Unsupervised Data set reading

Debanjan_Sengupta · August 28, 2017, 3:45pm

Is there any definite way in PyTorch for loading only images from a folder without any labels or ground truths as in a supervised classification task?

AashishV · August 28, 2017, 3:50pm

I hope this helps, http://pytorch.org/docs/master/torchvision/datasets.html#imagefolder

Debanjan_Sengupta · August 28, 2017, 4:02pm

Thanks. But I was asking how to load images in a folder into batches without any target (or labels) as in a supervised task.

Just like in autoencoders there are no image labels as such.

vabh · August 28, 2017, 4:28pm

Hi,

Have a look here to see how a generic Dataset can be implemented:

github.com

pytorch/vision/blob/master/torchvision/datasets/cifar.py

from __future__ import print_function
from PIL import Image
import os
import os.path
import errno
import numpy as np
import sys
if sys.version_info[0] == 2:
    import cPickle as pickle
else:
    import pickle

import torch.utils.data as data
from .utils import download_url, check_integrity


class CIFAR10(data.Dataset):
    """`CIFAR10 <https://www.cs.toronto.edu/~kriz/cifar.html>`_ Dataset.

    Args:

This file has been truncated. show original

In particular, the __getitiem__ method, which returns a tuple comprising (data, label)

The generic loop is something like:

for (data, labels) in dataloader:
    # train / eval code

You’re free to ignore the label here and you can train an autoencoder on cifar10, for example, pretty much out of the box.