ternarynet

ternarynet copied to clipboard

Reproduced ResNet18 CIFAR10 result is 10% lower than reported

6

Hi @czhu95 , Thanks for providing the codes! Recently I use your codes to ternarize a ResNet18 using CIFAR10. Firstly I use tensorpack to train a ResNet18 to validation error...

csyhhu

Hi @czhu95 , Thanks for your interesting work. I ran your code to reproduce your results of resnet on Cifar-10 from scratch. However I see the results are not similar...

haithanhp

Hi there, I'm trying to re-run p-cifar10-resnet.py, but I'm unable to find `build_prune` https://github.com/czhu95/ternarynet/blob/3fe8613767cdab084390f8894f4551a690111826/examples/Ternary-Net/p-cifar10-resnet.py#L15 Is this included in the repository? Thanks

skoppula

您好，论文中做ternary之前，先对weight进行归一化，即除以最大的abs(w)归一化到[-1, 1]。 根据论文中设定threshold的方法  weight已经归一化，最大的max(abs(w))是1，因此delta就是t，也就是文中的0.05。 请问是这样吗？谢谢

camel007

Hi Chenzhuo, I was studying your Ternarynet work and was trying to run the Alexnet+Imagenet training. `'python` ./tw-imagenet-alexnet.py --gpu 0,1,2,3 --data `MY_IMAGENET_PATH'` {I have arranged the data in the format...

srkm009