Code release for REPAIR: REnormalizing Permuted Activations for Interpolation Repair.
- The notebook Train-Merge-REPAIR-VGG11 is a self-contained, PyTorch-only script which trains two standard VGG11 networks on CIFAR-10. It then merges them in weight-space using neuron alignment. Finally, it applies REPAIR to reduce the merging penalty from ~20% to ~5%. The width of the VGG network can be scaled up to obtain zero-barrier connectivity.
- The notebook Merge-ResNet50-ImageNet applies REPAIR to merging ResNet50s trained on ImageNet. It is less self-contained, depending on FFCV ImageNet files, and two pretrained ResNet50 checkpoints.
These two notebooks contain complete implementations of REPAIR from which further experiments can be derived. The remainder of the notebooks in ./notebooks contain assorted such experiments.