提交

kongdeqiang

2026-03-13 28ccfbfc51068a663a80764e14074df5edf2b5ba

 funasr/models/sond/encoder/conv_encoder.py

@@ -17,24 +17,24 @@

class EncoderLayer(nn.Module):
    def __init__(
            self,
            input_units,
            num_units,
            kernel_size=3,
            activation="tanh",
            stride=1,
            include_batch_norm=False,
            residual=False
        self,
        input_units,
        num_units,
        kernel_size=3,
        activation="tanh",
        stride=1,
        include_batch_norm=False,
        residual=False,
    ):
        super().__init__()
        left_padding = math.ceil((kernel_size - stride) / 2)
        right_padding = kernel_size - stride - left_padding
        self.conv_padding = nn.ConstantPad1d((left_padding, right_padding), 0.0)
        self.conv1d = nn.Conv1d(
                input_units,
                num_units,
                kernel_size,
                stride,
            input_units,
            num_units,
            kernel_size,
            stride,
        )
        self.activation = self.get_activation(activation)
        if include_batch_norm:
@@ -71,23 +71,23 @@
    """

    def __init__(
            self,
            num_layers,
            input_units,
            num_units,
            kernel_size=3,
            dropout_rate=0.3,
            position_encoder=None,
            activation='tanh',
            auxiliary_states=True,
            out_units=None,
            out_norm=False,
            out_residual=False,
            include_batchnorm=False,
            regularization_weight=0.0,
            stride=1,
            tf2torch_tensor_name_prefix_torch: str = "speaker_encoder",
            tf2torch_tensor_name_prefix_tf: str = "EAND/speaker_encoder",
        self,
        num_layers,
        input_units,
        num_units,
        kernel_size=3,
        dropout_rate=0.3,
        position_encoder=None,
        activation="tanh",
        auxiliary_states=True,
        out_units=None,
        out_norm=False,
        out_residual=False,
        include_batchnorm=False,
        regularization_weight=0.0,
        stride=1,
        tf2torch_tensor_name_prefix_torch: str = "speaker_encoder",
        tf2torch_tensor_name_prefix_tf: str = "EAND/speaker_encoder",
    ):
        super().__init__()
        self._output_size = num_units
@@ -125,8 +125,8 @@
                activation,
                self.stride[lnum],
                include_batchnorm,
                residual=True if lnum > 0 else False
            )
                residual=True if lnum > 0 else False,
            ),
        )

        if self.out_units is not None:
@@ -137,7 +137,7 @@
                num_units,
                out_units,
                kernel_size,
        )
            )

        if self.out_norm:
            self.after_norm = LayerNorm(out_units)
@@ -172,4 +172,3 @@
            outputs = outputs + inputs

        return outputs, ilens, None