How to normalize the custom dataset

vision
1

Context: I am doing image segmentation using Pytorch, before feed the training data to the network, I need to do the normalisation
My image size is 256x256x3, and my mask size is 256x256x3
I have a TrainDataset class, and my sample is a dict type
for my image, I should use: sample['image'] for my image and sample['mask'] for the mask
The Question is: How can I do the normalization for a dict type dataset
I put my TrainDataset class below for information

class TrainDataset(Dataset):
    """Training dataset with mask image on gray scale/RGB"""
    def __init__(self, train_dir, semantic_dir, transform=None):
        """
        Args:
            train_dir (string): Directory with training images
            transform (callable): Optional transform to be applied on a sample
            semantic_dir (string): Directory with semantic segmentation training image
        """
        self.train_dir = train_dir
        self.transform = transform
        self.semantic_dir = semantic_dir
        
    def __len__(self):
        return len(os.listdir(self.train_dir))
    
    def __getitem__(self, idx):
        img_name = os.path.join(self.train_dir, os.listdir(self.train_dir)[idx])
        semantic_name = os.path.join(self.semantic_dir, os.listdir(self.semantic_dir)[idx])
        image = io.imread(img_name)
        semantic = io.imread(semantic_name)
        sample = {'image':image, 'semantic':semantic}
        
        if self.transform:
            sample = self.transform(sample)
        return sample

composed = transforms.Compose([RandomCrop(256), ToTensor()])

transformed_dataset = TrainDataset(train_dir=train, semantic_dir =train_mask, transform = composed)

dataloader = DataLoader(transformed_dataset, batch_size=4, shuffle = False, num_workers = 4)

for i_batch, (inputs, target) in enumerate(train_loader):
    print(i_batch, sample_batched['image'].size(), sample_batched['semantic'].size())
    print(sample_batched['image'].dtype)
    if i_batch == 3:   
        break

Here is the result for the data size:

0 torch.Size([4, 3, 256, 256]) torch.Size([4, 3, 256, 256])
torch.uint8
1 torch.Size([4, 3, 256, 256]) torch.Size([4, 3, 256, 256])
torch.uint8
2 torch.Size([4, 3, 256, 256]) torch.Size([4, 3, 256, 256])
torch.uint8
3 torch.Size([4, 3, 256, 256]) torch.Size([4, 3, 256, 256])
torch.uint8

Thank you in advance.

2

torchvision.transforms won’t take a dict, so you should call the transformations on your data and target directly or you could write an own transform method in your Dataset, which takes the specified dict as its input.

Besides that, note that the default tensor image shape in PyTorch is [batch_size, channels, height, width]. Currently it looks like the channel dimension is in a wrong position.
Also, could you give us some information on the segmentation mask? Based on the given shape it looks like the segmentation mask is a color image. If you would like to use a standard classification criterion like nn.CrossEntropyLoss, you would have to transform the colors to a class index.
If I’m correct in my assumption about the segmentation mask (different colors correspond to different classes), you should also be careful with the normalization, since this might distort your indices.

3

any special reason you set the shuffle to False?

shuffle = False
4

Hi, this is not a must, you could make the shuffle = True