YUI

Good questions. First, `param.data.mul_(1 - self.wd)` is weight decay.It is also used in the official code (https://github.com/google-research/mixmatch/blob/master/mixmatch.py#L92), but it is different from the normal weight decay of pytorch (which is...

In Sec. 3.1 of the paper, ''For each u in the batch of unlabeled data U, we generate K augmentations.'' And K is 2 in the paper, so u1 and...

They are from the same batch but with different augmentations. See here: https://github.com/YU1ut/MixMatch-pytorch/blob/master/dataset/cifar10.py#L7-L14

See this: https://github.com/YU1ut/MixMatch-pytorch/issues/20#issuecomment-583770295

In my opinion, this method needs to see the same sample with different augmentations a lot of times and get enough mixup samples to improve the performance. So, it is...

Did you try to initialize the bias in Conv2D to 0?

First, try lambda-u = 0 and change alpha to get the best performance, I think alpha=0.2 is a good start. Then, try to change lambda-u. You also should try different...

Thank you for reporting this. I will investigate this.

Because the score is logarithmic. https://github.com/sgrvinod/a-PyTorch-Tutorial-to-Image-Captioning/blob/master/eval.py#L111

No. Only base classes are used in pre-training. https://github.com/YU1ut/imprinted-weights/blob/master/loader.py#L27 And the network only classifies 100 classes. https://github.com/YU1ut/imprinted-weights/blob/master/models.py#L6