SkySegmentation

Dataset conversion

Dataset need to be in LabelMe format. Images and JSON are in one directory. To convert the images to a supported format, run create_dataset.py.

This requires different conda env witr LabelMe installed. VS Code seems to have problems running another conda env in the terminal, so make sure to run it from an outside terminal.

python create_dataset.py    path/to/labelme     \
                            path/to/save        \
                            path/to/labels.txt  

Training

Train the model using train.py.

Available architectures are:

• pfn
• unet
• deeplab

To train run:

python train.py     --data_dir    path/to/dataset         \
                    --save_dir    path/to/save/model      \
                    --num_classes 5