This is a Tensorflow implementation of A simple neural network module for relational reasoning. Currently only the simple Sort-of-CLEVR dataset (described in paper) is implemented. This dataset simplifies the problem by considering a smaller space of questions that can be encoded as binary strings and thus bypassing the need for a model to parse natural language but still retaining the need for relational reasoning.
$ python -m rn.data.sort_of_clevr.inspect
$ python -m rn.train$ python -m rn.predict --checkpoint path/to/checkpoint [--only-misclassifications] [--seed xxx]
TODO: Show accuracy and loss curves.