Architecture Overview¶
This diagram shows how ContextualConv2d works internally.
ContextualConv2d Architecture¶
Each convolution group handles a slice of the input and combines it with optional global context parameters (scale and/or bias).
Back to top
digraph ContextualConv2d {
rankdir=TB;
Input -> Unfold [label="im2col"];
Unfold -> SliceGroups [label="split by group"];
SliceGroups -> MatMul [label="grouped weights"];
MatMul -> AddContext [label="(optional) + scale/bias"];
AddContext -> BiasAdd [label="FiLM: γ * conv(x) + β"];
BiasAdd -> Reshape [label="(N, C_out, H, W)"];
}