Exploiting the Potential of Standard Convolutional Autoencoders for Image Restoration by Evolutionary Search

This repository contains the code for the following paper:

Masanori Suganuma, Mete Ozay, and Takayuki Okatani, "Exploiting the Potential of Standard Convolutional Autoencoders for Image Restoration by Evolutionary Search," International Conference on Machine Learning (ICML), 2018. [PDF] [arXiv]

If you find this work useful in your research, please cite:

@inproceedings{suganumaICML2018,
    Author = {M. Suganuma and M. Ozay and T. Okatani},
    Title = {Exploiting the Potential of Standard Convolutional Autoencoders for Image Restoration by Evolutionary Search},
    Booktitle = {ICML},
    Year = {2018}
}

Sample results on inpainting tasks:

example

Sample results on denoising tasks:

example

Requirement

Usage

Run the architecture search (denoising)

python exp_main.py -i

Run the architecture search (inpainting)

python exp_main.py -i -mask center

When you specify the -i option, an initial individual consists of a single convolution layer and a single deconvolution layer. To choose inpainting tasks, please specify the -mask option (center, pixel, half).

When you use the multiple GPUs, please specify the -g option (default:1):

python exp_main.py -g 2

You can set the number of offsprings with the -l option (default:2):

python exp_main.py -l 4

After the execution, the files, network_info.pickle and log_cgp.txt will be generated. The file network_info.pickle contains the information for Cartegian genetic programming (CGP) and log_cgp.txt contains the log of the optimization and discovered CAE architecture's genotype lists.

Some parameters (e.g., # rows and columns of CGP, and # epochs) can easily change by modifying the arguments in the script exp_main.py.

Re-training (fine-tuning)

To re-train the discovered architecture:

python exp_main.py -m retrain

Re-evolution

To re-start the evolution:

python exp_main.py -m reevolution

Dataset

The CelebA dataset is available here.

The cars dataset is available here.

The SVHN dataset is available here.

The BSDS500 dataset is available here.

To create the training, validation, and test sets of the CelebA dataset, please download "Align&Cropped Images" and save images to ./celebA_org, and download "list_landmarks_landmarks_align_celeba.txt" from here. Then please run a following code:

python create_data.py