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

• RAM model on Chainer
• script to train RAM and classify digits with RAM
• including Original MNIST, Translated MNIST, and Cluttered Translated MNIST tasks
• 1.13% error on 28x28 MNIST (RAM, 6 glimpses, 8x8, 1 scale)
• 1.58% error on 60x60 Translated MNIST (RAM, 6 glimpses, 12x12, 3 scales)
• 5.30% error on 60x60 Cluttered Translated MNIST (RAM, 6 glimpses, 12x12, 3 scales)

glimpses and output probabilities at each time step

• Original MNIST

• Translated MNIST

• Cluttered Translated MNIST

Python(2 or 3), Chainer, PIL, matplotlib, tqdm

train.py: optimizes weights of a RAM model and ouputs learned weights to *.chainermodel file every 100 epoch

• with "--original" for 28x28 original MNIST task, "--translated" for 60x60 translated MNIST, and "--cluttered" for 60x60 cluttered translated MNIST
• training for Cluttered Translated MNIST will take around half a day with CPU
• train on CPU with default setting or GPU with "-g YOUR_GPU_DEVICE_ID"

➜ python train.py --original  

infer.py: plot result of inference by a trained RAM model (result will show up with your matplotlib's backend)

➜ python infer.py --original -m models/ram_original_epoch800.chainermodel