impression2805 / cvpr21_pass Goto Github PK

Thanks for your great job at first!
In SSL，the number of the logits have multiplied 4. Is it necessary？
What if we just augment the data without multipling the number of the logits by 4 ?

Hello, I wonder why the output and soft_feat_aug should be divided by args.temp when compute ce loss?

Is there a plan to release the code of the TinyImageNet and ImageNet-Subset data loaders（e.g. orders）, which is convenient to follow?

First of all, thank you for the great code repository. I noticed that you have mentioned in the paper that you are using the nearest class mean (NCM) classifier to classify between the classes. However, I could not locate the NCM classifier within the code. Could you please clarify?

It seems that the prototype augmentation is very similar to semantic augmentation. Are there any specific differences in details?

Implicit semantic data augmentation for deep networks. NIPS.

请问下论文的公式是不是有点问题？
公式（2）求和下标为n，但没有用到
公式（8）总共的loss，但解释的Lt,ce项基于旋转后的数据作交叉熵，应该是对新数据+旋转数据还是单纯新数据？kD项也只是对旋转后的数据，原来数据没用？但图看起来是都用。

Hello , I wonder why the code in train() function writing like this?

target = torch.stack([target * 4 + k for k in range(4)], 1).view(-1)

instead of using this

target = torch.stack([target * k for k in range(4)], 1).view(-1)

代码里cifar的顺序是从0到99，这是论文里使用的顺序吗？对于其他的数据集呢？感谢～

Is PASS similar to the comparison method (ours Gaussian) in the GFR (generative feature replay) work, although SSL is not used in GFR. Many thanks.