Fastapi model serving

[metazool]

A rough prototype for an API which returns embeddings and classification from a choice of endpoints, each for a different model.

See #52 for notes on the thinking around adding this to one or more pipelines. Definitely an unsophisticated prototype, reusing existing functions plus small parts of https://github.com/alan-turing-institute/ViT-LASNet

To test

Note this assumes that the model weights for the new Turing models #45 are not available where the test is running!

py.test test/test_model_api.py

Run the API and send it an image URL, get back embeddings

python src/cyto_ml/models/api.py
curl -d url="https://cdn.oceanservice.noaa.gov/oceanserviceprod/facts/nasa-copepod.jpg" http://localhost:8081/resnet50/

To run with the Resnet18 weights, see this note https://github.com/alan-turing-institute/ViT-LASNet/tree/main/test and the linked issue here alan-turing-institute/ViT-LASNet#2 - for now a manual stage of downloading them from Google Drive and copying into ./data/weights - option is to package this as a container

[metazool]

This is still really a draft but I'd like to merge it and revise from there; I've got another branch taken from this one with a lot of aligned work on it which I'd like to request critical review on...