RAM

Recurrent Attention Model

Code

code: 'ram_modified.py'

Playing with loss function in RAM

This project is a modified version of https://github.com/jtkim-kaist/ram_modified. The previous version reported about 98% accuracy after 600,000 epoch. In the modified RAM, I have changed the loss function from the activation network to cross entropy loss and achieved similar accuracy with much less epoch, around 90,000. In this code, you can play with code by changing the loss function to triplet or saeimes which are mainly used for training the image search engines.

Reference

Recurrent Models of Visual Attention

http://papers.nips.cc/paper/5542-recurrent-models-of-visual-attention.pdf

https://arxiv.org/pdf/1412.7755.pdf