Update anomaly_detection model for new photometry

[matwey]

IMPORTANT: Please create an issue first before opening a Pull Request.
Linked to issue(s): Closes put link

What changes were proposed in this pull request?

Retrain anomaly_detection model on lc_features_20210617_photometry_corrected.parquet

How was this patch tested?

N/A

[JulienPeloton]

Hi @matwey -- thanks for the retraining. Any ideas why the model grew from 13MB to 30MB with a similar dataset?

[JulienPeloton]

@matwey -- we use a very specific version of sklearn (see here for the version used). Hence the CI is failing:

/home/libs/miniconda/lib/python3.9/site-packages/sklearn/base.py:329: UserWarning: Trying to unpickle estimator ExtraTreeRegressor from version 1.3.0 when using version 1.0.2. This might lead to breaking code or invalid results. Use at your own risk. For more info please refer to:
https://scikit-learn.org/stable/modules/model_persistence.html#security-maintainability-limitations
  warnings.warn(
Traceback (most recent call last):
  File "/__w/fink-science/fink-science/fink_science/anomaly_detection/processor.py", line 66, in <module>
    forest_r = pickle.load(forest_file)
  File "sklearn/tree/_tree.pyx", line 653, in sklearn.tree._tree.Tree.__setstate__
  File "sklearn/tree/_tree.pyx", line 1340, in sklearn.tree._tree._check_node_ndarray
ValueError: node array from the pickle has an incompatible dtype:
- expected: [('left_child', '<i8'), ('right_child', '<i8'), ('feature', '<i8'), ('threshold', '<f8'), ('impurity', '<f8'), ('n_node_samples', '<i8'), ('weighted_n_node_samples', '<f8')]
- got     : {'names': ['left_child', 'right_child', 'feature', 'threshold', 'impurity', 'n_node_samples', 'weighted_n_node_samples', 'missing_go_to_left'], 'formats': ['<i8', '<i8', '<i8', '<f8', '<f8', '<i8', '<f8', 'u1'], 'offsets': [0, 8, 16, 24, 32, 40, 48, 56], 'itemsize': 64}

Could you retrain using sklearn 1.0.2 ?

[matwey]

Hi @JulienPeloton

1. I've retrained using sklearn 1.0.2 and Python 3.9 and force-pushed the commit
2. Currently, I use notebooks provided by @VoidDruid and there is a hyper-parameter tuning stage with GridSearchCV. I suppose that hyper-parameters (trees number) appeared to be different for different data.

However, I think that it is desirable to migrate to ONNX to address both questions. First, it is size-efficient, second it is version compatible. I think this should be next step for this module.

[JulienPeloton]

Thanks for the retraining -- there is one unit test failing (expected as you changed the model):

File "/__w/fink-science/fink-science/fink_science/__init__.py", line ?, in __main__.anomaly_score
Failed example:
    df.filter(df["anomaly_score"] < -0.5).count()
Expected:
    15
Got:
    7

However, I think that it is desirable to migrate to ONNX to address both questions

I completely agree!

[matwey]

Thanks for the retraining -- there is one unit test failing (expected as you changed the model):

File "/__w/fink-science/fink-science/fink_science/__init__.py", line ?, in __main__.anomaly_score
Failed example:
    df.filter(df["anomaly_score"] < -0.5).count()
Expected:
    15
Got:
    7

This was from a docsting, and fixed now.

[codecov-commenter]

Codecov Report

Patch and project coverage have no change.

Comparison is base (68fa05e) 53.44% compared to head (05d090a) 53.44%.
Report is 1 commits behind head on master.

❗ Current head 05d090a differs from pull request most recent head 4149708. Consider uploading reports for the commit 4149708 to get more accurate results

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #329   +/-   ##
=======================================
  Coverage   53.44%   53.44%           
=======================================
  Files          33       33           
  Lines        1510     1510           
=======================================
  Hits          807      807           
  Misses        703      703           

Files Changed	Coverage Δ
fink_science/anomaly_detection/processor.py	77.35% <ø> (ø)

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[JulienPeloton]

Thanks! I'm merging.