VolPy Mask_RCNN does not work in CaImAn 1.9.9 or later

[KawashimaLab]

• Tell us a bit about your setup:

1. Linux CentOS7

2. Python 3.9.13, cuda 11, Tesla V100

3. Jupyter Lab (remote)

4. [demo_pipeline_voltage_imaging.py]

5. CaImAn version : 1.9.9 or later

6. mamba and pip

• Describe the issue that you are experiencing

We have been using VolPy and it has been working great with Mask_RCNN. We recently tried to upgrade CaImAn to newer versions and met trouble. The installation and most of the functionalities was OK. But MASK_RCNN module does not recognize neurons even in the demo code (0 neurons recognized).

There is no clear error code returned in the script. The code does not stop at MASK_RCNN part.

We reverted the version to CaImAN to 1.9.8 (Tensorflow 2.7.0) and it works. CaImAn 1.9.9 and Tensorflow 2.10.0 does not work. We tried to revert Tensorflow to 2.7.0, but couldn't figure out a way in caiman environment.

Thanks for your help!

Takashi

[j-friedrich]

Thanks Takashi for letting us know. I tried, and ran into the same issue using a fresh clone of the current master branch. Changing the environment.yml to force python==3.9 I ended up with tensorflow 2.6 and it works. Oddly, updating tensorflow from 2.6 to 2.7 breaks it for me, whereas you say 2.7 works for you. Don't have a fix yet for this curious issue.

[pgunn]

@caichangjia This might be worth digging into

[caichangjia]

I will look into that

[KawashimaLab]

Thank you Johannes and Changjia! I just double-checked the version. In the current working CaImAn 1.9.8, Python is 3.9.13. Both Tensorflow and Keras are 2.7.0.

[caichangjia]

Okay, I finally figure out the problem is because tensorflow changes function to load weights for the neural network. Will ask for pull request soon. @KawashimaLab

[KawashimaLab]

Thank you so much!