Test and review models in catalogue

[ots22]

Aim is to have all models in https://github.com/alan-turing-institute/scivision/blob/main/scivision/catalog/data/models.json working when they're loaded (or an indication that they are not supposed to, e.g. via "scivision_usable": false).

Start by working through the scivision examples gallery:

1. Clone each of the repos in https://github.com/scivision-gallery
2. Try to run each notebook according to the instructions
3. Fix any easy to fix problems or update the documentation with extra steps needed
   • Consider updating dependencies (in environment.yml) - to the latest scivision package in particular
   • General 'code review' suggestions and clarity of the text
4. Open issues (in those repos) for the harder problems and paste error messages/ screenshots etc: (example issue)
5. Try the same for the catalogue models that don't have a corresponding notebook example

• [x] https://github.com/scivision-gallery/plankton-classification
• [x] https://github.com/scivision-gallery/scivision-basic-usage-examples
• [x] https://github.com/scivision-gallery/plant-phenotyping-classification
• [ ] https://github.com/scivision-gallery/cryoEM-object-detection
• [x] https://github.com/scivision-gallery/tree-crown-detection
• [ ] https://github.com/scivision-gallery/coastalveg-edge-detection
• [x] https://github.com/scivision-gallery/connections-workshop-examples

(Note: transferred from private https://github.com/alan-turing-institute/DS4S-project-management/issues/83)

Also consider the full set of models in the model catalog (note some typos need to be corrected to get some of these to run):

• [x] https://sci.vision/#/model/stardist - runs with master not main
• [x] https://sci.vision/#/model/PlantCV - runs with main not master
• [x] https://sci.vision/#/model/greenotyper - runs as is
• [x] https://sci.vision/#/model/mapreader-plant - runs with main not master
• [x] https://sci.vision/#/model/resnet50-plankton - runs with main not master
• [ ] https://sci.vision/#/model/vedge-detector - main not master, but doesn't run
• [x] https://sci.vision/#/model/image-classifiers - main not master, but doesn't run
• [x] https://sci.vision/#/model/huggingface-classifiers
• [ ] https://sci.vision/#/model/huggingface-segmentation - needs torch
• [ ] https://sci.vision/#/model/huggingface-object-detection
• [ ] https://sci.vision/#/model/detectreeRGB-forest - needs torch
• [x] https://sci.vision/#/model/flower-classification-model - doesn't run through `load pretrained model'
• [x] https://sci.vision/#/model/butterfly-classification-model - doesn't run through `load pretrained model'

[ots22]

A note to include a list of models here, grouped into 'works with pip', 'works with documented steps', 'broken'.

@IFenton

[ots22]

Also discussed earlier:

• the gallery examples should all include a environment.yml (or requirements.txt, but might be some advantage to these all being consistent).
• there should be a specification of all packages and their complete versions in the environment somewhere, including secondary dependencies, and including scivision itself. This could be in the files above, or somewhere else (e.g. make an 'environment-lock.yml', and keep 'environment.yml' for describing the immediate dependencies with ranged versions).

[IFenton]

A note to include a list of models here, grouped into 'works with pip', 'works with documented steps', 'broken'.

@IFenton

These have now been added in the initial comment

[IFenton]

I've corrected the branch names where master was used rather than main and vice versa (see #585)

[ots22]

huggingface-classifiers

• dependencies in requirements.txt can be trimmed
• pip installing safetensors (dependency of transformers) attempts to use the source distribution, rather than the wheel (there seem to be several mac wheels on https://pypi.org/project/safetensors/#files) - fails when it can't find a rust compiler

Summary of notes below: conda default channel on mac only supports wheels tagged up to macosx_10, but there isn't one on PyPI for safetensors, making installation more difficult.

Suggest we don't include this example in the introductory notebook (since a chance it doesn't work smoothly first time), and maybe recommend using python from conda-forge in installation/setup instructions.

To-do before we can tick this one off:

• [x] update requirements.txt (I'm ignoring any environment.yml in any model repo)

---

Relevant docs https://peps.python.org/pep-0425/ https://packaging.python.org/en/latest/specifications/platform-compatibility-tags/

Attempting to install directly from the wheel:

$ pip install safetensors-0.3.3-cp311-cp311-macosx_12_0_universal2.whl 
ERROR: safetensors-0.3.3-cp311-cp311-macosx_12_0_universal2.whl is not a supported wheel on this platform.

Running pip debug --verbose shows my platform supports various cp311-cp311-macosx_10_*_universal2 (but no macos_12_0)

Maybe-relevant issues:

• https://github.com/pypa/packaging/issues/435
• https://github.com/pypa/pip/issues/10670
• https://github.com/pypa/pip/issues/11618

~

• Is this because pip/python are installed using conda?
  • looks like it could be: https://stackoverflow.com/questions/76105751/why-does-the-python-installed-by-condas-defaults-report-the-wrong-mac-platform
• There is no python 3.11 wheel on https://pypi.org/project/safetensors/#files that also has macosx 10.*\

[ots22]

Will mark this one one done

[ots22]

image-classifiers too (which also had some required versions relaxed)