ibelievecjm / pseudo_label-pytorch Goto Github PK

labeled_bs = self.labeled_bs
labeled_loss = torch.sum(self.loss_fn(outputs, targets)) / labeled_bs
with torch.no_grad():
pseudo_labeled = outputs.max(1)[1]
unlabeled_loss = torch.sum(targets.eq(NO_LABEL).float() * self.loss_fn(outputs, pseudo_labeled)) / (data.size(0)-labeled_bs +1e-10)
loss = labeled_loss + self.unlabeled_weight()*unlabeled_loss

pseudo_labeled = outputs.max(1)[1] #outputs should be unlabel data,but your code outputs is model(label_data)??????

'af' = 3 in the paper Pseudo-Label The Simple and Efficient Semi-Supervised Learning Method for Deep Neural Networks ,
I would like to know whether 0.3 in here help model perform better than 3 in paper.

paper shows that they pretraining use all data in a unsupervised way.I'm wondering where you pretrain your net in your codes.

The 'datadir' in util.datasets.py in line 27 shows that cifar10 should be stored in the route, but I can't find the file. Can you upload the file? I'm a green hand in deep learning and have no idea about distribution method in datasets. I would be obliged if you could help me solve the problem.

在linux按照指令操作为什么还是不行？

`# make the pseudo label for unlabeled data
output2 = net(img2);
_, label2 = torch.max(output2, 1);

#semi-supervised loss
loss = criterian(output1, label1) + alpha * criterian(output2, label2);`

Could you please explain why criterian() for unlabeled data works? Because label2 is actually calculated from output2, they are the same indeed? Why non-zero loss is obtained?