# Contributing CatBoost to a Python AutoML Library — My GitHub PR Walkthrough

## Introduction

In this post, I'll walk you through my recent open source contribution — adding CatBoost support to [SwiftPredict](https://github.com/ManasRanjanJena253/SwiftPredict), a fully local AutoML and experiment tracking library built with Python. I'll cover the exact code changes, challenges I faced, and how I tested it end-to-end.

* * *

## What is SwiftPredict?

SwiftPredict is an open source AutoML library that collapses the entire machine learning pipeline into a single `fit()` call. It handles:

*   Null imputation
    
*   Categorical and text encoding
    
*   Task detection (classification vs regression)
    
*   Class imbalance correction (SMOTE)
    
*   Multi-model training with 5-fold cross-validation
    
*   Experiment tracking via local MongoDB
    

```python
from swiftpredict import AutoML

model = AutoML()
results = model.fit(
    project_name="cancer-prediction",
    file_path="cancer.csv",
    target_column="target"
)
print(results)
```

* * *

## The Issue

The maintainer opened [Issue #7](https://github.com/ManasRanjanJena253/SwiftPredict/issues/2) requesting CatBoost to be added to the model zoo. The existing model zoo included XGBoost and LightGBM but was missing CatBoost — a gradient boosting library known for handling categorical features natively and performing well in competitive ML.

* * *

## What is a Model Zoo?

A model zoo is a collection of pre-defined ML models that SwiftPredict trains and evaluates automatically. When you call `model.fit()`, it trains every model in the zoo and picks the best one.

**Before my contribution — 5 models:**

*   GaussianNB, XGBClassifier, RandomForestClassifier, LGBMClassifier, LogisticRegression
    

**After my contribution — 6 models:**

*   GaussianNB, XGBClassifier, RandomForestClassifier, LGBMClassifier, LogisticRegression, **CatBoostClassifier** ✅
    

Same for regression — CatBoostRegressor was added alongside LinearRegression, XGBRegressor, LGBMRegressor, and RandomForestRegressor.

* * *

## The Changes

Only one file was modified: `backend/app/services/preprocessing.py`

### 1\. Import CatBoost

```python
from catboost import CatBoostClassifier, CatBoostRegressor  # Importing CatBoost models for classification and regression
```

### 2\. Add to model\_zoo()

```python
def model_zoo(task, model=None):
    if task == "classification":
        models = [GaussianNB, XGBClassifier, RandomForestClassifier,
                  LGBMClassifier, LogisticRegression, CatBoostClassifier]  # Added CatBoostClassifier
    else:
        models = [LinearRegression, XGBRegressor, LGBMRegressor,
                  RandomForestRegressor, CatBoostRegressor]  # Added CatBoostRegressor
```

### 3\. Handle CatBoost in train\_model()

CatBoost doesn't support `n_jobs` like sklearn models, and prints verbose training logs by default. The fix was to handle it separately with `verbose=0`:

```python
elif k.__name__ == "CatBoostClassifier":
    model = k(verbose=0)  # verbose=0 suppresses CatBoost console output
```

```python
if k.__name__ == "CatBoostRegressor":
    model = k(verbose=0)  # verbose=0 suppresses CatBoost console output
```

This is consistent with how LightGBM is handled using `verbose=-1`.

* * *

## Testing

I tested the pipeline end-to-end using the breast cancer dataset from scikit-learn:

```python
from swiftpredict import AutoML

model = AutoML()
results = model.fit(
    project_name="cancer-test",
    file_path="cancer.csv",
    target_column="target"
)

metrics = model.evaluate_performance(key="f1")
print(metrics)
```

**Results:**

```plaintext
Training the Models: 100%|████████| 6/6 ✅

Best Models: {'f1': 'LGBMClassifier', 'precision': 'LGBMClassifier',
              'accuracy': 'LGBMClassifier', 'overall': ['LGBMClassifier']}

Metrics: {'accuracy': 0.93, 'f1': 0.93, 'roc_auc': 0.925, 'precision': 0.93}
```

All 6 models trained successfully with no extra console output from CatBoost.

* * *

## Challenges I Faced

**1\. Disk space issues** Installing CatBoost (101.7 MB) on a nearly full C: drive caused `[Errno 28] No space left on device`. Fixed by moving the project to D: drive and clearing pip cache.

**2\. Virtual environment issues** Packages were installing to system Python instead of the venv. Fixed by recreating the venv on D: drive.

**3\. MongoDB connection** The library requires MongoDB for experiment tracking. Used MongoDB Atlas (cloud) instead of installing locally.

**4.** `n_jobs` **error** CatBoost doesn't accept `n_jobs` as a constructor argument unlike sklearn models. Required a separate `elif` block to initialize it correctly.

* * *

## Key Takeaways

*   Open source contributions don't have to be massive — adding one well-tested model with proper handling is valuable
    
*   Always read the maintainer's guidelines before submitting a PR
    
*   Test your changes end-to-end before opening a PR
    
*   Add comments to new code as requested by the maintainer
    

* * *

## PR Link

👉 [feat: Add CatBoost support to model zoo](https://github.com/ManasRanjanJena253/SwiftPredict/pulls)

* * *

## Conclusion

My first open source contribution was small but meaningful — expanding SwiftPredict's model zoo from 5 to 6 models by adding CatBoost support. If you're looking to make your first open source contribution, look for issues labeled `good first issue` or `help wanted` on GitHub. Start small, test thoroughly, and follow the maintainer's guidelines.

Happy contributing! 🚀