Reccurrent Attention Model

Introduction

This repo is an implementation of Reccurrent Attention Model (RAM) from Recurrent Models of Visual Attention.

I tested the model on $28 \times 28$ MNIST dataset and got the following results:

nums_glimpses	error rate	error rate in the paper
2	2.34%	3.79%
7	1.93%	1.07%

Requirements

• Python 3.6+
• PyTorch 0.4

Usage

The code has been tested in a CPU-only environment.

See detail in train.py for hyperparameters setting and run the following command with arguments:

python train.py --epochs 30 ...