Kesi Save

Knowledge distillation from Ensembles of Iterative pruning (BMVC 2020)

Project README

Gift from Iterative Network Pruning

This repository contains the code for reproducing the results for paper Paying more attention to Snapshot of Iterative Pruning: Improving Model Compression via Ensemble Distillation (BMVC 2020).

Summary

In short, the paper propose to leverage the snapshots of iterative pruning to construct ensembles and distilling knowledge from them. To stimulate the diversity between each snapshots, we use One-cycle schedule to retrain the pruned networks. Thus, each snapshot is encouraged to converge to different optimal solution.

The algorithm is summarized below:

  1. Train the baseline network to completion.
  2. Prune redundant weights (based on some criteria).
  3. Retrain with One-cycle learning rate.
  4. Repeat step 2 and 3 until desired compression ratio is reached.
  5. Distill knowledge from ensemble to desired network.

How to run

Please checkout example.pynb for detail instruction to reproduce the results on CIFAR. Instruction for running experiments on Tiny-Imagenet might be updated later.

We also provided the scripts for repeative pruning and knowledge distillation (read Sec.5 in Colab example). Disclamer: you might have to modify the checkpoint_paths variable in ensemble_finetune.py to appropriate paths (and by that I mean cifar/filter_pruning/ensemble_finetune.py, cifar/weight_pruning/ensemble_finetune.py,... depending on your chosen method/dataset).

Results

CIFAR-10 and CIFAR-100.

Results on CIFAR

Tiny-Imagenet

Results on Tiny-Imagenet

PFEC and MWP stand for Pruning Filters for Efficient ConvNets and Learning both Weights and Connections for Efficient Neural Networks respectively.

Acknowledgement

The code is mostly taken from Eric-mingjie's repository

Open Source Agenda is not affiliated with "Kesi" Project. README Source: lehduong/kesi

Open Source Agenda Badge

Open Source Agenda Rating