Release 0.4.1

[nnansters]

Hi everyone,

Niels here with just a quick update. We've just released version 0.4.1 of NannyML.
And yes, we're messing up the semantic versioning a bit for now, but I don't feel like this is a full minor version 🙂

Installing / upgrading

You can get this latest version by using pip:

pip install -U nannyml

Whats new?

We've added some very early support for regression use cases! Now hold your horses: we don't have anything implemented for performance estimation or calculation yet. We need a lot more research first.
But with this update you'll be able to create metadata for a regression model and use that metadata with all our drift calculators.

Some example code:

import nannyml as nml
import numpy as np
import pandas as pd

# Hijacking the binary dataset here    
ref_df, ana_df, tgt_df = nml.load_synthetic_binary_classification_dataset()

# Don't need probabilities
ref_df = ref_df.drop(columns=['y_pred_proba'])
ana_df = ana_df.drop(columns=['y_pred_proba'])

# We do need predicted values
ref_df['y_pred'] = np.random.randn(len(ref_df))
ana_df['y_pred'] = np.random.randn(len(ref_df))

# Extract metadata for regression model type
md = nml.extract_metadata(data, model_type='regression', exclude_columns=['identifier'])
md.target_column_name = 'work_home_actual' # Won't need this for drift calculations but metadata has to be complete

# Perform univariate drift detection
calc_univar = UnivariateStatisticalDriftCalculator(metadata, chunk_size=5000).fit(reference_data=ref_df)
drift = calc_univar.calculate(pd.concat([ref_df, ana_df]))
drift.plot(kind='prediction_drift', class_label=1).show()
drift.plot(kind='prediction_distribution', class_label=1).show()

# Perform multivariate drift detection
calc_multivar = DataReconstructionDriftCalculator(metadata, chunk_size=5000).fit(reference_data=ref_df)
drift = calc_multivar.calculate(pd.concat([ref_df, ana_df]))
drift.plot(kind='drift').show()

What's changed?

The DefaultChunker (used when not specifying any chunking preferences) no longer uses the minimum_chunk_size (read more in the docs) to split data into chunks. It will now split into 10 chunks of equal sizes.

the SizeBasedChunker used to discard any leftover observations that were too few to create a chunk with chunk_size. We've now made that behavior configurable, defaulting to keep the - possibly underpopulated - last chunk as well.

That was all there was to it. Told you, not really a minor release eh. We'll stick to semver principles better as soon as we hit that magical 1.0.0 release.

We're still wrapping up our development cycle, so expect some news on what we'll be working on next soon!

Thanks again for your wonderful support, stars, bug requests, reviews, ... It means the world to us! ❤️

Cheers!