Shufflenet-v2-Pytorch Introduction This is a Pytorch implementation of faceplusplus's ShuffleNet-v2. For details, please read the following papers: ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design Pretrained Models on ImageNet We provide pretrained ShuffleNet-v2 models on ImageNet,which achieve slightly better accuracy rates than the original ones reported in the paper. The top-1/5 accuracy rates by using single center crop (crop size: 224x224, image size: 256xN): Network Top-1 Top-5 Top-1(reported in the paper) ShuffleNet-v2-x0.5 60.646 81.696 60.300 ShuffleNet-v2-x1 69.402 88.374 69.400 Evaluate Models python eval.py -a shufflenetv2 --width_mult=0.5 --evaluate=./shufflenetv2_x0.5_60.646_81.696.pth.tar ./ILSVRC2012/ python eval.py -a shufflenetv2 --width_mult=1.0 --evaluate=./shufflenetv2_x1_69.390_88.412.pth.tar ./ILSVRC2012/ Version: Python2.7 torch0.3.1 torchvision0.2.1 Dataset prepare Refer to https://github.com/facebook/fb.resnet.torch/blob/master/INSTALL.md#download-the-imagenet-dataset
gdjmck/Shufflenet-v2-Pytorch
Perfect implement. Model shared. x0.5 (Top1:60.646) and 1.0x (Top1:69.402).
PythonBSD-2-Clause