Pytorch dataset loaders that can be cloned into any project. Currently provides:

- MNIST
- RotatedMNIST
- FashionMNIST
- ClutteredMNIST
- PermutedMNIST
- CIFAR10
- SVHN
- Omniglot
- SVHN Original (non-centered)
- Dataset Operators (see section below)

This is the main entrypoint to the datasets. It expects an argparse object as input with the following members:

- args.task : what dataset to load (eg: 'mnist')
- args.data_dir : where to save / load data from
- args.batch_size : batch size for train and test loaders
- args.cuda : this is needed in order to pin memory to GPU for data loaders and to create more workers

Note: all simple datasets are auto-downloaded into data_dir if they dont exist there. For larger datasets (eg: mini-imagenet) the loader will expect the data to exist in data_dir.

Each dataloader has a .train_loader and .test_loader that can be utilized in order to iterate over data, eg:

from datasets.loader import get_loader

# use argparse here to extract required members

mnist = get_loader(args)
for data, label in mnist.train_loader:
    # do whatever you want with the training data

Since all the loaders are also operating over images they have the following members:

- loader.img_shp : the size of the image as [C, H, W]
- loader.batch_size : the batch size of the loader
- loader.output_size : dimension of labels (eg: 10 for MNIST)

You can get a sequential loader by using the get_sequential_data_loaders function. This takes the dataset and splits it into several datasets (eg: MNIST --> 10 datasets with each individual digit).

Appending + between datasets in args.task will return a merged dataset, eg: mnist+fashion returns a mixture dataset. Currently it is hardcoded to reshape all data to (32, 32) (to be fixed in future). This can be used with batch OR sequential datasets

Appending rotated_ to args.task with rotate that entire dataset. This can be used with sequential or normal batch datasets.

The utils.py file houses dataset transformations such as bw_2_rgb and resize which allow to convert black and white images to RGB. The resize transform allows to resize all images in the dataset (particularly useful if you are merging two datasets)