Galaxy Zoo generation using DDPM

Installation

conda env create -f environment.yml

Dataset

Galaxy Zoo dataset is used

Training

Training of the classifier (used for guidance)

To run the training of the classifier, first fill the config file. Example of the detailed config is available configs/classifier.yaml

Then run:

python train_classifier.py --config=<path to config>

If you want to automatically select the batch size, add --auto_bs flag. If you want to automatically select learning rate, add --auto_lr flag.

Training of the conditional DDPM generator

To run the training of the classifier, first fill the config file. Example of the detailed config is available configs/generator.yaml

Then run

python train_generator.py --config=<path to config>

If you want to automatically select the batch size, add --auto_bs flag. If you want to automatically select learning rate, add --auto_lr flag.

Image generation

Generate images using classifier guidance

Run

python classifier_sample.py --config_gen=<generator config> \
                            --config_clas=<classifier config> \
                            --ckpt_gen=<generator ckpt> \
                            --ckpt_clas=<classifier ckpt> \
                            --classifier_scale=3 \
                            --batch_size=16 \
                            --output=<output path> \
                            --timestep_respacing=250

Generate images using classifier-free guidance

Run

python classifier_free_samlpe.py --config=<generator config> \
                                 --ckpt=<generator ckpt> \
                                 --output=<path where to save generated images> \
                                 --batch_size=16 \
                                 --guidance_scale=3 \
                                 --timestep_respacing=250

Evaluation

To run the generated images evaluation

python evaluate.py --path_data=<path to real images directory> \
                   --path_labels=<path to csv with labels> \
                    --path_gen_images=<path to .npy file with generated images> \ 
                    --path_gen_labels=<path to .npy file with labels used to generate images>

vkinakh / galaxy-zoo-generation-diffusion