ARC_Omniglot
The Attentive Recurrent Comparator (ARC) catches the difference between two characters in the same way that humans do: by iteratively glancing between the images.
Please see fellowship_submission.ipynb for analysis
To run the model
-
update pytorch to 0.4.1. Do this if you have conda:
- conda config --add channels soumith
- conda update pytorch torchvision
-
download the current repo
-
download omniglot.npy and omniglot_strokes.npz from here and place it under ./data/
- omniglot.npy is equivalent to vertically stacking the background and evaluation datasets from Brendan Lake's Omniglot repo
-
run: python ARC/train.py --name model
to continue training and read classification accuracies on test set.