Code used by the paper "What is the Role of Recurrent Neural Networks (RNNs) in an Image Caption Generator?". https://arxiv.org/abs/1703.09137
This paper investigates whether the RNN should be viewed as a sequence generator or as a sequence encoder. It does so by comparing the performance of a caption generator when the image features are included in the RNN or outside of the RNN. Paper will appear at the INLG 2017 conference.
Works on both python 2 and python 3 (except for the MSCOCO evaluation toolkit which requires python 2) and uses TensorFlow 1.0.
Python dependencies (install all with pip):
- tensorflow
- future