# Training baseline model
python train.py --name baseline_model \ # model name for folder + wandb
                --logdir ./ \ # parent folder to write experiments. will create new folder {logdir}/{name}/
                --baseline \ # tell training script to use the baseline dataloader (i.e. real images)
                --l2 50 \ # train using L2 loss for first 50 epochs
                --patch 128 \ # patch size for random crops in training
                --amp \ # auto-mixed precision - should make training faster + allow larger batch size
                --epochs 500 \ # number of epochs
                --epoch_length 200 \ # number of iters per epoch
                --lr 0.001 \ # initial learning rate
                --val_interval 2 \ # save weights and calculate val dice every 2 epochs

# Train fancy Synth model
python train.py --name fancy_synth_model \ # model name for folder + wandb
                --logdir ./ \ # parent folder to write experiments. will create new folder {logdir}/{name}/
                --mbhealthy \ # tell training script to use MultiBrain healthy labels
                --fade \ # use INU fields to mimic penumbra within lesion masks
                --lesion_weight 2 \ # upweight lesion class by 2 relative to healthy tissue in seg loss
                --l2 50 \ # train using L2 loss for first 50 epochs
                --patch 128 \ # patch size for random crops in training
                --amp \ # auto-mixed precision - should make training faster + allow larger batch size
                --epochs 500 \ # number of epochs
                --epoch_length 200 \ # number of iters per epoch
                --lr 0.001 \ # initial learning rate
                --val_interval 2 \ # save weights and calculate val dice every 2 epochs

# Test (infer) fancy Synth model on new data - this does not require ground truth data
python test.py --weights ./fancy_synth_model/checkpoint.pt \ # path to our model trained in previous script
               --tta \ # use test-time augmentation when predicting
               --mb \ # predict multi-brain labels (this determines the out channels in the model so please set this if used in training, even if you don't want healthy labels)
               --patch 128 \ # same size we used in training - will use Sliding Window Inferer
               --savedir /my/output/data/folder/ \ # folder to write predictions to
               --files /my/input/data/folder/*.nii.gz \ # regex to generate list of files. can also be a path to a .txt containing a single column of file paths