README.md

SKNet

Implemenation of Selective Kernel Networks by pytorch.

The architecture of SK is as follows

I trained SKNet50 on ImageNet-2012 from scratch and got an accuracy of 21.26, which did not reach the performance of 20.79 in the paper. If somebody know what caused the problem, please leave me a message.

The pretrained weights are provided below.

Requirement

• pytorch 1.4.0+
• torchvision
• tensorboard 1.14+
• numpy
• pyyaml
• tqdm
• pillow

Dataset

• ImageNet-2012

Pretrained Model on ImageNet-2012

Architecture	Top-1 error	Pretrained model
SKNet50 (My Imp.)	21.26	Google Drive Baidu Netdisk
SKNet50 (paper)	20.79	None

If you want to use my pretrained weight, you should do

1. place the downloaded pretrained model in runs/sknet_imagenet/86028 folder under this project
2. config the attribute of runid and cuda in the config file configs/sknet_imagenet.yml
3. run validata.py or test.py (For test, you should specify the img_path in the test.py)

The error curve of SKNet50 during my training process is shown below