peal.architectures.module_blocks

Building blocks from which PEAL’s configurable predictors are assembled.

Each block is an nn.Sequential parameterised by a pydantic layer config (FCConfig, VGGConfig, ResnetConfig) so that classifier architectures can be described entirely in yaml. create_cnn_layer stacks num_blocks VGG or ResNet blocks into one stage whose first block downsamples by stride 2.

Functions

create_cnn_layer

The create_cnn_layer function creates a CNN layer.

Classes

FCBlock	Fully connected block: a 1x1 projection, optional activation and dropout.
ResnetBlock	Basic residual block with two 3x3 convolutions and a skip connection.
TransformerBlock	Pre-activation transformer block: attention and position-wise MLP, each residual.
VGGBlock	VGG-style block: convolution, optional batchnorm, activation.