The final paper is named final_report.pdf and is included in the git repo.
To be able to run CycleGAN, Pix2Pix and StarGAN, use: git submodule update --init --recursive
The dataset can be found here with all explanation related to it.
To convert images from raw to the required pre-processed format, use:
python raw_to_adobe_png.py --origin <DNG images folder> --destination <Adobe PNG destination folder>
To resize images, use:
-
For StarGAN and CycleGAN
python resize.py --origin <images source folder> --destination <resized images destination folder> --extention <extention of the source images> --targetSize <size the longest edge should be> -
For Pix2Pix
python resize_pix2pix.py --origin <images source folder> --destination <resized images destination folder> --extention <extention of the source images> --targetSize <size the longest edge should be>
To use the baseline methods model, the file run_baseline.py needs to be opened and adapted. At the lines starting at 82 :
dir_raw = "raw_for_baseline/"
dir_a = "reducedDataSetPNG500/A/"
dir_b = "reducedDataSetPNG500/B/"
dir_c = "reducedDataSetPNG500/C/"
dir_d = "reducedDataSetPNG500/D/"
dir_e = "reducedDataSetPNG500/E/"
Change the "raw_for_baseline/" part for your folder with the raw images, and all other folders by your artists folders.
Finally, at the bottom, change the letter "D" by the artist for whom you want results.
#running + saving of test
for subdir, dirs, files in os.walk(dir_raw):
for file in files:
print(file)
test_plot2(file, "D")
CycleGAN works in two phases: first a train phase, and then a test phase. The code and the different options available are best explained directly in the cycleGAN subfolder.
To create a dataset that is usable by cycleGAN, a root folder has to contain 4 subfolders: trainA, trainB, testA, testB. trainA and trainB contain the images that are used to train the model, respectively in the A image domain and the B image domain. Same for testA and testB.
The exact command ran to get the best results presented in the report is
python train.py --dataroot <dataset root folder> --name <experiment_nam> --model cycle_gan --gpu_ids <0, 1, 2, 3> --verbose --batch_size 4 --num_threads 8 --dataset_mode unaligned --save_epoch_freq 25 --norm instance --preprocess resize_and_crop --load_size 286 --crop_size 256 --gan_mode vanillaTo use Pix2Pix, are needed a folder with two subfolder A/ and B/, which include each train/ and test/ folder.
A and B folders are the domains for images going from A to B, or B to A
To create dataset to use with Pix2Pix model :
python <cycleGAN_folder>/datasets/combine_A_and_B.py --fold_A <images>/A/ --fold_B <images>/B/ --fold_AB <combined_image_folder>
to train the models described in our paper :
For pix2pixA
python <cycleGAN_folder>/train.py --dataroot <combined image folder> --name pix2pixA --gpu 0,1 --model pix2pix
For pix2pixB
python <cycleGAN_folder>/train.py --dataroot <combined_image_folder> --name pix2pixB --gpu 0,1 --n_epochs 500 --n_epochs_decay 500 --preprocess resize_and_crop --save_epoch_freq 20 --save_latest_freq 20000 --model pix2pix --batch_size 64
For pix2pixC
python <cycleGAN_folder>/train.py --dataroot <combined_image_folder> --name pix2pixC --gpu 0,1 --n_epochs 200 --n_epochs_decay 300 --preprocess scale_width_and_crop --save_epoch_freq 20 --save_latest_freq 20000 --model pix2pix --batch_size 64
For pix2pixD
python <cycleGAN_folder>/train.py --dataroot <combined_image_folder> --name pix2pixD --gpu 0,1 --n_epochs 200 --n_epochs_decay 300 --preprocess scale_width_and_crop --save_epoch_freq 20 --save_latest_freq 20000 --model pix2pix --batch_size 64
To test the models :
python <cycleGAN_folder>/test.py --dataroot <combined_image_folder> --name pix2pix{A, B, C, D} --model pix2pix --num_test 1000 --gpu 0,1 --preprocess {scale_width, resize}_and_crop
Dependencies can be found here
To conform to the dataset structure required by StarGAN script, structure the data as described here.
To run the training of the network:
python main.py --mode train --dataset RaFD --rafd_crop_size CROP_SIZE --image_size IMG_SIZE \
--c_dim LABEL_DIM --rafd_image_dir TRAIN_IMG_DIR \
--sample_dir stargan_custom/samples --log_dir stargan_custom/logs \
--model_save_dir stargan_custom/models --result_dir stargan_custom/results
To test the model:
python main.py --mode test --dataset RaFD --rafd_crop_size CROP_SIZE --image_size IMG_SIZE \
--c_dim LABEL_DIM --rafd_image_dir TEST_IMG_DIR \
--sample_dir stargan_custom/samples --log_dir stargan_custom/logs \
--model_save_dir stargan_custom/models --result_dir stargan_custom/results
To compute the Fréchet Inception Distance, run
./fid_score.py <folder1> <folder2>
Where and contain the two sets of images that are going to be compared. The FID will be the distance between the two. Optional argument is --gpu x, with x being the gpu number that the computation is going to be ran on. If this argument is not provided, the computation will run on CPU.
To achieve this from the results of a run of the test.py script from cycleGAN, we have created the reorganise.py script. Is it run as
python reorganise.py <folder>
Where is the <experiment_name>/test_latest/images folder, that contains the output of the test.py script run beforehand. <experiment_name> is the name given in test.py
This will create subfolders with copies of the images, categorised as real_A, real_B, fake_A, fake_B. We then usually called
run ./fid_score.py <experiment_name>/test_latest/images/fake_B <experiment_name>/test_latest/images/real_B
To get the distance between the generated B images and the ground truth B images, ie the retouched images.