pytorch-cifar
pytorch-cifar copied to clipboard
kuangliu
can't achieve the reported accuracy on VGG19
Have anyone train a VGG19 network used this code? I used the total exactly the same code in the repository except using a lr_schuduler to change the learning rate automatically. But I can only achieve the accuracy around 88%. Can anyone tell me what's wrong?
For me it would achieve ~90.8%, but I still cannot get the reported 92.64%. @kuangliu care to share your idea?
I've got 93.37% with ubuntu 16.04, cuda 9.2, cudnn 7.2, pytorch 0.4.1, and 4 * gtx 1080 ti. I use batch_size=128 (means 32 for each gpu) epoch=450 learning rate schedule 0.1 when epoch < 150, 0.01 when epoch < 300, 0.001 when epoch < 450 momentum=0.9, weight_decay=5e-4
@zeal-github @EnJiang I think the main problem is that you didn't set the learning rate schedule right. I got 93.69% using the same learning rate schedule as default. You can see pull request #26 for a automatic learning rate schedule.
@hjinlee88 @gaosh Are you sure your results reproducible? I used the same learning rate schedule with yours, but none of my experiments could reach 92% acc. I especially have done many experiments on vgg16, the best result is 90.68%, which is much lower than 92.64% reported by the author.
@hjinlee88 @gaosh Are you sure your results reproducible? I used the same learning rate schedule with yours, but none of my experiments could reach 92% acc. I especially have done many experiments on vgg16, the best result is 90.68%, which is much lower than 92.64% reported by the author.
I think We should use The same random seed, or IT is not reproducible. any body share The random seed
@myhub @lhyfst if the base result is 90.68%, I don't think change random seed would help. The margin is near 3%, such huge difference can't be solely explained by random seed.
@cbasemaster , It's testing accuracy. Training accuracy is meaningless. I follow the hjinlee88 suggestion, the best test accuracy is 93.98%.
