Add ComponentGraph class

[eccabay]

Closes #1277 (design doc here)

[codecov]

Codecov Report

Merging #1415 (f7178a1) into main (06babaa) will increase coverage by 0.1%.
The diff coverage is 100.0%.

@@            Coverage Diff            @@
##             main    #1415     +/-   ##
=========================================
+ Coverage   100.0%   100.0%   +0.1%     
=========================================
  Files         230      232      +2     
  Lines       15831    16430    +599     
=========================================
+ Hits        15823    16422    +599     
  Misses          8        8             

Impacted Files	Coverage Δ
evalml/pipelines/__init__.py	100.0% <100.0%> (ø)
evalml/pipelines/component_graph.py	100.0% <100.0%> (ø)
...valml/tests/pipeline_tests/test_component_graph.py	100.0% <100.0%> (ø)
evalml/tests/pipeline_tests/test_graphs.py	100.0% <100.0%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 06babaa...f7178a1. Read the comment docs.

[dsherry]

@eccabay this is huge!! Nice going, this API is looking great. Very excited to get it into the pipelines! I really like the way we use that recompute method, and I think that from_list method is gonna be super helpful. The try-first-and-fail logic in graph is neat.

I left some comments and suggestions. I didn't dig into the tests yet, will do so on next pass. Some things to call out:

• I think we should have public methods for fitting and predicting. They could certainly share impl.
• I think we should keep the user-defined component_dict completely separate from whatever structure we use to hold the component instances. That way, the component_dict holds the template for the graph. I suppose add_node and add_edge would have to modify that structure.

I can also see us wanting to evaluate part(s) of the graph, i.e. call transform on all preprocessing features. We don't have to add that to this PR, but I think that would be great to file as something to add once the DAGs work is complete.

[angela97lin]

Just leaving tiny comments for now, would love to take a closer look when I get the chance later 😁

[bchen1116]

The implementation looks great, and the tests are awesome! Just had a few questions/comments.

I agree with Dylan about submitting an issue to track the caching for this class, since the current cache might run into memory issues.

There currently is no implementation for adding additional nodes or adding two graphs together, as mentioned in the original issue, is there? Not sure if I missed something, but might be nice to file an issue to track adding those two!

Also, not sure if this was intentional, but this class hasn't been added to the docs yet.

Otherwise, I left a few comments, but looks good to me!

[dsherry]

I agree with Dylan about submitting an issue to track the caching for this class, since the current cache might run into memory issues.

@bchen1116 good point, already filed! (idk if my prev comment mentioned this) #466

[eccabay]

There currently is no implementation for adding additional nodes or adding two graphs together, as mentioned in the original issue, is there? Not sure if I missed something, but might be nice to file an issue to track adding those two!

Great point @bchen1116! I did remove these capabilities as I iterated through this PR. With the decision for the ComponentGraph class to be completely internal/private (users shouldn't ever need to muck with it), any use cases for adding nodes or merging graphs disappeared. That being said, it may be worth it to file another issue to track adding that in the future, should use cases arise. (This PR is already so large, that work would probably be left to a separate issue)

[dsherry]

@eccabay solid work on this!! 🎉

Here's all that I found which I'd consider blocking:

• I proposed a simplification in compute_final_component_features / _compute_features, so that we have one evaluation impl instead of two.
• I left a few little suggested cleanups in _compute_features.

[dsherry]

If you have any example images handy it would be cool to see what this looks like!

[eccabay]

Sure! Here's an uninstantiated pipeline

[eccabay]

And here's the same pipeline after instantiation with default parameters!

[dsherry]

Ooooooh. This rocks. Thanks!