aishell example

游雁

2024-02-19 94de39dde2e616a01683c518023d0fab72b4e103

 funasr/models/paraformer/model.py

@@ -1,35 +1,30 @@
import os
import logging
from typing import Union, Dict, List, Tuple, Optional

import torch
import torch.nn as nn
#!/usr/bin/env python3
# -*- encoding: utf-8 -*-
# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
#  MIT License  (https://opensource.org/licenses/MIT)

import time

from funasr.losses.label_smoothing_loss import (
    LabelSmoothingLoss,  # noqa: H301
)

from funasr.models.paraformer.cif_predictor import mae_loss

from funasr.models.transformer.utils.add_sos_eos import add_sos_eos
from funasr.models.transformer.utils.nets_utils import make_pad_mask, pad_list
from funasr.metrics.compute_acc import th_accuracy
from funasr.train_utils.device_funcs import force_gatherable

from funasr.models.paraformer.search import Hypothesis

import torch
import logging
from torch.cuda.amp import autocast
from typing import Union, Dict, List, Tuple, Optional

from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank
from funasr.utils import postprocess_utils
from funasr.utils.datadir_writer import DatadirWriter
from funasr.register import tables
from funasr.models.ctc.ctc import CTC
from funasr.utils import postprocess_utils
from funasr.metrics.compute_acc import th_accuracy
from funasr.utils.datadir_writer import DatadirWriter
from funasr.models.paraformer.search import Hypothesis
from funasr.models.paraformer.cif_predictor import mae_loss
from funasr.train_utils.device_funcs import force_gatherable
from funasr.losses.label_smoothing_loss import LabelSmoothingLoss
from funasr.models.transformer.utils.add_sos_eos import add_sos_eos
from funasr.models.transformer.utils.nets_utils import make_pad_mask, pad_list
from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank


@tables.register("model_classes", "Paraformer")
class Paraformer(nn.Module):
class Paraformer(torch.nn.Module):
    """
    Author: Speech Lab of DAMO Academy, Alibaba Group
    Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition
@@ -38,7 +33,6 @@
    
    def __init__(
        self,
        # token_list: Union[Tuple[str, ...], List[str]],
        specaug: Optional[str] = None,
        specaug_conf: Optional[Dict] = None,
        normalize: str = None,
@@ -79,17 +73,17 @@
        super().__init__()

        if specaug is not None:
            specaug_class = tables.specaug_classes.get(specaug.lower())
            specaug_class = tables.specaug_classes.get(specaug)
            specaug = specaug_class(**specaug_conf)
        if normalize is not None:
            normalize_class = tables.normalize_classes.get(normalize.lower())
            normalize_class = tables.normalize_classes.get(normalize)
            normalize = normalize_class(**normalize_conf)
        encoder_class = tables.encoder_classes.get(encoder.lower())
        encoder_class = tables.encoder_classes.get(encoder)
        encoder = encoder_class(input_size=input_size, **encoder_conf)
        encoder_output_size = encoder.output_size()

        if decoder is not None:
            decoder_class = tables.decoder_classes.get(decoder.lower())
            decoder_class = tables.decoder_classes.get(decoder)
            decoder = decoder_class(
                vocab_size=vocab_size,
                encoder_output_size=encoder_output_size,
@@ -104,7 +98,7 @@
                odim=vocab_size, encoder_output_size=encoder_output_size, **ctc_conf
            )
        if predictor is not None:
            predictor_class = tables.predictor_classes.get(predictor.lower())
            predictor_class = tables.predictor_classes.get(predictor)
            predictor = predictor_class(**predictor_conf)
        
        # note that eos is the same as sos (equivalent ID)
@@ -169,6 +163,7 @@
        self.use_1st_decoder_loss = use_1st_decoder_loss
        self.length_normalized_loss = length_normalized_loss
        self.beam_search = None
        self.error_calculator = None
    
    def forward(
        self,
@@ -439,7 +434,7 @@
        #         scorer.to(device=kwargs.get("device", "cpu"), dtype=getattr(torch, kwargs.get("dtype", "float32"))).eval()
        self.beam_search = beam_search
        
    def generate(self,
    def inference(self,
             data_in,
             data_lengths=None,
             key: list=None,
@@ -456,7 +451,7 @@
            self.nbest = kwargs.get("nbest", 1)
        
        meta_data = {}
        if isinstance(data_in, torch.Tensor): # fbank
        if isinstance(data_in, torch.Tensor) and kwargs.get("data_type", "sound") == "fbank": # fbank
            speech, speech_lengths = data_in, data_lengths
            if len(speech.shape) < 3:
                speech = speech[None, :, :]
@@ -496,6 +491,8 @@
        b, n, d = decoder_out.size()
        if isinstance(key[0], (list, tuple)):
            key = key[0]
        if len(key) < b:
            key = key*b
        for i in range(b):
            x = encoder_out[i, :encoder_out_lens[i], :]
            am_scores = decoder_out[i, :pre_token_length[i], :]
@@ -533,9 +530,9 @@
                if tokenizer is not None:
                    # Change integer-ids to tokens
                    token = tokenizer.ids2tokens(token_int)
                    text = tokenizer.tokens2text(token)
                    
                    text_postprocessed, _ = postprocess_utils.sentence_postprocess(token)
                    text_postprocessed = tokenizer.tokens2text(token)
                    if not hasattr(tokenizer, "bpemodel"):
                        text_postprocessed, _ = postprocess_utils.sentence_postprocess(token)
                    
                    result_i = {"key": key[i], "text": text_postprocessed}