This is an simple implementation of the CVPR2019 paper Local to Global Learning for Deep Neural Networks. Current version is based on CIFAR100. The other experiments are coming soon.

##Dependency

python3, torch >= 0.3

Steps

Download CIFAR100 and transform the data to .npy form:

python cifar100_download.py

Run group number=1 to test the baseline (VGG-16 on CIFAR100):

python lgl.py --group-num=1

Run the code with group number 3 and random selection strategy to test lgl method:

python lgl.py --group-num=3

Normally, the accuracy rises as we increase group number. Run

python lgl.py --group-num=20