src folder includes all the project codes.
All the needed packages are listed in requirements.txt.
train.py is used for training the GAN.
inference.py is used for modify a non-acne image with a pretrained GAN model.
data folder is the location for training data.
The ACNE04 images can be downloaded from here. Please download the Classification.tar and unzip it in the data folder so that the images can be found in data/Classification/JPEGImages.
The extracted facial landmarks are stored in data/Results.
ckpt folder contains a checkpoint of a pretrained Generator.
To train the model, you can use command:
python -m src.train --gpu 0 --exp train \
--modify_loss_weight 1.0 --log_every 20 --epoch 1000 --batch_size 128 \
--Dprestep 20 --Gprestep 0 --Gstep 2
To modify a no-acne image, you can use command:
python -m src.inference \
--image_path 'data/example/levle0_2.jpg' \
--landmark_path 'data/example/levle0_2.pkl' \
--save_dir 'generate' \
--Gckpt 'ckpt/Gen_5600.pth' \
--num_patch 5
It will randomly choose a patch, add acnes to the patch and save the modified full face image and a patch comparison figure to save_dir. With --num_patch X, it will randomly sample X patches and generate X modifications.
To modify a single non-acne image with multiple patches, you can use this command:
python -m src.inference_combined \
--image_name 'levle0_2' \
--save_dir 'generate' \
--Gckpt 'ckpt/Gen_5600.pth' \
--num_patch 15
To modify a random selection of non-acne images (level 0) with multiple patches each, you can use this command:
python -m src.inference_dataset \
--save_dir 'generate/dataset1' \
--Gckpt 'ckpt/Gen_5600.pth' \
--num_patch 10 \
--num_images 10
Lu, Z & Krishnamurthy, S (2021). SkinGAN: Medical image Synthetic data generation using Generative methods