DataLoader for a VDSR network

Hello all
I’m trying to run a VDSR network in Pytorch and I wonder how I should use the DataLoader.

I have made 4 directories:

  • HR: the ground truth images
  • LR_div2: each image from the HR folder is downsampled by a factor 2 then upsampled again using bilinear interpolation. They have the same name as the original image
  • LR_div3 : the same with factor 3
  • LR_div4 : the same with factor 4

Each folder has subfolders ‘train’ and ‘test’. Now I’m wondering how I can write the DataLoader.

I have found this piece of code:

class AutoEncoderDataSet(Dataset):
    def __init__(self, dir_lr, dir_gt, traintest, transform=None):
        self.dir_lr = self.load_dir_single(join(dir_lr, traintest))
        self.dir_gt = self.load_dir_single(join(dir_gt, traintest))
        self.transform = transform

    def is_image_file(self, filename):
        return any(filename.endswith(extension) for extension in [".png", ".PNG", ".jpg", ".JPG", ".jpeg", ".JPEG"])

    def load_img(self, filename):
        img =
        return img

    def load_dir_single(self, directory):
        return [join(directory, x) for x in listdir(directory) if self.is_image_file(x)]

    def __len__(self):
        return len(self.dir_in)

    def __getitem__(self, index):
        img_lr = self.load_img(self.dir_lr[index])
        img_gt = self.load_img(self.dir_gt[index])
        sample = {'img_in': img_lr, 'img_gt': img_gt}

        if self.transform:
            sample = self.transform(sample)

        return sample

together with

ps = {
    'DIR_LR2': PATH + 'LR_div2',
    'DIR_LR3': PATH + 'LR_div3',
    'DIR_LR4': PATH + 'LR_div4',
    'DIR_HR':  PATH + 'HR'
train_set = AutoEncoderDataSet(ps['DIR_LR2'], ps['DIR_HR'], 'train', composed)
train_loader = DataLoader(train_set, batch_size=BATCH_SIZE_TRAIN, shuffle=True, num_workers=4)

But how can I take into account all 3 LR folders?
Thanks for your help.

If you want to concatenate the samples from all 3 directories, you could add another class method which is able to accept multiple folders, could call load_dir_single on each one of them, and concatenate the samples in the end.

Thanks for your answer. I’m not sure to understand what you mean, as I’m quite new to pytorch.

I thought that the first argument of DataLoader should be a list of couples, made of an input image and its associated ground truth.
Do you suggest that it could be a list of 4 images (3 LR and the ground truth) ?

The first argument to a DataLoader is a dataset.
In your current code snippet you are already working with a custom Dataset and you can adapt it as needed. My suggestion was to pass all folders into AutoEncoderDataset and to load all directories into a single list.