This is a PyTorch implementation of [1611.06440 Pruning Convolutional Neural Networks for Resource Efficient Inference].
3x model size reducing can be achieved in the Kaggle's cat-vs-dog dataset with rather tiny accuracy loss.
This project was modified from https://github.com/jacobgil/pytorch-pruning and some main differences are:
- Pruning is done in one single pass rather than sequentially.
- FilterPruner and PruningFineTuner in the referred original project are merged to FilterPruner class for clarity and simplicity.
- InceptionV3, Inception_Resnet_V1, Resnet50 are pruned as examples and you can define a new FilterPruner for your custom model.
Note:
- You can use PyTorch's pretrained Resnet50 or InceptionV3 as a base model and prune them on the cat-vs-dog dataset mentioned before.
(See prune_InceptionV3_example.py and prune_Resnet50_example.py) - To prune a new model, you need to define a forward function and a prune_conv_layer function under FilterPruner according to your model's architecture.
TODO: Change pruner's forward to a more readable style.