Context

MNIST is famous for being a deep learning beginner-friendly dataset for classification of handwritten digits. This is usually solved using CNNs. However, the same dataset can be used in reverse fashion to train a neural network like GAN to generate 'handwritten' digit images.

Content

The dataset contains 4 files corresponding to the features and labels for both the training and test set. This allows for easy load-and-train and relieves the user from doing data manipulation. As mentioned here, the images (labels in the image generation case) are represented in a flattened manner with each column representing a pixel in the 28x28 sized image. Also, the original labels (features in the image generation case) have been one-hot encoded for use in training for image generation.

Acknowledgements

The CSV format of the original MNIST dataset has been taken from here