FST

• Paper: Towards Fully Sparse Training, AAAI2022

Install CUDA Extension

• The extension implements the Im2col + GEMM based convolution.
• Our CUDA extension are build with gcc version 5.3.1

cd Sparse_Conv_Extension_CUDA
sh install.sh

Training scripts

python main_imagenet.py -a resnet18 --data path_to_ImageNet --checkpoint checkpoints/imagenet/res18

Check Sparse Mask

The trained model should be 2:4 structured sparsed (requiring four consecutive values containing at least two zeros).

python check_mask.py

Results on ResNet-18

Model	Pattern				Accuracy	Download
Float	Dense	-	-	-	71.2
ASP	2:4 structured		-	-	70.7
Ours	2:4 structured				71.0	google cloud

Evaluate with 2:4 sparsed model

python main_imagenet_eva.py -a resnet18 --data path_to_ImageNet --resume checkpoints/imagenet/res18/model_best.pth.tar --evaluate