peal.architectures.module_blocks.ResnetBlock

class peal.architectures.module_blocks.ResnetBlock

Bases: Sequential

Basic residual block with two 3x3 convolutions and a skip connection.

The residual branch is conv-(bn)-act-conv-(bn). When stride > 1 the identity path is replaced by AvgPool2d(2) followed by a 1x1 convolution so that shapes match; the sum is followed by a final activation.

Parameters:

• input_channels (int) – Number of input channels.

• activation (type of nn.Module) – Activation class, instantiated without arguments.

• stride (int) – Stride of the first convolution (2 downsamples).

• conv (type of nn.Module) – Convolution class (nn.Conv1d/Conv2d/Conv3d).

• batchnorm (type of nn.Module) – Batchnorm class matching conv; used if config.use_batchnorm.

• config (ResnetConfig) – Provides num_neurons (output channels) and use_batchnorm.

Notes

The downsampling path always uses nn.AvgPool2d regardless of tensor_dim.

Methods

__init__(input_channels, activation, stride, conv, batchnorm, config)

The __init__ method initializes the ResnetBlock class. :param input_channels: The number of input channels. :param activation: The activation function class. :param stride: The stride of the convolution. :param conv: The convolution class. :param batchnorm: The batchnorm class. :param config: The config of the block.

Parameters:

• input_channels (pydantic.types.PositiveInt)

• activation (Type[torch.nn.Module])

• stride (pydantic.types.PositiveInt)

• conv (Type[torch.nn.Module])

• batchnorm (Type[torch.nn.Module])

• config (ResnetConfig)