cvblab/progleason-gan

conditional progressive growing GAN approach for prostatic cancer Gleason Grade patch synthesis

INFORMATION

This is the code implementation for: "ProGleason-GAN: Conditional Progressive Growing: GAN for prostatic cancer Gleason Grade patch synthesis"

TRAINING INSTRUCTIONS

To initiate the training, the following parameters must be set in the config.py file.

Arguments	Description
--START_TRAIN_AT_IMG_SIZE	Start resolution
--CHECKPOINT_GEN	Path for generator checkpoint
--CHECKPOINT_CRITIC	Path for discriminator checkpoint
--RESULTS_DIR	Output directory for the recontructed slides
--PATH_CSV_SICAP	Path with SICAPv2 partition annotations
--PATH_IMAGES_SICAP	Path containing SICAPv2 patches
--DEVICE	DEVICE INFO (cpu or cuda)
--SAVE_MODEL	Flag to allow the training to save the model in the RESULTS_DIR
--LOAD_MODEL	Flag to allow the training to load previous checkpoints
--LEARNING_RATE_GENERATOR	Learning rate for the generator model
--LEARNING_RATE_DISCRIMINATOR	Learning rate for the discriminator model
--BATCH_SIZES	List of batch sizes for each resolution
--CHANNELS_IMG	The number of channels in the input images
--Z_DIM	Size of the input noise vector
--IN_CHANNELS	The number of channels in the generator's input noise vector
--LAMBDA_GP	The weight factor for the gradient penalty term used in the Wasserstein GAN (WGAN) loss
--PROGRESSIVE_EPOCHS	List of training epochs for each resolution
--N_CLASSES	Number of classes in the dataset
--NUM_WORKERS	The number of parallel workers for data loading during training

After that, you only need to call

$ python train.py