The unoffical PyTorch implementation of WeightNet, based on OpenMMLab mmclassification.
See install.md
and getting_started.md
.
Learn more at mmcls's documentation.
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
This project is released under the Apache 2.0 license.