weightnet.pytorch

The unoffical PyTorch implementation of WeightNet, based on OpenMMLab mmclassification.

Installation

See install.md and getting_started.md. Learn more at mmcls's documentation.

Experiments

Model	#Params.	FLOPs	Top-1 err. (offical)	Top-1 err. (ours)
ShuffleNetV2 (1×)	2.2M	138M	30.9	30.5
+ WeightNet (1×)	2.4M	139M	28.8	29.1

bash tools/dist_train.sh configs/shufflenet_v2/shufflenet_1x_b256x4_imagenet.py 4
# ckpts/shufflenet_1x_b256x4_imagenet
# top1 accuracy: 69.5 top5 accuracy: 88.8
# reference top1 accuracy: 69.1
bash tools/dist_test.sh configs/shufflenet_v2/shufflenet_1x_b256x4_imagenet.py ckpts/shufflenet_1x_b256x4_imagenet/shufflenet_1x_b256x4_imagenet.pth 4 --metrics accuracy

bash tools/dist_train.sh configs/shufflenet_v2/weightnet_1x_b256x4_imagenet.py 4
# ckpts/weightnet_1x_b256x4_imagenet
# top1 accuracy: 70.9 top5 accuracy: 89.9
# reference top1 accuracy: 71.2
bash tools/dist_test.sh configs/shufflenet_v2/weightnet_1x_b256x4_imagenet.py ckpts/weightnet_1x_b256x4_imagenet/weightnet_1x_b256x4_imagenet.pth 4 --metrics accuracy

Acknowledgement

• WeightNet
• ShuffleNet-Series
• mmclassification

License

This project is released under the Apache 2.0 license.