Custom Datasets

To train own models from multiple images, you need to create a dataset first. This dataset is just a file structure so PanSeg knows how to load the training data. It should look like this:

mydataset/
├── train/
│   ├── first.h5
│   └── second.h5
└── val/
    ├── val_one.h5
    └── val_two.h5 

Each h5 file must contain the input image under the raw key, and the segmentation under the label key. Both 2D and 3D data are supported, see Training data layouts.

The recommended way to train new models is through the napari training GUI.