yeonghyeon / dcraw

DCRAW: Deep Convolutional Recurrent Attentive Writer

DCRAW: Deep Convolutional Recurrent Attentive Writer

Implementation of DRAW based Auto-Encoder. First of all, feature maps are extracted from Deep Convolutional Neural Network (DCNN). Then, encoding and decoding are recurrently conducted by DRAW. Code after decoding is used to reconstruct original image. (When the purpose is not reconstructing of original image, it is possible to set the other target image.)

Requirements

• Python 3.5.2
• Tensorflow 1.4.0
• Numpy 1.13.3
• Scipy 1.2.0
• Matplotlib 3.0.2

Usage

Just run the following command in the terminal.

$ python run.py // with default options
$ python run.py --attention True // for using attentive read and write.
$ python run.py --help // for confirming the option

Architecture

Whole architecture of the DCRAW.

Architecture of the DRAW.

Results

Following figures are sequentially generated by DCRAW for each epoch.

Without attention

Epoch 0

Epoch 500

Epoch 1000

With attention

Epoch 0

Epoch 500

Epoch 1000

Reference

• DRAW: Deep Recurrent Attentive Writer
• Stacked Convolutional Auto-Encoders for Hierarchical Feature Extraction
• An End-to-End Trainable Neural Network for Image-Based Sequence Recognition and Its Application to Scene Text Recognition

Author

YeongHyeon Park