[FEATURE] Tensorflow Compatibility
Closed this issue · 2 comments
Is your feature request related to a problem? Please describe.
Right now we using a rather unusual setup for MacOS:
conda create -n mirar
conda activate mirar
conda config --env --set subdir osx-64
conda install python=3.11
pip install --upgrade pip
The reason was that some astronomy software was not available in the standard arm64 mode.
However, I have tragically discovered that tensorflow is really really hard to install for such a setup. Can we transition to a standard installation? Otherwise I am not sure how to use tensorflow with our current set up.
Issues with arm64:
- pyfftw not available. This can be overcome by using pip rather than poetry.
- The following packages are not available on arm64 via conda:
- astromatic-scamp
- astromatic-source-extractor
- astromatic-psfex
- astromatic-swarp
I'm really not sure what the best thing to do here is.
- We could abandon tensorflow in favour of some simpler ML set up
- We could require users install astromatic software via brew? rather than conda
- We could run mirar on MacOS via Docker somehow (cc #790)
My thoughts: nearly every mirar user will need source-extractor. Most will not need ML from tensorflow. So I'd favour 1 or 3.
I had completely forgotten but apparently I did try making an issue for this in the astromatic Github page (conda-forge/astromatic-source-extractor-feedstock#9) one year ago.
My working assumption is that this is not likely to be fixed any time soon.
Possible solution: pytorch