Inquiries About Cellpose Model Training and New Functionalities

[WhenFlyWang]

Congratulations on the official publication of Cellpose 3 in Nature Methods! I have read the relevant papers and watched the tutorial videos you published. However, when trying to train my own model using Cellpose, I have encountered a few issues and would greatly appreciate your insights on the following:

When using the human-in-the-loop approach, if the image training cannot be completed in one session, how can I resume training from where I left off after disconnecting? I aim to use as many images as possible to train human-level segmentation results and ensure accuracy.

I am training my model using the GUI, but I seem to be unable to exclude boundary cells from the training. Is there a way to address this?

My dataset includes multiple cell types. I’m wondering if Cellpose will eventually support both classification and segmentation functionalities. While Cellotype has already implemented this, it requires four A100 GPUs, which is a challenging setup for me to meet. Is there a more feasible solution within Cellpose?

I have reviewed the notebook at cellpose/notebooks/run_cellpose_2.ipynb at main · MouseLand/cellpose · GitHub, but it seems there are no curves for evaluating model accuracy or other metrics. Is this something I would need to address on my own?

As English is not my first language, I apologize for any inaccuracies in my expression. Please feel free to ask for clarification if needed. I look forward to your reply and appreciate your time and assistance.

Thank you again!

[Jiadalee]

@WhenFlyWang the notebook has average_precision function for you to evaluate your model. Are you from China? we can discuss so via wechat if you are ok to post your wechat ID.

[WhenFlyWang]

@Jiadalee Yes, I am Chinese, my wechat ID is goodluck23o. Thank you for the answer to average_precision. Do you know any other questions, or should I wait for other kind people?
@

[WhenFlyWang]

@Jiadalee Thank you for your suggestion. I tested the average_precision function, and it returns a result for IoU threshold 0.5. However, I would like to plot the relationship between Average Precision (AP) and different IoU thresholds. Could you guide me on how to implement this? May I ask if you encountered any difficulties while trying to add me on WeChat?