Implementation and comparison of different neural network models for image generation in low data setting with objective high FOM and generation diversity. The Denoising Difusion model with a UNet is the clear winner. The comparison focuses on different entropy based diversity metrics.
Download data with
wget --no-check-certificate 'https://drive.google.com/file/d/1-ZTBtsQsQ6Sk2VR-zNRhTDjlmzc6bsd2/view?usp=sharing' -O images.zip
unzip images.zip
path = ~/scratch/nanophoto/topoptim/fulloptim/
mkdir -R path
mv images.npy path
To train models, use train.py or to run on slurm use sbatch setoff.sh main.py.
main.py uses hydra. Default configs can be overridden by
python train.py model=wgan
To compare several trained models, use photo_gen/evaluation/compare_models.py with argument evaluation.datasets=datasets.yaml and datasets.yaml a config file containing for each dataset
datasets:
- name: my_dataset
- path: path/to/dataset/images.npy
- n_samples: number_of_training_samples
The script will save plots in output/<time>/metric_plots.