Can't I use datastore and trinnet function when training pre-trained cellpose model?

Hello!

I am trying to use the cellpose model to track irregularly shaped cells.

But before that, I want to fine-tune the pre-trained model because it is not very accurate.

I've read the matlab documentation on this, and it looks like I need to prepare the images and ground truth for training in a folder in advance.

I only have 14 images, so I would like to train it with real-time augmentation and use ground truth to terminate early.

When I used other models before, I used the trainnet function in matlab, which was easy to adapt.

However, there doesn't seem to be an option for this in the trainCellpose function.

Do I have no choice but to create my own training function?

Answer 1

[Aneela]

Hi Junhyeon,

“trainCellpose” does not directly support for real-time augmentation. Having a pre-augmented dataset is essential.

• To apply augmentations on-the-fly during training, you might consider implementing a custom datastore by subclassing “matlab.io.Datastore”.

“trainCellpose” does not support early stopping in a way that “trainnet” supports. To achieve early stopping using “trainCellpose”, here’s a workaround:

• Ensure validation dataset is separate from training dataset.
• Train the model for a small number of epochs at a time, manually checking the performance on the validation set after each iteration.
• If the validation performance does not improve over successive iterations, stop the training process.

For more information, please refer to the following MathWorks documentation links:

“trainCellpose” - https://www.mathworks.com/help/medical-imaging/ref/cellpose.traincellpose.html

“matlab.io.Datastore” - https://www.mathworks.com/help/matlab/ref/matlab.io.datastore-class.html

[Junhyeon]

Thank you for your reply. I'll try the methods you suggested.