TypeError: default_collate: batch must contain tensors, numpy arrays, numbers, dicts or lists; found <class 'NoneType'>

vision
1 
torch.manual_seed(101)
num_epochs=30
for epoch in tqdm(range(num_epochs)):
    for step,i in enumerate(train_loader):
        img,label_one_hot,label = i
        img = Variable(img).cuda()
        label_one_hot= Variable(label_one_hot.float()).cuda()
        pred = cnn_model(img)
        loss_value = loss_fn(pred,label_one_hot)
        optimizer.zero_grad()
        loss_value.backward()
        optimizer.step()
        print('epoch:', epoch+1, 'step:', step+1, 'loss:', loss_value.item())

This is my training loop

class Mydataset(Dataset):
    def __init__(self,path,is_train=True,transform=transform):
        self.path = path
        if is_train: self.img = os.listdir(self.path)
        self.transform = transform
    def __getitem__(self,idx):
        img_path = self.img[idx]
        files = os.listdir(self.path)
        for file in files:
            if (file.endswith(".jpg") and file[:-4] == length_of_captcha):
                img = cv2.imread(img_path,cv2.IMREAD_GRAYSCALE)
                kernel = cv2.getStructuringElement(cv2.MORPH_RECT,(5,12))
                img = cv2.morphologyEx(img, cv2.MORPH_CLOSE, kernel)
                label = Path(self.path/img_path).name[:-4]
                label_ = []
            else:
                continue
            for i in label:
                label_ += encode(i)
            if self.transform is not None:
                img = self.transform(img)
            return img,np.array(label_),label
    def __len__(self):
        return (len(self.img))

Here I am trying for character recognition for 6 digit captcha image after some preprocessing. I am getting the error message in the title
This is my dataset class and here I have passed successfully creating train_loader. But I am still encountering this and I don’t know what is going wrong. Can anyone help me?

2

Based on the error message I guess that some return values are empty.
Could you add a condition into __getitem__ before the return statement and check if img, label_, and label have valid values?

3 Likes
3

Hello there, i am new to PyTorch and get the same error from my dataloader. I am not sure if i should open a new discussion or post it here?
Anyway, @ptrblck by check if the variables have valid values, you mean to print their shape?
Many thanks for any help already in advance!

4

The error message claims that NoneTypes are found in the default_collate, so you would need to check why some samples are loaded as None.

5

Ok, i found the bug. Thanks for your help.