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
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