How to Create a Custom Dataset with Multiple Labels and Masks?

vision
1

Pytorch newbie here! Hi all!

I’m trying to create a custom dataset to plug into DataLoader that is composed of single-channel images (20000 x 1 x 28 x 28), single-channel masks (20000 x 1 x 28 x 28), and three labels (20000 X 3).

Following the documentation, I thought I would test creating a dataset with a single-channel image and a single-channel mask, using the following code:

class CustomDataset(Dataset):

    def __init__(self, images, masks, transforms=None, train=True): 
        self.images = images
        self.masks = masks
        self.transforms = transforms

    def __getitem__(self, index):
        image = self.images.iloc[index, :]
        image = np.asarray(image).astype(np.float64).reshape(28, 28, 1)
        mask = self.masks.iloc[index, :]
        mask = np.asarray(mask).astype(np.float64).reshape(28, 28, 1)
        transformed_image = self.transforms(image)
      
        return transformed_image, mask

    def __len__(self):  

       return len(self.images)

Using the class, I form the dataset from two pandas dataframes and plug into DataLoader.

transform = transforms.Compose(
    [transforms.ToPILImage(),
     transforms.ToTensor(),
     transforms.Normalize((0.5, ), (0.5, ))
])

train_images = pd.read_csv('train.csv')
train_masks = pd.read_csv('masks.csv')

train_data = CustomDataset(train_images, train_masks, transform)
trainloader = DataLoader(train_data, batch_size=128, shuffle=True)

I would expect the shape of a single batch in trainloader to be ([128, 1, 28, 28], [128, 1, 28, 28]), for both the image on the left and the mask on the right.

Instead the shape of a single batch trainloader is ([128, 1, 28, 28], [128]), which makes me think that the masks have somehow been transformed into labels.

How do I fix this, and how do I add in the actual labels? Thanks in advance for your help!

2

I cannot reproduce the issue using random inputs:

class CustomDataset(Dataset):
    def __init__(self, transforms=None): 
        self.images = np.random.randn(1000, 28, 28, 1).astype(np.float32)
        self.masks = np.random.randn(1000, 28, 28, 1).astype(np.float32)
        self.transforms = transforms

    def __getitem__(self, index):
        image = self.images[index]
        mask = self.masks[index]
        transformed_image = self.transforms(image)
      
        return transformed_image, mask

    def __len__(self):  
       return len(self.images)
   
    
transform = transforms.Compose(
    [transforms.ToPILImage(),
     transforms.ToTensor(),
     transforms.Normalize((0.5, ), (0.5, ))
])

train_data = CustomDataset(transform)
trainloader = DataLoader(train_data, batch_size=128, shuffle=True)

data, mask = next(iter(trainloader))
print(data.shape)
> torch.Size([128, 1, 28, 28])

print(mask.shape)
> torch.Size([128, 28, 28, 1])

and get the expected shapes.
Could you check, what the difference between the codes might be?