This command trains a ResNet-50 with GN on ImageNet:
./imagenet-resnet-gn.py --data /path/to/imagenet --depth 50
This script tries to follow the exact setting in the original paper.
Trained on 8 GPUs with a total batch size of 256, ResNet-50 in this script achieves 24.0, 24.0, 24.1 top-1 validation error in 3 independent runs, evaluated by the median of last 5 epochs. This matches the performance in the paper.
Training a ResNet-50 takes ~27 hours on 8 V100s.
Training a ResNet-101 with this script should reach 22.5~22.6 top-1 validation error.
Training logs for ResNet-50 & ResNet-101 are included.
The code is modified slightly from Tensorpack ResNet examples.
Trained ResNet-50 & ResNet-101 models are available at Tensorpack model zoo. They can be evaluated with:
./imagenet-resnet-gn.py --data /path/to/imagenet -d [50/101] --eval --load model.npz