Repository files navigation

Paper:https://arxiv.org/abs/2102.06171.pdf

Original code:https://github.com/deepmind/deepmind-research/tree/master/nfnets

I recommend using Docker to run the code:

docker build -t nfnets/imagenet:latest --build-arg USER_ID=$(id -u) --build-arg GROUP_ID=$(id -g) .

To train NFNets on imagenet dataset:

docker run --rm -it --gpus all -v $(pwd):/tf -p 8889:8888 -p 6006:6006 nfnets/imagenet:latest python train.py --variant F0 --batch_size 4096 --num_epochs 360

Please see thetrain.py module to get more arguments.

Pre-trained weights have been converted to be compatible with my models' implementation. You can download them fromhere

To evaluate NFNets on test set of imagenet dataset:

docker run --rm -it --gpus all -v $(pwd):/tf -p 8889:8888 -p 6006:6006 nfnets/imagenet:latest python evaluate_imagenet.py --variant F0 --batch_size 50

You can also check the notebook in the repo showing how to run an NFNet to classify an image.

• WSConv2d
• Clipping Gradient module
• Documentation
• NFNets
• NF-ResNets
• Update pretrained weights
• How to find-tune

To cite the original paper, use:

@article{brock2021high,  author={Andrew Brock and Soham De and Samuel L. Smith and Karen Simonyan},  title={High-Performance Large-Scale Image Recognition Without Normalization},  journal={arXiv preprint arXiv:},  year={2021}}