marcbelmont / gan-image-similarity

InfoGAN inspired neural network trained on zap50k images (using Tensorflow + tf-slim). Intermediate layers of the discriminator network are used to do image similarity.

gan-image-similarity

InfoGAN inspired network trained on images from zap50k. This network is using continuous latent codes + noise as input to the generator. For the image similarity part, I extract the features of last layer before the classification layer of the discriminator for all images and calculate the l2 distances.

Installation

This project is using the following python packages:

tensorflow==0.11.0rc0
scipy==0.18.1

You will also need zap50k dataset:

wget http://vision.cs.utexas.edu/projects/finegrained/utzap50k/ut-zap50k-images.zip

Training

To train the network on all the images: python main.py --logdir=logs/exp1 --batch_size=128 --file_pattern="ut-zap50k-images/*/*/*/*.jpg"

To generate the intermediate features of all the images and calculate similarity for randomly picked images: python main.py --similarity --batch_size=128 --logdir=logs/exp1 --file_pattern="ut-zap50k-images/*/*/*/*.jpg"

To get images from the generator: python main.py --logdir=logs/exp1 --sampledir=samples

Results

For each line, the first image is the seed and the rest are most similar images according to l2 distance of the intermediate features.

Samples from the generator after a coulpe of epochs.