DeformableModulationNormalization

Enums · AiDotNet.Enums

How a deformable convolution normalizes its modulation scalars.

How It Works

DCNv2 (Zhu et al., “Deformable ConvNets v2”, CVPR 2019) passes each sample point’s modulation scalar through a sigmoid independently, giving every point a weight in (0, 1) with no relationship between them.

DCNv3 (Wang et al., “InternImage: Exploring Large-Scale Vision Foundation Models with Deformable Convolutions”, CVPR 2023) changes this deliberately: the scalars are normalized with a softmax ACROSS the K sample points, so they sum to one at every location. The paper gives the reason directly – an unbounded sum of modulation weights makes gradients and feature magnitudes grow with model scale, and normalizing them stabilizes training for large models. Using sigmoid under a DCNv3 configuration is therefore not a small deviation; it removes one of the three changes that define DCNv3.

Fields

Field Summary
DeformableModulationNormalization Auto Choose from the layer’s configuration: softmax when it is set up as DCNv3 (a separable offset projection), sigmoid otherwise.
DeformableModulationNormalization Sigmoid Per-point sigmoid, as in DCNv2.
DeformableModulationNormalization Softmax Softmax across the K sample points, as in DCNv3.
int value__