Models

from cellpose import models

Each model will be downloaded automatically to your models.MODELS_DIR (see Installation instructions for more details on MODELS_DIR). You can also directly download a model by going to the URL, e.g.:

https://www.cellpose.org/models/MODEL_NAME

All built-in models were trained with the ROIs resized to a diameter of 30.0 (diam_mean = 30), except the ‘nuclei’ model which was trained with a diameter of 17.0 (diam_mean = 17). User-trained models will be trained with the same diam_mean as the model they are initalized with. The models will internally take care of rescaling the images given a user-provided diameter (or with the diameter from auto-diameter estimation in full models).

Full built-in models

These models have Cellpose model weights and a size model. This means you can run with diameter=0 or --diameter 0 and the model can estimate the ROI size. However, we recommend that you set the diameter for your ROIs rather than having Cellpose guess the diameter.

These models can be loaded and used in the notebook with models.Cellpose(model_type='cyto3') or in the command line with python -m cellpose --pretrained_model cyto3.

We have a nuclei model and a super-generalist cyto3 model. There are also two older models, cyto, which is trained on only the Cellpose training set, and cyto2, which is also trained on user-submitted images.

FYI we are no longer using the 4 different versions and --net_avg is deprecated.

Cytoplasm model ('cyto3', 'cyto2', 'cyto')

The cytoplasm models in cellpose are trained on two-channel images, where the first channel is the channel to segment, and the second channel is an optional nuclear channel. Here are the options for each: 1. 0=grayscale, 1=red, 2=green, 3=blue 2. 0=None (will set to zero), 1=red, 2=green, 3=blue

Set channels to a list with each of these elements, e.g. channels = [0,0] if you want to segment cells in grayscale or for single channel images, or channels = [2,3] if you green cells with blue nuclei.

The ‘cyto3’ model is trained on 9 datasets, see the Cellpose3 paper for more details.

These models are downloadable from the website with MODEL_NAME: cytotorch_0, cyto2torch_0, and cyto3. The size models are size_cytotorch_0.npy, size_cyto2torch_0.npy, and size_cyto3.npy.

Nucleus model (‘nuclei’)

The nuclear model in cellpose is trained on two-channel images, where the first channel is the channel to segment, and the second channel is always set to an array of zeros. Therefore set the first channel as 0=grayscale, 1=red, 2=green, 3=blue; and set the second channel to zero, e.g. channels = [0,0] if you want to segment nuclei in grayscale or for single channel images, or channels = [3,0] if you want to segment blue nuclei.

The ‘nuclei’ model is downloadable from the website with MODEL_NAME nucleitorch_0, and size model size_nucleitorch_0.npy.

Other built-in models

The main built-in models are dataset-specific models trained on one of the 9 datasets in the Cellpose3 paper. These models do not have a size model. If the diameter is set to 0.0, then the model uses the default diam_mean for the diameter (30.0).

These models can be loaded and used in the notebook with e.g. models.CellposeModel(model_type='tissuenet_cp3') or models.CellposeModel(model_type='livecell_cp3'), or in the command line with python -m cellpose --pretrained_model tissuenet_cp3.

The dataset-specific models were trained on the training images from the following datasets:
There are also legacy models which remain on the website with MODEL_NAME:

  • Cellpose2 style-specific models: CP, CPx, TN1, TN2, TN3, LC1, LC2, LC3, LC4, style_choice.npy.

  • Cellpose2 general models: tissuenet, livecell, general.

  • Cellpose1 mxnet models: cyto_0, nuclei_0, size_cyto_0.npy, size_nuclei_0.npy.

We see no improvement with transformer models in the Cellpose3 paper and the Neurips challenge response, but the models are available as transformer_cp3 and neurips_cellpose_transformer. These model take three channels as input. The u-net based Neurips challenge model is available as neurips_cellpose_default and also takes three channels as input.

User-trained models

By default, models are trained with the ROIs resized to a diameter of 30.0 (diam_mean = 30) – this is necessary if you want to start from a pretrained cellpose model. If you want to use a different diameter and use pretraining, we recommend performing training yourself on the cellpose dataset with that diameter so the model learns objects at that size. All user-trained models will save the diam_mean so it will be loaded automatically along with the model weights.

Each model also saves the diam_labels which is the mean diameter of the ROIs in the training images. This value is auto-loaded into the GUI for use with the model, or will be used if the diameter is 0 (diameter=0 or --diameter 0).

These models can be loaded and used in the notebook with e.g. models.CellposeModel(model_type='name_in_gui') or with the full path models.CellposeModel(pretrained_model='/full/path/to/model') . If you trained in the GUI, you can automatically use the model_type argument. If you trained in the command line, you need to first add the model to the cellpose path either in the GUI in the Models menu, or using the command line: python -m cellpose --add_model /full/path/to/model.

Or these models can be used in the command line with python -m cellpose --pretrained_model name_in_gui or python -m cellpose --pretrained_model /full/path/to/model.

Finding models on BioImage.IO

BioImage.IO is a repository for sharing AI models, datasets and tools for bioimage analysis. You may look for Cellpose models on BioImage.IO Model Zoo by searching for the tag cellpose. To download a model, click on the model card, click the download icon, and choose “Download by Weight Format” - “Pytorch State Dict”.

Sharing models on BioImage.IO

You can also share your trained Cellpose models on the BioImage.IO Model Zoo. To do this, you need to export your model in the BioImage.IO format using cellpose/export.py and then upload the packaged model to the Model Zoo.

Detailed steps:

  1. Train a Cellpose model and check if it works well on your data.

  2. Create an environment python -m pip install 'cellpose[bioimageio]' or 'cellpose[all]' if you haven’t already. Note that most users installed 'cellpose[gui]' without the bioimageio packages.

  3. Export the model using export.py script. Use python export.py --help to see the usage, or check the example in its docstring. In short, you need to name your models, specify if the model runs on cytoplasm/nuclei/both, and provide:

    1. a model filepath,

    2. a README.md filepath,

    3. a cover image filepath(s),

    4. a short description string,

    5. a license name like MIT,

    6. a link to your GitHub repo (or Cellpose repo),

    7. information about the authors and what to cite,

    8. tags including cellpose, 2d and 3d (Cellpose models handle both).

  4. If you are updating a uploaded model, you should also specify the model ID and icon. Don’t forget to increment the version number.

  5. Go to BioImage.IO, click “Upload”, and follow the instructions there.