Address dependency problems caused by M1 Macs

[edwardchalstrey1]

From a scivision workshop this week we encountered the following issues, a couple of notes on problems encountered yesterday when they were going through the example notebook, all on M1 Macs:

• we already mentioned the tensorflow one, but to clarify a bit, if they see the error when doing load_pretrained_model for the "resnet" model, we had to clone the repo https://github.com/alan-turing-institute/scivision_classifier.git then remove tensorflow from requirements.txt, and then pip install . (and install tensorflow some other way e.g. via conda
• some people also had problems with load_pretrained_model for the huggingface classifier. For this, the approach was similar: clone the repo https://github.com/alan-turing-institute/scivision_huggingface, remove all the "transformer" lines from the requirements.txt, then do pip install git+https://github.com/huggingface/transformers (i.e. install from source), and then do pip install . .

[edwardchalstrey1]

I'm not sure if this is something that can be fixed by fixing the version numbers of tensorflow or transformers - we should investigate

some investigation: https://github.com/AoifeHughes/scivision_classifier/blob/requirements-fix/requirements.txt

[edwardchalstrey1]

Since the issue doesn't occur when installing tensorflow via conda, this might be a nice time to think about updating scivision load_pretrained_model function to install dependencies for models that have an environment.yml -> #331

[edwardchalstrey1]

Could be useful: https://towardsdatascience.com/python-conda-environments-for-both-arm64-and-x86-64-on-m1-apple-silicon-147b943ffa55

[ots22]

Closing, but linking the discussion from #444