| | |
|---|
| | | import torch.nn as nn |
|---|
| | | |
|---|
| | | from funasr.export.utils.torch_function import MakePadMask |
|---|
| | | from funasr.train.abs_espnet_model import AbsESPnetModel |
|---|
| | | from funasr.export.utils.torch_function import sequence_mask |
|---|
| | | 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 |
|---|
| | |
|---|
| | | **kwargs, |
|---|
| | | ): |
|---|
| | | super().__init__() |
|---|
| | | onnx = False |
|---|
| | | if "onnx" in kwargs: |
|---|
| | | onnx = kwargs["onnx"] |
|---|
| | | if isinstance(model.encoder, SANMEncoder): |
|---|
| | | self.encoder = SANMEncoder_export(model.encoder) |
|---|
| | | self.encoder = SANMEncoder_export(model.encoder, onnx=onnx) |
|---|
| | | 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.decoder = ParaformerSANMDecoder_export(model.decoder, onnx=onnx) |
|---|
| | | |
|---|
| | | self.feats_dim = feats_dim |
|---|
| | | self.model_name = model_name |
|---|
| | | self.onnx = False |
|---|
| | | if "onnx" in kwargs: |
|---|
| | | self.onnx = kwargs["onnx"] |
|---|
| | | |
|---|
| | | |
|---|
| | | 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 forward( |
|---|
| | | self, |
|---|
| | | speech: torch.Tensor, |
|---|
| | |
|---|
| | | 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() |
|---|
| | | pre_token_length = pre_token_length.round().type(torch.int32) |
|---|
| | | |
|---|
| | | decoder_out, _ = self.decoder(enc, enc_len, pre_acoustic_embeds, pre_token_length) |
|---|
| | | decoder_out = torch.log_softmax(decoder_out, dim=-1) |
|---|
| | | # sample_ids = decoder_out.argmax(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() |
|---|
| | | speech_lengths = torch.tensor([6, 30], dtype=torch.int32) |
|---|
| | | return (speech, speech_lengths) |
|---|
| | | |
|---|
| | | def get_dummy_inputs_txt(self, txt_file: str = "/mnt/workspace/data_fbank/0207/12345.wav.fea.txt"): |
|---|
| | | import numpy as np |
|---|
| | | fbank = np.loadtxt(txt_file) |
|---|
| | | fbank_lengths = np.array([fbank.shape[0], ], dtype=np.int32) |
|---|
| | | speech = torch.from_numpy(fbank[None, :, :].astype(np.float32)) |
|---|
| | | speech_lengths = torch.from_numpy(fbank_lengths.astype(np.int32)) |
|---|
| | | return (speech, speech_lengths) |
|---|
| | | |
|---|
| | | def get_input_names(self): |
|---|
