Replicating the experiments of "The lottery ticket hypothesis"

This repository contains the implementation to replicate the experiments "Lottery Ticket Hypothesis". The detail explanation of the concepts are described in this site

References

THE LOTTERY TICKET HYPOTHESIS: FINDING SPARSE, TRAINABLE NEURAL NETWORK

Requirements

pip install -r requirements.txt

Experiments

Dataset

Cifar-10

Network

VGG-based 6 Convolutional Netowrok and 2 Full Connected Network.The trained model reached 84.82% accuracy on testing set. Detail parameter configs and performance described in [to be linked!].

Accuracy of initial weighted pruned model.

11.60% accuracy with original data, on the other hand, a pruned accuracy reached 19.10%. However it is found that even not so much rate of pruning can decrease accuracy. Detail result described in [to be linked!].

Retrained result

T.B.C

How to use

Check performance of original network

In root directory of this repository,

python experiments\original_perfoemance\original_performance.py

Check the accuracy of initial weighted model with different pruning rate

In root directory of this repository,

python experiments\original_perfoemance\original_performance.py

Check the accuracy of retrained model with different pruning rate

In root directory of this repository,

python experiments\original_perfoemance\retrain.py