Reccurent Attention Model with Chainer based on the following paper
arXiv:1406.6247: Recurrent Models of Visual Attention [Volodymyr Mnih+ 2014]

• RAM model difinition file (Chainer)
• script for training the model on MNIST
• script to run the model on MNIST

• hyper-params to get the best accuracy in the paper
• multi-scale glimpse
• models to solve "Translated MNIST" task

Training the model without LSTM takes a day with CPU (reaches 96% accuracy)

Training the model with LSTM takes ??? with CPU
(still searching for the hyper-parameters to get the best accuracy in the paper...)

Python(2 or 3), Chainer, scikit-learn, PIL, tqdm

➜ python train.py   

If you use a GPU, add the option "-g deviceID".
When you use LSTM units in core RNN layer, add the option "--lstm".
(better performance but a little time consuming with LSTMs)

➜ python train.py -g 0 --lstm  

After training, you can get predictions by the trained model.

➜ python predict.py -m ram_wolstm.chainermodel