The numpy version in the installation of py4dstem-aiml

[PhysXu]

Hi, all,

I encountered an issue when installing the py4dstem with AI/ML functionality by following the steps below:

conda update conda
conda create -n py4dstem-aiml
conda activate py4dstem-aiml
conda install -c conda-forge cudatoolkit=11.0 cudnn=8.1 cupy 
pip install "py4dstem[aiml-cuda]"

The snapshot of this issue:

and I check the version of the installed python, 4dstem and numpy, please see below:

Is the version of numpy too high?

[chiahao3]

Not from the py4DSTEM team so just my 2 cents. It's been a month since the post so you might've solve it already.
I'll suggest restrict your numpy to be < 2.0 during the conda install. Numpy 2.0 was released a couple months back and it will take a while for cupy and scipy to keep up. You might need a newer python as well, probably 3.10 or 3.11. I'm suprised that conda provides a 3.9.19 to you. Overall speaking, I think you just need to find a sweet combination of python, numpy, and cupy. If you're feeling adventerous, you may experiment with different version of cudatoolkit as well.

[sezelt]

I tried this on my machine, and with cupy 13.3.0 and numpy 2.0.2 it is possible to import py4DSTEM. So it seems that whatever incompatibility was present has been resolved by upstream packages.

It does seem odd that py4DSTEM 0.12.8 is the version that it picks. The aiml-cuda option is still in the setup.py, so I'm not sure why newer releases don't get chosen. Is there some step in the packaging process we've been skipping? In any case, this may be the latest version where the aiml features worked, so perhaps it's intentional, but I'm still confused at what step the limit is imposed.

[PhysXu]

I am still confused by this issue.
Even though I installed a very new python (3.11 or 3.12), the pip install "py4dstem[aiml-cuda]" still installed a very old version of py4dstem (0.12.8).

If I want to use the GPU-version of py4dstem, what is the standard steps of the installation?

[sezelt]

The aiml version of py4DSTEM is only necessary if you want to use the FCU-net (https://www.nature.com/articles/s41524-022-00939-9). If you want to use the other GPU features (ptychography, disk detection, ACOM, etc), just install py4DSTEM normally and also specify cupy:
conda create -n py4dstem -c conda-forge py4dstem cupy
Conda should take care of all the other details.