Quick attempt at model interpretation

[metazool]

Considered leaving this on a branch as the results were so inconclusive, but there's some useful utils and improved test coverage in here as well that seem worth a merge - would appreciate eyes on it.

What's in here

• Tiny bit of refactoring for ease of loading the scivision model either with or without its final layer (the classification one)
• Associated tests, increased overall coverage
• A notebook which leans heavily on the Captum tutorial walkthroughs, applying a range of methods to heatmap areas of plankton images which have influenced the predicted class

The visualisations don't look very meaningful to me, wrote a bit about probable reasons why, and what's worth doing next, in #6

To test

• run the python tests as per the README (minimum: export PYTHONPATH=.; py.test
• if you've run through the outstanding PR in https://github.com/NERC-CEH/plankton_ml/pull/5 then you'll have the chromadb index with the images and embeddings in it
• The notebook should be a matter of Run all - it needs the location of the s3 object store set in .env - see also the README

[github-actions[bot]]

☂️ Python Coverage

current status: ✅

Overall Coverage

Lines	Covered	Coverage	Threshold	Status
115	101	88%	0%	🟢

New Files

File	Coverage	Status
cyto_ml/data/s3.py	0%	🟢
cyto_ml/tests/test_image_embeddings.py	100%	🟢
cyto_ml/tests/test_intake_utils.py	100%	🟢
cyto_ml/tests/test_vector_store.py	100%	🟢
TOTAL	75%	🟢

Modified Files

File	Coverage	Status
cyto_ml/data/intake.py	67%	🟢
cyto_ml/data/vectorstore.py	81%	🟢
cyto_ml/models/scivision.py	96%	🟢
TOTAL	81%	🟢

updated for commit: c278b48 by action🐍

[metazool]

Switched this to draft as I have a small set of changes testing fallback approaches which didn't look right to me either!

[metazool]

I've switched this back out of draft status and added a few final touches with non-plankton images and a generic ImageNet set of weights. The outcome's still really open, but there are higher priority things to work on - the fact that I can't easily connect this line of work to anything active on the RSE group discussions forum is a helpful indicator of that!

[metazool]

See also #6

[metazool]

Abandoning this, the outcome wasn't very illuminating, can easily be reconstructed, and the underlying changes (dropping scivision, shifting the project layout) would make reconstruction less effort than merging