How can I merge the two parts(training and extra) of SVHN dataset?

Hi, I noticed we can use SVHN dataset using the official dataset wrapper..
my question is, how can I merge the train and extra parts together and use them for training?
The API does not provide any options in this case.

Thanks alot in advance

I did it and here is the full sourcecode :
However, I noticed it consumes a lot of memory. Is there an easy way that I can dump what I have read and then read it batch by batch from there so that it does not consume that much memory?

from __future__ import print_function
import as data
from PIL import Image
import os
import os.path
import numpy as np
from utils import download_url, check_integrity

class SVHN(data.Dataset):
    """`SVHN <>`_ Dataset.
    Note: The SVHN dataset assigns the label `10` to the digit `0`. However, in this Dataset,
    we assign the label `0` to the digit `0` to be compatible with PyTorch loss functions which
    expect the class labels to be in the range `[0, C-1]`

        root (string): Root directory of dataset where directory
            ``SVHN`` exists.
        split (string): One of {'train', 'test', 'extra'}.
            Accordingly dataset is selected. 'extra' is Extra training set.
        transform (callable, optional): A function/transform that  takes in an PIL image
            and returns a transformed version. E.g, ``transforms.RandomCrop``
        target_transform (callable, optional): A function/transform that takes in the
            target and transforms it.
        download (bool, optional): If true, downloads the dataset from the internet and
            puts it in root directory. If dataset is already downloaded, it is not
            downloaded again.

    url = ""
    filename = ""
    file_md5 = ""

    split_list = {
        'train': ["",
                  "train_32x32.mat", "e26dedcc434d2e4c54c9b2d4a06d8373"],
        'test': ["",
                 "test_32x32.mat", "eb5a983be6a315427106f1b164d9cef3"],
        'extra': ["",
                  "extra_32x32.mat", "a93ce644f1a588dc4d68dda5feec44a7"]

    def __init__(self, root, split='train-full',
                 transform=None, target_transform=None, download=False):
        self.root = os.path.expanduser(root)
        self.transform = transform
        self.target_transform = target_transform
        self.split = split  # training set or test set or extra set

        # if self.split not in [self.split_list,'train-full']:
        #     raise ValueError('Wrong split entered! Please use split="train" or train-full '
        #                      'or split="extra" or split="test"')

        if split=='train-full':
          for splt in enumerate(self.split_list):
            if(splt[1] not in ['test']): 
        if download:

        if not self._check_integrity():
            raise RuntimeError('Dataset not found or corrupted.' +
                               ' You can use download=True to download it')

        # import here rather than at top of file because this is
        # an optional dependency for torchvision
        import as sio
        for i in range(len(self.filenames)):
          # reading(loading) mat file as array
          loaded_mat = sio.loadmat(os.path.join(self.root, self.filenames[i]))
 = np.concatenate((, loaded_mat['X']), axis=3)
          # loading from the .mat file gives an np array of type np.uint8
          # converting to np.int64, so that we have a LongTensor after
          # the conversion from the numpy array
          # the squeeze is needed to obtain a 1D tensor
          y = loaded_mat['y'].astype(np.int64).squeeze()
          self.labels = np.concatenate((self.labels, y),axis=0)
 = loaded_mat['X']

        #self.labels = loaded_mat['y'].astype(np.int64).squeeze()

        # the svhn dataset assigns the class label "10" to the digit 0
        # this makes it inconsistent with several loss functions
        # which expect the class labels to be in the range [0, C-1], self.labels == 10, 0) = np.transpose(, (3, 2, 0, 1))

    def __getitem__(self, index):
            index (int): Index

            tuple: (image, target) where target is index of the target class.
        img, target =[index], int(self.labels[index])

        # doing this so that it is consistent with all other datasets
        # to return a PIL Image
        img = Image.fromarray(np.transpose(img, (1, 2, 0)))

        if self.transform is not None:
            img = self.transform(img)

        if self.target_transform is not None:
            target = self.target_transform(target)

        return img, target

    def __len__(self):
        return len(

    def _check_integrity(self):
        root = self.root
        for i in range(len(self.filenames)):
          md5 = self.file_md5s[i]
          fpath = os.path.join(root, self.filenames[i])
        return check_integrity(fpath, md5)

    def download(self):
      for i in range(len(self.filenames)):
        md5 = self.file_md5s[i]
        download_url(self.urls[i], self.root, self.filenames[i], md5)

    def __repr__(self):
        fmt_str = 'Dataset ' + self.__class__.__name__ + '\n'
        fmt_str += '    Number of datapoints: {}\n'.format(self.__len__())
        fmt_str += '    Split: {}\n'.format(self.split)
        fmt_str += '    Root Location: {}\n'.format(self.root)
        tmp = '    Transforms (if any): '
        fmt_str += '{0}{1}\n'.format(tmp, self.transform.__repr__().replace('\n', '\n' + ' ' * len(tmp)))
        tmp = '    Target Transforms (if any): '
        fmt_str += '{0}{1}'.format(tmp, self.target_transform.__repr__().replace('\n', '\n' + ' ' * len(tmp)))
        return fmt_str

I have another example, does pytorch shuffle the info by itself or should I be doing the shuffling as well?
Do I still have to create a new transform lambda function for making the input to be in the range 0-1 or the the transform.ToTensor() suffices?

Thanks in advance

I see that you have done the merge/fusion yourself by tweaking the original class. Nonetheless, if you want to do this ‘train’-‘extra’ merge, you can inherit the SVHN Dataset class and combine the two ‘splits’ as one. This would be a much neater solution. Here’s an example of how this could be done:

As for your second question, the shuffle is done when you initiate the Dataloader, something like:
train_loader =, batch_size, shuffle=True, num_workers=2)

1 Like

Thanks a lot, I’ll keep that in mind. I’m new to python and know just the basic stuff that’s why my methods are not efficient at all.
However, Can you help me on somehow making this more memory efficient? the current(original) implementation tries to read everything into the memory and when both training and extra are merged, they take a huge amount of memory (more than 15+Gigabyte!) what are my options here?