pytorch_Filter_Response_Norm
usage
BN:(for one GPU)
python cifar.py --arch resnet --depth 20 --epoch 400 --train-batch 32
FRN:
python cifar.py --arch resnet_frn --depth 20 --epoch 400 --cos true --train-batch 32 --ramp-up 50
reproduce FRN on resnet_20/cifar_10
result
In paper, 32 images per gpu is the biggest batch size. So the reproduction matches paper's conclusion.
| images per gpu | 128 | 32 |
|---|---|---|
| batch_norm | 93.01% | 91.2% |
| FRN | 92.3% | 92.3% |