Add basic polygon metrics to nucleus repo

[phil-scale]

Adds IoU, precision, and recall metrics to the Nucleus repository. Creates framework to support more general metrics.

Note: adding the Shapely package also adds a dependency on the geos library, which can be installed in MacOS using brew install geos.

Design doc here: https://docs.google.com/document/d/1uOjZQ9wRMdpsOaYN_w6eTqPvQu0Nwtl5wRNj0HkL0EU/edit#heading=h.raendxyxrl42
ch318548
ch303550

[shortcut-integration]

This pull request has been linked to:

• Shortcut Story #303550: Move public evaluation functions to nucleus-python-client repo
• Shortcut Story #318548: Add info about evaluation criteria

[phil-scale]

Note: this bumps the python requirement to >= 3.7, < 3.11 in order to support scipy - is that ok?

Edit: Scipy, numpy, and Shapely all will be installed only if the python version is >= 3.7, < 3.11. The required python version is still ^3.6

[gatli]

Very clear code to read 🙂

I've requested changes since we need to be sure that Python 3.6 users see a warning message to upgrade. Until then I'm not comfortable rolling this out since I know that python3.6 is default in some widely used distros.

[phil-scale]

Very clear code to read 🙂

I've requested changes since we need to be sure that Python 3.6 users see a warning message to upgrade. Until then I'm not comfortable rolling this out since I know that python3.6 is default in some widely used distros.

Moved the new dependencies to dev only, with the Python requirement applied specifically to those dependencies.

[yixu34]

Do we want to have defaults for the IOU and confidence thresholds? Those seem like things we'd want the user to specify.

[phil-scale]

My thinking is that we can expose information about what defaults are used in Model CI in two ways:

1. Currently, we can keep these defaults in code so that users who want to see what the values are can find them directly here as defaults.
2. Once the UI shows the source code of evaluation functions, we can get rid of the defaults here and instead show them in the source code (e.g. PolygonIOU(enforce_label_match=True, iou_threshold=0.2, confidence_threshold=0.5)).

[phil-scale]

In that case, I'll add a TODO here in polygon_metrics.py to remove the defaults once we can surface these defaults more directly to the user.

[yixu34]

Same here. Also wondering why this is 0.5.

[phil-scale]

The idea is that in Nucleus and the most commonly reported metrics, it's usually 0.5 by default. Also if we're exposing these evaluation functions to users, we would want users to know the default values that are actually used in Model CI.

[gatli]

LGTM 👍 🔥

These tests are failing on master as well so they shouldn't block you. Hopefully I'll have a fix today and then you can simply rebase on top.

[gatli]

Rebased on top of master ...