This is a wrapper repo that shows how to use the autoprune library which uses PyTorch JIT Tracing functionality to extract an execution graph of a network to then automatically detect dependencies between layers in order to prune dependent layers correctly. As it uses torch.jit.trace functionality, the models passed in must be written to be traceable.

Most things you want to do can be done without editing the autoprune library itself, and so before running anything here run the command git submodule update --init to point to the latest version of the library.

Currently, the models (connectivity) type supported by autoprune are: - No special connectivity : AlexNet, VGG, etc. - Residual connectivity : ResNets etc. - Depthwise convolution connectivity : MobileNets etc. - Concatenataion connectivity : SqueezeNets, GoogleNets etc.

The file basic_usage.py shows how to use autoprune without any modifications for the supported l1-norm pruning metric.

Eg: To l1-norm prune resnet_18 by 20%, run

python basic_usage.py --pl 0.2 --net resnet_18 --metric l1-norm

The file taylor_pruning.py shows how to add custom functionality to the library without modifying the library itself to implement the more complex Taylor-first-order pruning method.

Eg: To taylor prune resnet_50 by 20%, run

python basic_usage.py --pl 0.2 --net gated_resnet_50

If you require to make even more detailed changes, you can change the library itself as this is also open-sourced.