espnet2.asr.encoder.avhubert_encoder.downsample_basic_block

source

espnet2.asr.encoder.avhubert_encoder.downsample_basic_block(inplanes, outplanes, stride)

Construct a downsample block for a neural network.

This function creates a sequential block consisting of a 1x1 convolution followed by batch normalization. It is typically used in architectures that require downsampling of feature maps, such as ResNet variants.

• Parameters:
  • inplanes (int) – Number of input channels.
  • outplanes (int) – Number of output channels.
  • stride (int) – The stride of the convolution.
• Returns: A sequential block containing a convolution and batch normalization.
• Return type: nn.Sequential

Examples

>>> downsample_block = downsample_basic_block(64, 128, stride=2)
>>> print(downsample_block)
Sequential(
  (0): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)
  (1): BatchNorm2d(128)
)