alanchou / deep-co-training-for-semi-supervised-image-recognition Goto Github PK

Hey Alan,
Thanks for your great work.
I ran your code on python 3.6 (you are using 3.5 but I don't see this is a problem), I could only get accuracy around 88%. I have also tried to add .detach() to perturbed example when creating pseudo label, result is still lower than 89%.
I am wondering if there is any other tricks to get 89% acc?
Thanks again for your work.

Cheers

Hi,

I am reading your code, which is a great implementation.
But I have a question. In the following code, why the batch_size in testloader is set to 100 but 1 in trainloader.

testloader = torch.utils.data.DataLoader(testset, batch_size=batch_size, shuffle=True, num_workers=2)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=1, shuffle=False, num_workers=2)

Hi Zhou,
Very happy to see your repo. I am noticing something and want to ask for the details.

The main idea of adversarial loss for two view is to let a hard example of model1 to be easy for model2. So I reduce the loss function KL(predict_model2(adversarial image targeted at model1), predict_model1(image) ). If the two models are the same, predict_model1(image) should perform well than predict_model2(adversarial image targeted at model1). I would wonder why there is no .detach() operation for the target tensor? Without detach(), you would reduce the performance of easy example.

I noticed that normalization of data is not used(the code is commented out) when using CIFAR10 . Is it needed?