Exploring Latent Dimensions of Crowd-sourced Creativity

(Accepted at Machine Learning for Creativity and Design (NeurIPS Workshop), 2021)

Our code is based on the official GANalyze repository

Overview

• Requirements and Installation
• Training
• Testing
• Reference

Requirements and Installation

We have tested our framework and produced the official results with the libraries and their corresponding versions below. However, newer versions will very likely work without a problem. If you experience a problem, don't hesitate to open an issue.

• CUDA 10.2
• PyTorch >= 1.6.0 and torchvision >= 0.7.0
• efficientnet-pytorch
• numpy, scipy, PIL

One possible way of installing the libraries on a linux server is as follows:

# CUDA 10.2
pip install torch==1.6.0 torchvision==0.7.0

# Efficientnet Pytorch
pip install efficientnet-pytorch==0.7.0

Finally, to clone this repo, run:

git clone https://github.com/catlab-team/latentcreative.git

Training

Before training, pretrained BigGAN models should be downloaded using the script below:

cd GANalyze
sh download_pretrained.sh

After download, you can use the training script below to train the proposed framework:

python train_pytorch.py --assessor_path ../CreativeClassifier/models/best.pth \
--experiment_name experiment --artbreeder_class 0 --class_direction 1 --num_samples 400000 \
--clipped_step_size 1 --transformer AdaptiveDirectionZAdaptiveDirectionY_noise_nonlinear class_reg=1000,noise_dim=8 \
--batch_size 8 --learning_rate 0.0001 --multiway_linear 0

Testing

After training, you can use the testing script below to produce images.

python test_artbreed.py --checkpoint 400000 --checkpoint_dir Checkpoints/experiment/biggan512_None/creative_classifier_best.pth/AdaptiveDirectionZAdaptiveDirectionY_noise_nonlinear_class_reg=1000,noise_dim=8/artbreeder_class:0/[HASH]

Reference

BibTeX reference of our work is as follows:

@article{kocasari2021exploring,
  title={Exploring Latent Dimensions of Crowd-sourced Creativity},
  author={Kocasari, Umut and Bag, Alperen and Atici, Efehan and Yanardag, Pinar},
  journal={arXiv preprint arXiv:2112.06978},
  year={2021}
}