export model

游雁

2023-02-07 59f184a622be316b6a75ce053ee8e19e6a7b50ec

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
import logging
 
 
import torch
import torch.nn as nn
 
from funasr.export.utils.torch_function import MakePadMask
from funasr.train.abs_espnet_model import AbsESPnetModel
from funasr.models.encoder.sanm_encoder import SANMEncoder
from funasr.export.models.encoder.sanm_encoder import SANMEncoder as SANMEncoder_export
from funasr.models.predictor.cif import CifPredictorV2
from funasr.export.models.predictor.cif import CifPredictorV2 as CifPredictorV2_export
from funasr.models.decoder.sanm_decoder import ParaformerSANMDecoder
from funasr.export.models.decoder.sanm_decoder import ParaformerSANMDecoder as ParaformerSANMDecoder_export
 
class Paraformer(nn.Module):
    """
    Author: Speech Lab, Alibaba Group, China
    Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition
    https://arxiv.org/abs/2206.08317
    """
 
    def __init__(
            self,
            model,
            max_seq_len=512,
            feats_dim=560,
            model_name='model',
            **kwargs,
    ):
        super().__init__()
        if isinstance(model.encoder, SANMEncoder):
            self.encoder = SANMEncoder_export(model.encoder)
        if isinstance(model.predictor, CifPredictorV2):
            self.predictor = CifPredictorV2_export(model.predictor)
        if isinstance(model.decoder, ParaformerSANMDecoder):
            self.decoder = ParaformerSANMDecoder_export(model.decoder)
        self.make_pad_mask = MakePadMask(max_seq_len, flip=False)
        self.feats_dim = feats_dim
        self.model_name = model_name
        self.onnx = False
        if "onnx" in kwargs:
            self.onnx = kwargs["onnx"]
    
    def forward(
            self,
            speech: torch.Tensor,
            speech_lengths: torch.Tensor,
    ):
        # a. To device
        batch = {"speech": speech, "speech_lengths": speech_lengths}
        # batch = to_device(batch, device=self.device)
    
        enc, enc_len = self.encoder(**batch)
        mask = self.make_pad_mask(enc_len)[:, None, :]
        pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index = self.predictor(enc, mask)
        pre_token_length = pre_token_length.round().long()
 
        decoder_out, _ = self.decoder(enc, enc_len, pre_acoustic_embeds, pre_token_length)
        decoder_out = torch.log_softmax(decoder_out, dim=-1)
 
        return decoder_out, pre_token_length
    
    # def get_output_size(self):
    #     return self.model.encoders[0].size
 
    def get_dummy_inputs(self):
        speech = torch.randn(2, 30, self.feats_dim)
        speech_lengths = torch.tensor([6, 30]).long()
        return (speech, speech_lengths)
 
    def get_input_names(self):
        return ['speech', 'speech_lengths']
 
    def get_output_names(self):
        return ['logits', 'token_num']
 
    def get_dynamic_axes(self):
        return {
            'speech': {
                0: 'batch_size',
                1: 'feats_length'
            },
            'speech_lengths': {
                0: 'batch_size',
            },
            'logits': {
                0: 'batch_size',
                1: 'logits_length'
            },
        }