tinyshift 0.4.0

A small toolbox for mlops

TinyShift

TinyShift is a small experimental Python library designed to detect data drifts and performance drops in machine learning models over time. The main goal of the project is to provide quick and tiny monitoring tools to help identify when data or model performance unexpectedly change. For more robust solutions, I highly recommend Nannyml.

Technologies Used

• Python 3.x
• Scikit-learn
• Pandas
• NumPy
• Plotly
• Scipy

Installation

To install TinyShift in your development environment, use pip:

pip install tinyshift

If you prefer to clone the repository and install manually:

git clone https://github.com/HeyLucasLeao/tinyshift.git
cd tinyshift    
pip install .

Note: If you want to enable plotting capabilities, you need to install the extras using Poetry:

poetry install --all-extras

Usage

Below are basic examples of how to use TinyShift's features.

1. Data Drift Detection

To detect data drift, simply score in a new dataset to compare with the reference data. The DataDriftDetector will calculate metrics to identify significant differences.

from tinyshift.detector import CategoricalDriftDetector

df = pd.DataFrame("examples.csv")
df_reference = df[(df["datetime"] < '2024-07-01')].copy()
df_analysis = df[(df["datetime"] >= '2024-07-01')].copy()

detector = CategoricalDriftDetector(df_reference, 'discrete_1', "datetime", "W", drift_limit='mad')

analysis_score = detector.score(df_analysis, "discrete_1", "datetime")

print(analysis_score)

2. Performance Tracker

To track model performance over time, use the PerformanceMonitor, which will compare model accuracy on both old and new data.

from tinyshift.tracker import PerformanceTracker

df_reference = pd.read_csv('refence.csv')
df_analysis = pd.read_csv('analysis.csv')
model = load_model('model.pkl') 
df_analysis['prediction'] = model.predict(df_analysis["feature_0"])

tracker = PerformanceTracker(df_reference, 'target', 'prediction', 'datetime', "W")

analysis_score = tracker.score(df_analysis, 'target', 'prediction', 'datetime')

print(analysis_score)

3. Visualization

TinyShift also provides graphs to visualize the magnitude of drift and performance changes over time.

tracker.plot.scatter(analysis_score, fig_type="png")

tracker.plot.bar(analysis_score, fig_type="png")

4. Outlier Detection

To detect outliers in your dataset, you can use the models provided by TinyShift. Currently, it offers the Histogram-Based Outlier Score (HBOS), Simple Probabilistic Anomaly Detector (SPAD), and SPAD+.

from tinyshift.outlier import SPAD

df = pd.read_csv('data.csv')

spad_plus = SPAD(plus=True)
spad_plus.fit(df)

anomaly_scores = spad_plus.decision_function(df)

print(anomaly_scores)

5. Anomaly Tracker

The Anomaly Tracker in TinyShift allows you to identify potential outliers based on the drift limit and anomaly scores generated during training. By setting a drift limit, the tracker can flag data points that exceed this threshold as possible outliers.

from tinyshift.tracker import AnomalyTracker

model = load_model('model.pkl') 

tracker = AnomalyTracker(model, drift_limit='mad')

df_analysis = pd.read_csv('analysis.csv')

outliers = tracker.score(df_analysis)

print(outliers)

In this example, the AnomalyTracker is initialized with a reference model and a specified drift limit. The score method evaluates the analysis dataset, calculating anomaly scores and flagging data points that exceed the drift limit as potential outliers.

Project Structure

The basic structure of the project is as follows:

tinyshift
├── LICENSE
├── README.md
├── poetry.lock
├── pyproject.toml
└── tinyshift
    ├── examples
    │   ├── outlier.ipynb
    │   └── tracker.ipynb
    ├── outlier
    │   ├── __init__.py
    │   ├── base.py
    │   ├── hbos.py
    │   └── spad.py
    ├── plot
    │   ├── __init__.py
    │   └── plot.py
    ├── tests
    │   ├── test_hbos.py
    │   └── test_spad.py
    └── tracker
        ├── anomaly.py
        ├── base.py
        ├── categorical.py
        ├── continuous.py
        └── performance.py      

License

This project is licensed under the MIT License - see the LICENSE file for more details.