keras implementation of A simple neural network module for relational reasoning
Relation network is a noval neural network introduced by deepmind in A simple neural network module for relational reasoning. It can achieve super-human performance in challenging visual question answering datasets such as CLEVR.
I implement Relation network using keras and train it on a challenging visual question answering dataset called Cornell NLVR. The training is in progress. The temporal test accuracy is 89.10%, which is much higher than the previous state of the art (61.99%).