peal.architectures.basic_modules

Small nn.Module building blocks used by PEAL’s configurable architectures.

The predictor and generator builders in peal.architectures assemble networks from layer lists; these modules wrap tensor operations (transpose, squeeze, mean, one-hot) so they can appear in nn.Sequential, plus a zennit-friendly residual block and a few attention layers with hand-rolled positional encodings for sequence and image inputs.

Classes

DimensionSwitchAttentionLayer	Cross-attention that maps a spatial feature map to output_size slots.
ImgSelfAttentionLayer	Self-attention over the spatial positions of an image feature map.
Mean	Average over a set of dimensions (global average pooling by default).
OneHotEncoding	One-hot encode integer inputs that arrive as a 2D tensor.
SelfAttentionLayer	Multi-head self-attention over a (B, L, C) sequence.
SkipConnection	Residual block module(x) + downsample(x) with a zennit Sum layer.
Squeeze	Remove singleton dimensions, one torch.squeeze per entry of dims.
Transpose	Module form of x.transpose(dim1, dim2).
Unsqueeze	Insert singleton dimensions, one torch.unsqueeze per entry of dims.