Value error when using Dataloader

vision
1

Recently, when I am using pytorch to build a simple image classifier, a value error occurred when using Dataloader. Here is the error message:

Epoch: 1/20

Train Loss 0.5450 Acc0.3640
Traceback (most recent call last):
  File "ms_resnet.py", line 135, in <module>
    optimizer_resnet,lr_scheduler_resnet,num_epochs=20)
  File "ms_resnet.py", line 82, in get_model
    for data in test_loader:
  File "/home/guest/.local/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 259, in __next__
    batch = self.collate_fn([self.dataset[i] for i in indices])
  File "/home/guest/.local/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 259, in <listcomp>
    batch = self.collate_fn([self.dataset[i] for i in indices])
  File "/home/jiahao/anaconda3/envs/guest/lib/python3.6/site-packages/torchvision/datasets/folder.py", line 124, in __getitem__
    img = self.transform(img)
  File "/home/jiahao/anaconda3/envs/guest/lib/python3.6/site-packages/torchvision/transforms/transforms.py", line 42, in __call__
    img = t(img)
  File "/home/jiahao/anaconda3/envs/guest/lib/python3.6/site-packages/torchvision/transforms/transforms.py", line 286, in __call__
    i, j, h, w = self.get_params(img, self.size)
  File "/home/jiahao/anaconda3/envs/guest/lib/python3.6/site-packages/torchvision/transforms/transforms.py", line 271, in get_params
    i = random.randint(0, h - th)
  File "/home/jiahao/anaconda3/envs/guest/lib/python3.6/random.py", line 220, in randint
    return self.randrange(a, b+1)
  File "/home/jiahao/anaconda3/envs/guest/lib/python3.6/random.py", line 198, in randrange
    raise ValueError("empty range for randrange() (%d,%d, %d)" % (istart, istop, width))
ValueError: empty range for randrange() (0,-15, -15)

and here is my code:

train_set=datasets.ImageFolder(os.path.join(data_dir,'train'),train_transform)
test_set=datasets.ImageFolder(os.path.join(data_dir,'val'),test_transform)
train_loader=torch.utils.data.DataLoader(train_set,batch_size=5,shuffle=True,num_workers=5)
test_loader=torch.utils.data.DataLoader(test_set,batch_size=5,shuffle=True,num_workers=5)
...

for epoch in range(num_epochs):
    ...
    model.train(True)
    for data in train_loader:
    ...
    #training process
    loss.backward()
    optimizer.step()

    model.train(False)
    for data in test_loader:
    ...
    #validation process

As you can see, the DataLoader in the training process works well, but it fails in the validation. I’ve googled it and I think maybe the something is wrong in the random.py file in python. So can anyone give some suggestion on how to deal with this?

2

Could you share some information about test_transform?
It seems you would like to perform something like a random cropping on you images and self.size seems to be misdefined.

3

Here is my image transforms code:


train_transform=transforms.Compose([
		transforms.CenterCrop(224),
		transforms.RandomRotation(5),
		transforms.ToTensor(),
#for transfer learning 
		transforms.Normalize(mean = [0.485, 0.456, 0.406],std = [0.229, 0.224, 0.225])])
test_transform=transforms.Compose([
		transforms.RandomCrop(224),
		transforms.RandomRotation(3),
		transforms.ToTensor(),
		transforms.Normalize(mean = [0.485, 0.456, 0.406],std = [0.229, 0.224, 0.225])])
4

Ok, it seems some images are smaller than your desired crop size of 224.
Could you please check this? Appenrently some images are 208 x 208, since the error message states randrange is called with -15. You can see the line of code here.

Code to reproduce the error:

test_transform=transforms.Compose([
    transforms.ToPILImage(),
	transforms.RandomCrop(224),
	transforms.RandomRotation(3),
	transforms.ToTensor(),
	transforms.Normalize(
        mean = [0.485, 0.456, 0.406],
        std = [0.229, 0.224, 0.225])]
)

x = torch.randn(3, 208, 208)
x_ = test_transform(x)
1 Like
5

Just like @ptrblck said, usually we resize the image a bit larger than the network inputs then we could use the crop fun correctly.

6

Oh yes, you are correct. I check some images and they do appear to be smaller than my desired size. I’ve changed the code to RandomResizedCrop() and it worked.
Thank you again for your kindness help.