[BUG] CellPose 3 Training Not Detecting ROI on M1 Mac

[david-an-work]

Describe the bug
I installed CellPose 3 on my computer and have been trying to get it to work. The cyto3 model works well for my images, though not perfect, so I was trying to train my own. However, when I train, it doesn't identify any reasonable ROIs (instead, it finds these seemingly random spots, as pictured below). I'm not sure if this is important, but the training file is saved as a document, not as an executable. Not sure if that makes a difference or not. Also, I installed most things (such as numpy and pyqt) through conda, not pip, since I'm using an environment through anaconda navigator.

To Reproduce
I follow the recommended instructions for training a model: inputting a .tif photo, then running cyto3 with a set diameter, then clicking "train new model." It's after this step, when it moves onto the next image, that there are no relevant ROIs.

Run log
Info:
cellpose version: 3.1.1.1
platform: darwin
python version: 3.10.10
torch version: 2.6.0

What the terminal says when training: 2025-02-22 17:41:48,462 [INFO] 160, train_loss=0.9803, test_loss=0.0000, LR=0.100000, time 67.14s

Screenshots

Would greatly appreciate any advice!

[Aaardaaalaaan]

There has been a similar biviour for Windows user as well, which is documented at the following link: https://github.com/MouseLand/cellpose/issues/1071.
Nevertheless, I believe it might be worthwhile to try running a script (instead of using the GUI) and adjusting the values for the parameters n_epoch and nimg_per_epoch, as discussed in this thread:
https://forum.image.sc/t/cellpose-v2-2-3-constantly-zero-0-roi-after-training/101801/32.
I also found these comparisons of performance using different parameters' values extremely useful:
https://github.com/True-North-Intelligent-Algorithms/tnia-python/blob/main/notebooks/imagesc/2024_09_29_training_cellpose/37_cellpose_training_trials.ipynb

[david-an-work]

Thanks for sharing this! I did increase n_epoch to 1000 once and saw that it was taking a very long time (I aborted at 20 mins+), would it work better if I wasn't working through the GUI?