Pytorch implementation of Memory Augmented Neural Network(Santoro et al.)

Download the omniglot Dataset from here and put all the images(evaluation + background) in one folder. Then run resize_images.py there.

Basic implementaion is in mann_pytorch.ipynb

Tasks Completed

• Basic Implementation of MANN using LSTM
• Fix Training loss error
• PreFetching Dataloader added
• Training for:-
  • 5 shot 1 way
  • 4 shot 2 way
• Use Bi-LSTM/multi layered LSTM
• Use a siamese network as its head(wont work most probably)
• Use NTM module

When ran for 259steps model shows expected behavior.Loss is decreasing and accuracy is varied(expected). Must run this for about 20k steps. Time taken is quite high rn.