UNet Training

With PanSeg you can train bespoke segmentation models! This is especially useful to get great results on a big dataset for which the build-in models work not perfectly.
First proofread some images from the dataset. Then train a new model on this high-quality data, and run it on the whole dataset.

You can also fine-tune existing models on your data.

Training from a dataset

For training from an dataset stored on disk, create the directories train and val. Your training images must be stored as h5 files in these directories. The h5 files must contain the input image under the raw key, and the segmentation under the label key. The label must be channel-less.

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

Training data layouts

Each h5 file may hold 2D or 3D data; which models can be trained depends on the layout of the raw/label pair:

raw	label	2D model	3D model
yx (2D)	yx (2D)	✓	–
cyx (2D)	yx (2D)	✓	–
zyx (3D)	zyx (3D)	✓, per z-slice	✓
czyx (3D)	zyx (3D)	✓, per z-slice	✓

Any other combination is rejected with a clear error. The label must always be channel-less.

PanSeg resolves the layout of each dataset from its axis_order attribute when present (PanSeg's own h5 exports carry it); otherwise it is inferred from the shapes of the raw/label pair as listed in the table. See Advanced Training for the details.

When training a 2D model on zyx/czyx data, every z-slice is an independent 2D sample. The patch z-extent must then be 1 (the z-stride follows automatically).

Train from GUI

To train from images loaded in the GUI, you need a layer containing the input image and one containing the segmentation. Make sure the quality of the segmentation is as good as possible by using the proofreading tool.

For 2D (yx/cyx) data only 2D training is possible. For volumetric (zyx/czyx) data, the model settings offer a 2D/3D choice (default 3D): 2D trains a 2D model per z-slice as described above and freezes the patch size in z to 1.

Widgets

• Train from: Whether to train from a dataset on disk or from loaded layers in Napari.
• Dataset: Dataset directory. It must contain a `train` and a `val` directory, each containing h5 files. Input/Output keys must be `raw` and `label` respectively.
• In and Out Channels: Number of input and output channels
• Data Resolution: Voxel size in um of the training data. Is initialized correctly from the chosen data if possible.
• Pretrained model: Optionally select an existing model to retrain. Hover over the name to show the model description. Leave empty to create a new model.
• Feature dimensions: Number of feature maps at each level of the encoder. If it's one number, the number of feature maps is given by the geometric progression: f_maps ^ k, k=1,2,3,4 Can't be modified for pretrained models.
• Max iterations: Maximum number of iterations after which the training will be stopped. Stops earlier if the accuracy converges.