lior1990 / figr

Few-shot Image Generation with Reptile

FIGR

Few-shot Image Generation with Reptile

The gist of this project is that the Reptile meta-learning algorithm is compatible with the GAN setup, unlike the more popular MAML meta-learning algorithm. We train GAN's for few-shot image generation on previously unseen classes on images through this approach.

FIGR-8

The project also includes a new dataset for few-shot image generation, FIGR-8. A dataset containing 18,409 classes of at least 8 images each for a totla of 1,548,944 images. It can be found here https://github.com/marcdemers/FIGR-8 and here bit.ly/FIGR-8 and is downloaded automatically when running this code like so:

$ python train.py --dataset FIGR8

Installation

$ git clone https://github.com/OctThe16th/FIGR.git
$ cd FIGR
$ pip install -r requirements.txt

Usage

$ python train.py --dataset Mnist & tensorboard --logdir Runs/

For the different command line options, simply write:

$ python train.py --help

If you use this code for your own projects, please consider citing the following paper:

@article{FIGR2019,
author = {Louis Clouâtre and Marc Demers},
title = {FIGR: Few-shot Image Generation with Reptile},
journal = {CoRR},
volume = {abs/1901.02199},
year = 2019,
ee = {http://arxiv.org/abs/1901.02199},
month = jan,
archiveprefix = “arXiv”,
number = “1901.02199v1”,
eprint = “1901.02199v1”,
primaryclass = “cs.CV”,
nonrefereed = “true”
}