winycg / hsakd Goto Github PK

Hi, thanks for your wonderful work. We wonder about the detailed training GPU device and training time. And we train the model on 6 GeForce 3090 for now, it will take about 2 hours per epoch.

I found the ss_acc1 and class_acc1 in your output .I want to ask a question the meanings of ss_acc1 ?

You have submitted the 80.7 for wrn-40-2 (Teacher Model), is it class_accuracy or ss_accuracy? I got this accuracy but as a class_accuracy, but you referreing best accuracy to the ss_accuracy that is confusing me.

Thanks for your work. I want to know how to use self-supervised learning in this code. Do you implement self-supervised leanring through following code?

Hi, has the article been released?

Hi, thanks for your cool work!

The released resnet18 checkpoint ' resnet18_imagenet_aux_best.pth.tar' has a size of about 452M, which is much larger than the original one (44.7M). What is the difference between them?

Thanks!

I have also asked question there on your email, please reply me.

Thanks for the great work! However, I found a mismatch error when importing your pre-trained weights（ResNet50_aux.pth.tar）:

RuntimeError: Error(s) in loading state_dict for ResNet_Auxiliary:
        size mismatch for backbone.conv1.weight: copying a param with shape torch.Size([64, 3, 3, 3]) from checkpoint, the shape in current model is torch.Size([64, 3, 7, 7]).

Is it caused by the difference between the size of the convolution and the pre-trained model?

# models/resnet_imagenet.py   line32
self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False)

How can I go about solving it? Looking forward to your reply!

I reproduced your method on Res50->MobileNetV2 and got 79.24 top-1 acc, that's awesome. But I reproduce other KD methods SSKD 、CRD... ，I found that my results were nearly to their paper, but were very different from your paper results.(For example, I used SSKD and got acc 72.46, SSKD paper 72.57, your paper 78.21). Could you tell me why? Thanks!

There are 3 values after ss_accuracy, 4 values after class_accuracy, and finally there is a best_accuracy. What are the meanings of these values

?

Hi, Thanks for your work , it is helpfue for me.

I read the auxiliary classifiers code for different model and find there is various between different model. could you tell me the rule for modifying new auxiliary classifiers? Thx a lot