Create smaller model from pruning mask


We know that pytorch pruning does not actually speed up inference.

I plan to do the following:

  • train a faster rcnn resnet18 on my dataset
  • apply group pruning (from the pruning module) on the backbone, in order to select kernel filters to be pruned
  • iterate on layers, and create a new backbone containing only non-pruned filters, and copy weights
  • retrain on the smaller model

In this way, I hope that I can speed up inference.
Is my approach reasonnable ?

Thank you for your help :slight_smile: