| | |
|---|
| | | StreamingRelPositionalEncoding, |
|---|
| | | ) |
|---|
| | | from funasr.modules.layer_norm import LayerNorm |
|---|
| | | from funasr.modules.normalization import get_normalization |
|---|
| | | from funasr.modules.multi_layer_conv import Conv1dLinear |
|---|
| | | from funasr.modules.multi_layer_conv import MultiLayeredConv1d |
|---|
| | | from funasr.modules.nets_utils import get_activation |
|---|
| | |
|---|
| | | feed_forward: torch.nn.Module, |
|---|
| | | feed_forward_macaron: torch.nn.Module, |
|---|
| | | conv_mod: torch.nn.Module, |
|---|
| | | norm_class: torch.nn.Module = torch.nn.LayerNorm, |
|---|
| | | norm_class: torch.nn.Module = LayerNorm, |
|---|
| | | norm_args: Dict = {}, |
|---|
| | | dropout_rate: float = 0.0, |
|---|
| | | ) -> None: |
|---|
| | |
|---|
| | | |
|---|
| | | return x, cache |
|---|
| | | |
|---|
| | | class ConformerChunkEncoder(torch.nn.Module): |
|---|
| | | class ConformerChunkEncoder(AbsEncoder): |
|---|
| | | """Encoder module definition. |
|---|
| | | Args: |
|---|
| | | input_size: Input size. |
|---|
| | |
|---|
| | | default_chunk_size: int = 16, |
|---|
| | | jitter_range: int = 4, |
|---|
| | | subsampling_factor: int = 1, |
|---|
| | | **activation_parameters, |
|---|
| | | ) -> None: |
|---|
| | | """Construct an Encoder object.""" |
|---|
| | | super().__init__() |
|---|
| | |
|---|
| | | ) |
|---|
| | | |
|---|
| | | activation = get_activation( |
|---|
| | | activation_type, **activation_parameters |
|---|
| | | activation_type |
|---|
| | | ) |
|---|
| | | |
|---|
| | | pos_wise_args = ( |
|---|
| | |
|---|
| | | simplified_att_score, |
|---|
| | | ) |
|---|
| | | |
|---|
| | | norm_class, norm_args = get_normalization( |
|---|
| | | norm_type, |
|---|
| | | ) |
|---|
| | | |
|---|
| | | fn_modules = [] |
|---|
| | | for _ in range(num_blocks): |
|---|
| | |
|---|
| | | PositionwiseFeedForward(*pos_wise_args), |
|---|
| | | PositionwiseFeedForward(*pos_wise_args), |
|---|
| | | CausalConvolution(*conv_mod_args), |
|---|
| | | norm_class=norm_class, |
|---|
| | | norm_args=norm_args, |
|---|
| | | dropout_rate=dropout_rate, |
|---|
| | | ) |
|---|
| | | fn_modules.append(module) |
|---|
| | |
|---|
| | | self.encoders = MultiBlocks( |
|---|
| | | [fn() for fn in fn_modules], |
|---|
| | | output_size, |
|---|
| | | norm_class=norm_class, |
|---|
| | | norm_args=norm_args, |
|---|
| | | ) |
|---|
| | | |
|---|
| | | self.output_size = output_size |
|---|
| | | self._output_size = output_size |
|---|
| | | |
|---|
| | | self.dynamic_chunk_training = dynamic_chunk_training |
|---|
| | | self.short_chunk_threshold = short_chunk_threshold |
|---|
| | |
|---|
| | | self.jitter_range = jitter_range |
|---|
| | | |
|---|
| | | self.time_reduction_factor = time_reduction_factor |
|---|
| | | |
|---|
| | | def output_size(self) -> int: |
|---|
| | | return self._output_size |
|---|
| | | |
|---|
| | | def get_encoder_input_raw_size(self, size: int, hop_length: int) -> int: |
|---|
| | | """Return the corresponding number of sample for a given chunk size, in frames. |
|---|
| | |
|---|
| | | x = x[:,::self.time_reduction_factor,:] |
|---|
| | | olens = torch.floor_divide(olens-1, self.time_reduction_factor) + 1 |
|---|
| | | |
|---|
| | | return x, olens |
|---|
| | | return x, olens, None |
|---|
| | | |
|---|
| | | def simu_chunk_forward( |
|---|
| | | self, |
|---|
