Zero Shot Learning in Scene Graph Generation

Following this paper and it's code, we added a zero-shot learning approach where our model could predict words even out of training data, by making the word predict word embedding instead of one-hot vector.

You can find the detailed report here