🌲 PrettyForest

Interactive 2.5D forest visualization for tree-based ML ensembles. Explore your model's structure, trace predictions through individual trees, and understand how the ensemble makes decisions.

Features

• 2.5D isometric forest — trees rendered with depth perspective, growth animation, seasonal themes
• All major frameworks — scikit-learn, LightGBM, CatBoost (RandomForest, GradientBoosting, single DecisionTree)
• Prediction tracing — select a data point, see per-tree corrections/votes, get the ensemble's actual prediction
• True label comparison — pass target=y_test to see if the model got it right
• Interactive detail view — double-click a tree to drill into its decision structure with per-node expansion
• Boosted tree awareness — correct display of gradient corrections vs class probabilities, with explanatory notes
• Dark mode — toggle with 🌙 button
• Scales to thousands — pagination, sorting, spotlighting for large ensembles

Installation

pip install prettyforest

# With optional frameworks
pip install prettyforest[all]  # includes lightgbm + catboost

Quick Start

from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
import polars as pl
from prettyforest import prettygrow

# Train
iris = load_iris()
X_train, X_test, y_train, y_test = train_test_split(
    iris.data, iris.target, test_size=0.3, random_state=42
)
model = RandomForestClassifier(n_estimators=20, max_depth=5, random_state=42)
model.fit(X_train, y_train)

# Visualize
X_test_pl = pl.DataFrame({name: X_test[:, i] for i, name in enumerate(iris.feature_names)})
prettygrow(model, data=X_test_pl, target=y_test, output_path="forest.html")

API

from prettyforest import prettygrow

prettygrow(
    model,              # Trained tree model (sklearn, LightGBM, or CatBoost)
    *,
    data=None,          # Polars DataFrame or NumPy array for prediction tracing
    target=None,        # True labels/values — shown alongside predictions
    output_path=None,   # Write HTML to file; if None, returns string or displays in notebook
    feature_names=None, # Column names when data is a NumPy array
    seed=42,            # Random seed for layout
    season=None,        # Color theme: "summer", "autumn", "winter", or None (metric-based)
)

Returns: Interactive AnyWidget instance when running inside a notebook environment (JupyterLab, Marimo, VS Code, Google Colab); raw HTML string if outside a notebook and no output_path; None if output_path is specified.

Supported Models

Interaction Guide

Seasons & Themes

Switch the visual theme live in the browser or set it via Python:

prettygrow(model, season="spring")   # 🌸 Light greens + pink blossoms
prettygrow(model, season="summer")   # 🌿 Deep lush greens
prettygrow(model, season="autumn")   # 🍂 Warm oranges, reds, golds
prettygrow(model, season="winter")   # ❄️ Bare branches, blue-grey
prettygrow(model)                    # 🌳 Natural (metric-based coloring)

You can also switch seasons on the fly using the dropdown in the header — no need to re-run Python. The canopies, ground, sky, and grass patches all update instantly.

Prediction Display

When you trace a sample:

• Per-tree badges show on each tree's canopy:
  • Random Forest: the class vote (colored by class)
  • Boosted models: the raw correction value (green = positive, red = negative)
• Ensemble prediction: the actual model.predict() output — always correct
• True label: shown in green if target was provided

Marimo & Jupyter Notebook Integration

PrettyForest uses native AnyWidget support for bi-directional synchronization and interactive rendering across all notebook environments (JupyterLab, Marimo desktop/cloud, VS Code, and Google Colab).

from sklearn.ensemble import RandomForestClassifier
from sklearn.datasets import load_iris
import polars as pl
from prettyforest import prettygrow

iris = load_iris()
model = RandomForestClassifier(n_estimators=20, max_depth=5, random_state=42)
model.fit(iris.data, iris.target)

X_pl = pl.DataFrame({name: iris.data[:, i] for i, name in enumerate(iris.feature_names)})

# Directly returns and displays an interactive AnyWidget inside Jupyter or Marimo
prettygrow(model, data=X_pl, target=iris.target)

Understanding Boosted vs Bagged Trees

Random Forest (bagged): Each tree trains independently on a random data subset. Leaves contain real class proportions or target means. Ensemble averages/votes.

Gradient Boosting / LightGBM / CatBoost (boosted): Trees train sequentially — each corrects the previous ensemble's errors. Leaf values are small gradient adjustments, not standalone predictions. Ensemble sums corrections.

When you double-click a boosted tree, a warning appears:

⚠️ This is a boosted tree — leaf values are gradient corrections, not final predictions.

The splits and features are fully interpretable — they show which features matter and how the space is partitioned. The leaf values just represent "how much to adjust" rather than "what to predict."

Development

git clone https://github.com/fabioscantamburlo/prettyforest.git
cd prettyforest
uv sync

# Run tests
uv run pytest

# Run experiments (all models)
uv run run_experiments.py --n-samples 1000 --n-trees 50 --max-depth 8

# Iris example (quick, all frameworks)
uv run examples/iris_forest.py

# MNIST example (larger, RF only)
uv run examples/mnist_forest.py

Pre-commit

pre-commit install  # set up hooks
pre-commit run --all-files  # manual run

Hooks: trailing whitespace, end-of-file, YAML check, large file guard (500KB), ruff lint + format.

License

MIT