simple-unet-2d

simple unet with NeurIPS'19 topoloss

Commands:

• Make sure to populate train.json and test.json with appropriate hyprerparameters

Train:

CUDA_VISIBLE_DEVICES=3 python3 main.py --params ./datalists/DRIVE/train.json

• Ensure crop_size in train.json is divisible by 16

Test/Inference:

CUDA_VISIBLE_DEVICES=4 python3 main.py --params ./datalists/DRIVE/test.json

Compute Evaluation Metrics (Quantitative Results):

python3 compute-eval-metrics.py

Dataset properties:

GT: Foreground should be 255 ; Background should be 0

• First do pretrain (1000-2000 epochs) by setting "topo_weight": 0 in train.json
• Then, load the best model from pretrain and train using topoloss by setting topo_weight to a non-zero value. Change the output_folder and checkpoint_restore in train.json too