EDRAM

Enriched Deep Recurrent Visual Attention Model

The repository contains implementation of Enriched Deep Recurrent Visual Attention Model for MNIST Cluttered (https://github.com/deepmind/mnist-cluttered). The original paper in proceeding to the WACV 2017 conference.

Dependencies

• Blocks follow the install instructions. This will install all the other dependencies for you (Theano, Fuel, etc.).
• MNIST Cluttered
• Bokeh 0.8.1+

Dataset

To be able to train the model you need to build MNSIT Cluttered dataset by running script from the fork (https://github.com/ablavatski/mnist-cluttered)

luajit mnist_cluttered_gen.lua
python mnist_cluttered.py --path

and setup the location of your data directory:

export FUEL_DATA_PATH=/home/user/data

Training

To train the model with a basic setup you need to run the script

cd edram
python train_mnist_cluttered.py

Usually it takes around 60 epoch to train the model up to accuracy 1.24% of error. After training in the folder you can find

• a pickle of the best model
• a pickle of the log
• the txt log file of the training