Self implementation of ThinMNIST (in a DataLoader for PyTorch).
I strongly recommend using the preprocessed version with the thin_mnist.py
DataLoader, instead of using thin_mnist_slow.py
.
NOTE that images are padded with 2 pixels on each border, changin the MNIST image size from 28x28
to 32x32
!
Check test.ipynb
for some output examples.