/DRAW-chainer

DRAW implementation using chainer

Primary LanguagePython

DRAW-chainer

Reproduction of Deep Recurrent Attentive Writer (Recurrent VAE with attention) using chainer

todo

  • Minimum implementation
  • Show result
  • Add visualizer
  • Complete description

Original Paper

"DRAW: A Recurrent Neural Network For Image Generation"

Develop environment

Python 3.5.2

(requirement libraries using pip)

chainer 1.16.0
scipy
pillow
sklearn (only for downloading mnist)

How to run

[no attention]

python train_draw_noattention.py save_path [options]

[with attention]

python train_draw_withattention.py save_path [options]

(options example)
with gpu, sequence time length = 32

python train_draw_withattention.py save_path -g 0 -T 32

if you want to see more, use

python train_draw_withattention.py --help

Architecture

DRAWarchitecture

results

without attention (reconstruction of train data during learning)

noA
epoch: 0 -> 20

with attention (reconstruction of train data during learning)

wA
epoch: 0 -> 20