GhostNet Segmentation

Example commands:

Train

• Every existing hyperparameters are in the configuration files. After a model is trained, a ckpt file will be saved under the weights folder.

python main.py --config ./configs/ghostseg_de_plain_111824.yaml

Validation

• Validation will not output npz file for post process. Just outputs evaluation metrics.

python main.py --config ./configs/ghostseg_de_plain_111824.yaml --evaluate ./weights/GhostSeg_DE/Plain/GhostSeg_DE_Plain_111824-0-epoch\=2514-valid_mean_IoU\=0.77.ckpt --val

Test

• Test will save an npz file that has images, labels, logits, and all relevent data for post-processing.

python main.py --config ./configs/ghostseg_de_plain_111824.yaml --evaluate ./weights/GhostSeg_DE/Plain/GhostSeg_DE_Plain_111824-0-epoch\=2514-valid_mean_IoU\=0.77.ckpt

Prediction

• Prediction only loads a data root and outputs polygon predictions, center lines of the polygons, and centroids of these polygons. All coordinates are in percentages of image widths and heights.

 python main.py --config ./configs/ghostseg_de_plain_111824.yaml --evaluate ./weights/GhostSeg_DE/Plain/GhostSeg_DE_Plain_111824-0-epoch\=2514-valid_mean_IoU\=0.77.ckpt --predict --predict_root ~/datasets/wwf_seg/Germany/val/