Train a Cellpose model

Use your corrected image masks to fine-tune Cellpose-SAM, then apply the saved model to another image folder. From Home → Tools → Make Masks, open Cellpose Workbench and choose Train.

To follow along, download the six example image/mask pairs and extract them. Set Source to the extracted training folder. Review the supplied cell masks in Make Masks, then follow the settings and Run steps below. The archive’s apply folder holds three more images from wells that are not in the training set; use it to try the trained checkpoint in Apply.

Prepare images and masks

Place training images in one folder and their masks in its masks subfolder:

training/
    field_01.tif
    field_02.tif
    masks/
        field_01.tif
        field_02.tif

Each mask must have the same height and width as its image. Use integer object labels: zero for background and a different positive number for each object. A mask filename can match its image or add _masks before the extension, such as field_01_masks.tif. Keep one matching mask per image.

Use Make Masks to inspect and correct object boundaries before training. Keep separate fields for validation. You can reuse a legacy project containing train/images and train/masks by selecting that project as Source.

Choose inputs and settings

1. Set Source (src) to the training image folder.

2. Leave mask_src empty to use <src>/masks, or select another mask folder.

3. For validation, set test_src to a separate image folder. Its masks belong in <test_src>/masks unless you select test_mask_src.

4. Choose base_model: cpsam starts from Cellpose-SAM; a model-zoo key or checkpoint path continues training an existing model.

5. Enter model_name as a filename prefix. Use save_path to choose an output folder, or leave it empty to save under the training project.

For multichannel images, channels selects up to three zero-based channel indices. For example, [0, 2] selects the first and third channels. Leave channel_axis empty for automatic detection, or set it explicitly when the image layout is ambiguous. Training accepts two-dimensional images with optional channels.

Main training controls

Setting	Default	What it controls
n_epochs	100	Number of training epochs.
batch_size	1	Images per training minibatch. Larger batches require more memory.
learning_rate	0.00001	Size of the model’s training updates.
weight_decay	0.1	Strength of weight regularization during training.
normalize / percentiles	On / [1, 99]	Percentile normalization of each image channel.
min_train_masks	5	Minimum labelled objects in an image included in training.
scale_range	0.5	Variation in image scale during training augmentation.
save_every / save_each	100 / Off	Checkpoint interval and whether periodic checkpoints get separate names.

For a smaller initial run, max_train_images limits the loaded training images. nimg_per_epoch and nimg_test_per_epoch limit how many images are used per epoch. Leave these limits empty to use all available images.

Run and use the output

Click Run and follow progress in the console. When training finishes, the console reports the saved checkpoint path. With the default output location, checkpoints are under <src>/models/cellpose_model/models. With a chosen save_path, they are under <save_path>/models.

Open Apply, select a separate image folder, such as the example’s apply folder, and check that Custom model points to the checkpoint you want. Turn on Save, which is off by default. Inspect a preview before processing the folder. Apply writes label masks into that image folder’s masks subfolder. Use separate labelled images to evaluate segmentation before applying the model to an entire experiment.

The Python entry point spacr.submodules.train_cellpose() accepts the same settings and returns the checkpoint path, training losses and validation losses. Continue with Make Masks to inspect or edit masks, or Mask to use the model in an image pipeline.