        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
        if x_mask is None:
            return x, None
        return x, x_mask[:, :-2:2][:, :-2:2][:, :-2:2]

 funasr/export/models/language_models/transformer.py

New file
@@ -0,0 +1,110 @@
import os

import torch
import torch.nn as nn
from funasr.modules.vgg2l import import VGG2L
from funasr.modules.attention import MultiHeadedAttention
from funasr.modules.subsampling import (
    Conv2dSubsampling, Conv2dSubsampling6, Conv2dSubsampling8)

from funasr.export.models.modules.encoder_layer import EncoderLayerConformer as OnnxEncoderLayer
from funasr.export.models.language_models.embed import Embedding
from funasr.export.models.modules.multihead_att import OnnxMultiHeadedAttention

from funasr.export.utils.torch_function import MakePadMask

class TransformerLM(nn.Module, AbsExportModel):
    def __init__(self, model, max_seq_len=512, **kwargs):
        super().__init__()
        self.embed = Embedding(model.embed, max_seq_len)
        self.encoder = model.encoder
        self.decoder = model.decoder
        self.make_pad_mask = MakePadMask(max_seq_len, flip=False)
        # replace multihead attention module into customized module.
        for i, d in enumerate(self.encoder.encoders):
            # d is EncoderLayer
            if isinstance(d.self_attn, MultiHeadedAttention):
                d.self_attn = OnnxMultiHeadedAttention(d.self_attn)
            self.encoder.encoders[i] = OnnxEncoderLayer(d)

        self.model_name = "transformer_lm"
        self.num_heads = self.encoder.encoders[0].self_attn.h
        self.hidden_size = self.encoder.encoders[0].self_attn.linear_out.out_features

    def prepare_mask(self, mask):
        if len(mask.shape) == 2:
            mask = mask[:, None, None, :]
        elif len(mask.shape) == 3:
            mask = mask[:, None, :]
        mask = 1 - mask
        return mask * -10000.0

    def forward(self, y, cache):
        feats_length = torch.ones(y.shape).sum(dim=-1).type(torch.long)
        mask = self.make_pad_mask(feats_length)  # (B, T)
        mask = (y != 0) * mask

        xs = self.embed(y)
        # forward_one_step of Encoder
        if isinstance(
            self.encoder.embed,
            (Conv2dSubsampling, Conv2dSubsampling6, Conv2dSubsampling8, VGG2L),
        ):
            xs, mask = self.encoder.embed(xs, mask)
        else:
            xs = self.encoder.embed(xs)

        new_cache = []
        mask = self.prepare_mask(mask)
        for c, e in zip(cache, self.encoder.encoders):
            xs, mask = e(xs, mask, c)
            new_cache.append(xs)

        if self.encoder.normalize_before:
            xs = self.encoder.after_norm(xs)

        h = self.decoder(xs[:, -1])
        return h, new_cache

    def get_dummy_inputs(self):
        tgt = torch.LongTensor([1]).unsqueeze(0)
        cache = [
            torch.zeros((1, 1, self.encoder.encoders[0].size))
            for _ in range(len(self.encoder.encoders))
        ]
        return (tgt, cache)

    def is_optimizable(self):
        return True

    def get_input_names(self):
        return ["tgt"] + ["cache_%d" % i for i in range(len(self.encoder.encoders))]

    def get_output_names(self):
        return ["y"] + ["out_cache_%d" % i for i in range(len(self.encoder.encoders))]

    def get_dynamic_axes(self):
        ret = {"tgt": {0: "tgt_batch", 1: "tgt_length"}}
        ret.update(
            {
                "cache_%d" % d: {0: "cache_%d_batch" % d, 1: "cache_%d_length" % d}
                for d in range(len(self.encoder.encoders))
            }
        )
        ret.update(
            {
                "out_cache_%d"
                % d: {0: "out_cache_%d_batch" % d, 1: "out_cache_%d_length" % d}
                for d in range(len(self.encoder.encoders))
            }
        )
        return ret

    def get_model_config(self, path):
        return {
            "use_lm": True,
            "model_path": os.path.join(path, f"{self.model_name}.onnx"),
            "lm_type": "TransformerLM",
            "odim": self.encoder.encoders[0].size,
            "nlayers": len(self.encoder.encoders),
        }

 funasr/modules/vgg2l.py

New file
@@ -0,0 +1,92 @@
"""VGG2L module definition for custom encoder."""

from typing import Tuple, Union

import torch


class VGG2L(torch.nn.Module):
    """VGG2L module for custom encoder.

    Args:
        idim: Input dimension.
        odim: Output dimension.
        pos_enc: Positional encoding class.

    """

    def __init__(self, idim: int, odim: int, pos_enc: torch.nn.Module = None):
        """Construct a VGG2L object."""
        super().__init__()

        self.vgg2l = torch.nn.Sequential(
            torch.nn.Conv2d(1, 64, 3, stride=1, padding=1),
            torch.nn.ReLU(),
            torch.nn.Conv2d(64, 64, 3, stride=1, padding=1),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d((3, 2)),
            torch.nn.Conv2d(64, 128, 3, stride=1, padding=1),
            torch.nn.ReLU(),
            torch.nn.Conv2d(128, 128, 3, stride=1, padding=1),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d((2, 2)),
        )

        if pos_enc is not None:
            self.output = torch.nn.Sequential(
                torch.nn.Linear(128 * ((idim // 2) // 2), odim), pos_enc
            )
        else:
            self.output = torch.nn.Linear(128 * ((idim // 2) // 2), odim)

    def forward(
        self, feats: torch.Tensor, feats_mask: torch.Tensor
    ) -> Union[
        Tuple[torch.Tensor, torch.Tensor],
        Tuple[Tuple[torch.Tensor, torch.Tensor], torch.Tensor],
    ]:
        """Forward VGG2L bottleneck.

        Args:
            feats: Feature sequences. (B, F, D_feats)
            feats_mask: Mask of feature sequences. (B, 1, F)

        Returns:
            vgg_output: VGG output sequences.
                   (B, sub(F), D_out) or ((B, sub(F), D_out), (B, sub(F), D_att))
            vgg_mask: Mask of VGG output sequences. (B, 1, sub(F))

        """
        feats = feats.unsqueeze(1)
        vgg_output = self.vgg2l(feats)

        b, c, t, f = vgg_output.size()

        vgg_output = self.output(
            vgg_output.transpose(1, 2).contiguous().view(b, t, c * f)
        )

        if feats_mask is not None:
            vgg_mask = self.create_new_mask(feats_mask)
        else:
            vgg_mask = feats_mask

        return vgg_output, vgg_mask

    def create_new_mask(self, feats_mask: torch.Tensor) -> torch.Tensor:
        """Create a subsampled mask of feature sequences.

        Args:
            feats_mask: Mask of feature sequences. (B, 1, F)

        Returns:
            vgg_mask: Mask of VGG2L output sequences. (B, 1, sub(F))

        """
        vgg1_t_len = feats_mask.size(2) - (feats_mask.size(2) % 3)
        vgg_mask = feats_mask[:, :, :vgg1_t_len][:, :, ::3]

        vgg2_t_len = vgg_mask.size(2) - (vgg_mask.size(2) % 2)
        vgg_mask = vgg_mask[:, :, :vgg2_t_len][:, :, ::2]

        return vgg_mask

 funasr/tasks/asr.py

@@ -105,7 +105,7 @@
    name="specaug",
    classes=dict(
        specaug=SpecAug,
        specaug_lfr=SpecAugLFR,
        specaug_lfr=FSpecAugLR,
    ),
    type_check=AbsSpecAug,
    default=None,