peal.generators.deeplift_resnet¶

ResNet variants with one nn.ReLU instance per activation site.

Captum’s DeepLift hooks every ReLU module and requires each to be used exactly once per forward pass; torchvision’s ResNet reuses one relu inside every block, which breaks the attribution. These classes mirror torchvision’s BasicBlock / Bottleneck / ResNet with distinct non-inplace ReLUs so that DeepLift can attribute through them. They are used by the Stable Diffusion 3 generator when it rebuilds a predictor for attribution-guided sampling (DeepLiftResNet18 … 152).

Functions

`DeepLiftResNet101`	DeepLift-compatible ResNet-101 (bottlenecks `[3, 4, 23, 3]`).
`DeepLiftResNet152`	DeepLift-compatible ResNet-152 (bottlenecks `[3, 8, 36, 3]`).
`DeepLiftResNet18`	DeepLift-compatible ResNet-18 (basic blocks `[2, 2, 2, 2]`).
`DeepLiftResNet34`	DeepLift-compatible ResNet-34 (basic blocks `[3, 4, 6, 3]`).
`DeepLiftResNet50`	DeepLift-compatible ResNet-50 (bottlenecks `[3, 4, 6, 3]`).

Classes

`DeepLiftBasicBlock`	BasicBlock with separate ReLU instances for DeepLift compatibility
`DeepLiftBottleneck`	Bottleneck with separate ReLU instances for DeepLift compatibility
`DeepLiftResNet`	Base ResNet class with unique ReLU instances for DeepLift compatibility