Dimension issue in correct prediction

[ildoonet]

Dimension issue in correct prediction

[ildoonet]

@NomiMalik0207 You need to debug line by line, especially I suggest you to look tensor sizes. In my experience, when the accuracy is above 1.0, it was (mostly) due to the wrong processed tensors. I always check the tensor sizes for this reason.

Also, this is not related to cutmix technically, but I will open this for now to help you. Let me know the result after you debug.

[ildoonet]

you can change predicted vector to one hot [64, 7]. or you should check the prediction accuracy only on validation set which would have target size [64]. If codes are provided, it will be much easier to answer. Thanks.

[NomiMalik0207]

i solved this error by matching the size but now in while testing model, the accuracy of test set start from 1200 or above out of 100. the code for test set is

for batch_idx, (inputs, targets) in enumerate(Testloader):                                                                                                     
        bs, ncrops, c, h, w = np.shape(inputs)                                                                                                                           
        inputs = inputs.view(-1, c, h, w)                                                                                                                                
        if use_cuda:                                                                                                                                                     
            inputs, targets = inputs.cuda(), targets.cuda()                                                                                                              
        inputs, targets = Variable(inputs, volatile=True), Variable(targets)                                                                                             
        outputs = net(inputs)                                                                                                                                                                                                                                                                                                    
        outputs_avg = outputs.view(bs, ncrops, -1).mean(1)                                                                                                                                                                                                                               
        loss = criterion(outputs_avg, targets)                                                                                                                           
        Test_loss += loss.item()                                                                                                                                   
        _, predicted = torch.max(outputs_avg.data, 1)                                                                                                                    
        predicted= predicted.unsqueeze(1) // for matchin size                                                                                                                               
        total += targets.size(0)                                                                                                                                         
        correct += predicted.eq(targets.long().data).cpu().sum()// target should be long tensor                                                                                                                                                                                                                               
        utils.progress_bar(batch_idx, len(Testloader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)'                                                                          
                           % (Test_loss / (batch_idx + 1), 100. * correct / total, correct, total))

[ildoonet]

@NomiMalik0207 Don't use CutMix wrapper for Testloader. Seems to me, predicted= predicted.unsqueeze(1) is not the solution for you.

1. Remove CutMix wrapper for Testloader, not Trainloader.
2. no .unsqueeze(1). Use torch.argmax to get predicted class indices and count the number of correctness with targets.

I hope this will help you.

[NomiMalik0207]

HERE is the results of your suggested modification

1. i already didnot use CutMix wrapper for testloader, So i removed predicted= predicted.unsqueeze(1) from testing solved the problem.
2. when i use torch.agrmax with trainloader, the train accuracy started from more then 100%. but its going normal with unsqueeze.
   Seeing farward to your helpful suggestions .
   Thank you mate

[NomiMalik0207]

Thank you for your time and help.
Here i showed the training loop for my network. The size of tensors are given as comment at the end of line.
for batch_idx, (inputs, targets) in enumerate(trainloader):
import pdb
if use_cuda:
inputs, targets = inputs.cuda(), targets.cuda()
inputs, targets = Variable(inputs, volatile=True), Variable(targets)
optimizer.zero_grad()
outputs = net(inputs)
loss = criterion(outputs, targets)
train_loss += loss.item()
pdb.set_trace()
_, predicted = torch.max(outputs.data, 1)// predicted.size()=torch.Size([64])
total += targets.size(0)
predicted = torch.argmax(predicted)//predicted.size()=torch.Size([])
// here when i print the value of predcted, it shows predicted= tensor(38)
#predicted= predicted.unsqueeze(1)// if i use this line, this shows //predicted.size()=torch.Size([64, 1])
correct += predicted.eq(targets.long().data).cpu().sum()// targets.size()=torch.Size([64, 7])
loss.backward()
optimizer.step()