Pin pytorch requirement

[morriscb]

Pin torch to 1.13.1.

The version of suite2p we are using grabs torch version >=1.7.1. There is a more recent version that seems to be causing the issue. Subsequent suite2p releases (latest is 0.14 ours is 0.10.2) pin the version of torch to a similar value. Attempted to update suite2p to 0.14 and 0.11 but found failures in testing one of which is non trivial. Suggest deploying this for now and creating a ticket to bump the suite2p version to 0.14 so we can unpin in our requirements.

[codecov[bot]]

Codecov Report

Merging #620 (747a8de) into main (51166ba) will increase coverage by 0.14%. The diff coverage is n/a.

@@            Coverage Diff             @@
##             main     #620      +/-   ##
==========================================
+ Coverage   96.13%   96.27%   +0.14%     
==========================================
  Files         105      105              
  Lines        7944     8249     +305     
  Branches      775      775              
==========================================
+ Hits         7637     7942     +305     
  Misses        229      229              
  Partials       78       78              

Flag	Coverage Δ
general_tests	96.27% <ø> (+0.14%)	:arrow_up:

Flags with carried forward coverage won't be shown. Click here to find out more.

see 83 files with indirect coverage changes

[aamster]

I dug into it also. The issue is a conflict with

deepinterpolation @ git+https://github.com/danielsf/deepinterpolation@staging/cleaner/ophys_etl

this installs tensorflow 2.5.0 which is pretty old, and has typing_extensions pinned to ~=3.7 https://github.com/tensorflow/tensorflow/blob/v2.5.0/tensorflow/tools/pip_package/setup.py#L93C5-L93C33

this conflicts with pytorch which is trying to install the latest typing_extensions. You can see that in the build output.

Only typing_extensions >= 3.10 apparently has the ParamSpec https://github.com/optuna/optuna/issues/3558

Another fix would be to instead of using Scott's fork of deepinterpolation (it's not even being used in the main branch ?) you could change it to deepinterpolation @ git+https://github.com/AllenInstitute/deepinterpolation_ai_internal which has tensorflow unpinned. More recent versions of tensorflow have a more relaxed constraint on that library: https://github.com/tensorflow/tensorflow/blob/v2.14.0/tensorflow/tools/pip_package/setup.py#L103

Sorry that was a long comment.

[morriscb]

Not at all. Sounds like a better solution. Just changed the deepinterpolation requirement and seeing what happens.

[morriscb]

Still getting the crash on 3.8. I'll keep the switch from Scott's fork and also pin torch and see if that fixes.

[aamster]

It looked like a different issue? You might to add https://github.com/AllenInstitute/ophys_etl_pipelines/blob/workflow_v2/src/ophys_etl/workflows/Dockerfile#L19 to the Dockerfile