Add KNN outlier detector

[mauicv]

What is this:

This PR implements the kNN outlier detector. See simple example.

• This detector has a PyTorch backend which can be directly serialized using torchscript. To do so the user should use:

  knn_detector = KNN(k=5)
  knn_detector.fit(X_ref)
  knn_detector.infer_threshold(X_ref, 0.1)
  knn_detector = torch.jit.script(knn_detector.backend)
  knn_detector.save('./knn_detector.pt')

• The kNN detector can be used as a single detector for a single k or an ensemble if passed multiple k. In the latter case, the detector must also be passed an aggregator to accumulate the multiple scores output for the multiple ks. For example:

  from alibi_detect.od.backend import ShiftAndScaleNormaliserTorch, AverageAggregatorTorch
  
  knn_detector = KNN(
      k=[5, 8, 12], 
      aggregator=AverageAggregatorTorch(), 
      normaliser=ShiftAndScaleNormaliserTorch()
  )
  
  knn_detector.fit(X_ref)
  knn_detector.infer_threshold(X_ref, 0.1)

• This PR also implements multiple aggregators and normalisers that can be used to normalise and aggregate ensemble outlier scores.

Out of scope

This PR does not Implement human-readable configuration or a pykeops backend. Neither does it implement documentation (i.e. example or method pages.) These will all be implemented in a later PR.

Notes:

Sub Module responsibilities:

A feature we want is torchscript-able components in order to ease deployment (See notes here). In order to ensure that we can do so while also ensuring that the torchscript language constraints don't contaminate too much of the code base we scope this functionality to the relevant subcomponents. In particular, each module has separate responsibilities:

KNNTorch:

• Implements torch KNN functionality, specifically forward method
• Extends torch.nn.Module
• Torchscript-able
• has fit, infer_threshold and predict methods.

KNN:

• Wrapper for the KNNTorch functionality. Specifically, fit, infer_threshold and predict
• Deals with validation
• Config <- (NOT IMPLEMENTED HERE)
• Does not extend torch.nn.module
• Uses backend functionality to convert inputs to correct backend tensor type.

mypy issue:

For some reason, mypy interprets the type of the accumulator attribute and check_fitted method as torch.Tensor This raises an error in mypy because torch.Tensor doesn't have a call method and we call both of these objects within the predict method. Explicitly defining these objects types in the class fixes the issue but then causes problems with torch script compiling.

Todo:

• Add more here
• Fix pytest -s --no-cov alibi_detect/od/backend/tests/test_knn_backend.py::test_knn_kernel test
• Decide on fit/infer-threshold error behaviour
• Add torchscript serliaziation tests
• Add integration tests
• Add validation in KNN class
• Protect optional dep imports in `KNN
• correctly reorder imports
• Add optional dep tests
• convert all normaliser to normalizer.
• Fix mypy

---

More Todo:

• I've just realised that the detectors define an outlier detector return dictionary:

  def outlier_prediction_dict():
      data = {
          'instance_score': None,
          'feature_score': None,
          'is_outlier': None
      }
      return copy.deepcopy({"data": data, "meta": DEFAULT_META})

  This should be used instead of the custom one I'm using in the kNN detector.
• Consider alternatives/unifications of _to_numpy pattern!
• test torch.compile on detector
• expose backend aggregator and normalizer class and rename backend to PyTorch. i.e. 1 and 3 in this comment.
• remove from utils.types import Protocol logic and replace with from typing_extensions import Protocol
• Revert _to_numpy pattern as it messes up typing 🤦
  • I've Replaced this with a singular dispatch pattern. This is a little exotic 😬 so happy to remove if peeps want!

[codecov-commenter]

Codecov Report

Merging #677 (a535e8d) into master (40f4121) will increase coverage by 0.46%.
The diff coverage is 94.98%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #677      +/-   ##
==========================================
+ Coverage   80.33%   80.79%   +0.46%     
==========================================
  Files         137      144       +7     
  Lines        9304     9619     +315     
==========================================
+ Hits         7474     7772     +298     
- Misses       1830     1847      +17     

Flag	Coverage Δ
macos-latest-3.9	?
ubuntu-latest-3.10	80.72% <94.98%> (+0.50%)	⬆️
ubuntu-latest-3.7	80.62% <94.92%> (+0.50%)	⬆️
ubuntu-latest-3.8	?
ubuntu-latest-3.9	80.67% <94.98%> (+0.50%)	⬆️
windows-latest-3.9	?

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
alibi_detect/od/__init__.py	100.00% <ø> (ø)
alibi_detect/base.py	85.45% <50.00%> (ø)
alibi_detect/od/base.py	75.75% <75.75%> (ø)
alibi_detect/od/pytorch/base.py	97.10% <97.10%> (ø)
alibi_detect/od/pytorch/ensemble.py	97.11% <97.11%> (ø)
alibi_detect/od/_knn.py	97.95% <97.95%> (ø)
alibi_detect/exceptions.py	100.00% <100.00%> (ø)
alibi_detect/od/pytorch/__init__.py	100.00% <100.00%> (ø)
alibi_detect/od/pytorch/knn.py	100.00% <100.00%> (ø)

... and 2 files with indirect coverage changes

[ojcobb]

Small number of comments around defaults and threshold inference, but otherwise looks good to me. Nice one!

[ascillitoe]

LGTM! Just a few very minor comments remaining...

[ascillitoe]

LGTM!