Models¶
All 2D models originally published in the Cellpose and Omnipose papers use nchan=2. This is because Cellpose defaults are set to train models that use two channels for segmentation (usually cytoplasm and nucleus). Images without a second channel are just padded with 0s. I think most users will train Omnipose on mono-channel images, so now nchan=1 by default.
Tip
Always specify nchan and nclasses when training and evaluating models.
Omnipose used to have a boundary prediction, so nclasses=3 (flow field, distance field, and boundary field in 2D). The current version of Omnipose no longer needs a boundary prediction, so nclasses=2 is the default.
See the table below for named models and their corresponding nchan, nclasses.
Pretrained models¶
model |
|
|
|
|---|---|---|---|
|
2 |
3 |
2 |
|
2 |
3 |
2 |
|
2 |
3 |
2 |
|
2 |
3 |
2 |
|
2 |
3 |
3 |
|
1 |
2 |
2 |
Cellpose models all have nchan=2, nclasses=2, and dim=2 (3D Cellpose uses 2D models to approximate 3D output). This means that if you wanted to, you could train an Omnipose model based on a Cellpose model using these hyperparameters (see Transfer learning).