retrofit 0.2.0

High-Performance ML Training, Scoring & Evaluation (Polars + GPU-Ready)

Installation

# Most up-to-date
pip install git+https://github.com/AdrianAntico/RetroFit.git#egg=retrofit

# From pypi
pip install retrofit==0.1.7

📦 RetroFit

High-Performance ML Training, Scoring & Evaluation (Polars + GPU-Ready)

Table of Contents

• Installation
• RetroFit
  • Key Features
  • Polars-Native Modeling Pipeline
  • Target Transformations (Regression)
  • GPU-Ready Training
  • Unified Scoring Engine
  • Full Evaluation Suite
  • Calibration Tables & Plot
  • ROC / PR / PR-ROC Curves
  • QuickEcharts Visuals
  • Partial Dependence Plots (PDP)
• Code Examples
  • CatBoost Examples
  • XGBoost Examples
  • LightGBM Examples
• Model Evaluation Visuals

RetroFit is a fast, production-oriented machine learning framework designed for training, scoring, and evaluating models.

Built from the ground up with Polars, GPU acceleration, and a scalable evaluation engine, RetroFit provides a unified interface for:

• 🧪 Model data creation & preprocessing (Polars-first)
• 🚀 Fast training with automatic CPU/GPU switching
• 📊 Unified scoring engine with inverse-transformation support
• 📈 Full evaluation suite for regression, binary, and multiclass
• 🎯 Calibration plots & tables
• 📉 ROC, PR, and PR-ROC curves (QuickEcharts visuals)
• 🔧 Automatic label encoding for classification/multiclass
• 🔄 Target variable transformations

RetroFit is designed for data scientists who want speed, modern tooling, and high-quality diagnostics without boilerplate.

---

🔥 Key Features

⚙️ 1. Polars-Native Modeling Pipeline

RetroFit uses Polars internally for:

• Numeric & categorical handling
• Efficient grouping and slicing
• Scored-data postprocessing
• Evaluation data wrangling
• Data preparation for CatBoost/XGBoost/LightGBM

Everything is vectorized whenever possible.

---

🎯 2. Target Transformations (Regression)

RetroFit supports:

• "none"
• "log" (auto detects ≤0 and applies min-shift)
• "sqrt"
• "standardize"

Transformation is applied automatically in create_model_data() and reversed in score(). User could also apply themselves before running create_model_data() and after running score().

---

⚡ 3. GPU-Ready Training

Enable GPU training with:

model = RetroFit(Algorithm="catboost", TargetType="regression", GPU=True)

RetroFit automatically:

• Switches tree construction method
• Adjusts booster settings
• Removes CPU-only parameters
• Ensures full CatBoost/XGBoost/LightGBM compatibility

---

📊 4. Unified Scoring Engine

model.score(DataName="test")

Or score external data:

model.score(NewData=df)

Outputs a Polars DataFrame with:

• Predictions (Predict_target)
• Probabilities (p1, class_k)
• Inverse-transformed regression predictions

---

🧮 5. Full Evaluation Suite

Regression Metrics

• R²
• MAE / MedianAE
• MAPE
• MSE / RMSE
• MSLE (auto-disabled if invalid)

Binary Classification

• Accuracy, Recall, Precision
• TPR, FPR, TNR, FNR
• F1, F0.5, F2
• MCC
• Threat score
• Utility (custom cost matrix)
• Full 101-point threshold curve

Multiclass Classification

• Overall: Accuracy, macro/micro/weighted F1
• One-vs-all threshold evaluation
• Uses label decoder to restore original class names

---

🎛️ 6. Calibration Tables & Plot

Regression + Classification calibration:

• Equal-width or quantile binning
• MACE, RMSE, MAE, R²
• Per-group calibration
• QuickEcharts visualization
• Metadata: timestamp, model name, grouping vars

---

📈 7. ROC / PR / PR-ROC Curves

RetroFit generates:

• ROC
• Precision-Recall
• AUC and Average Precision
• QuickEcharts Area plots with gradient shading

---

🎨 8. QuickEcharts Visuals

All plots are powered by QuickEcharts:

• Line / Area plots
• Gradient fills
• Auto subtitles with metrics
• HTML export
• Themes: 'chalk', 'dark', 'essos', 'halloween', 'infographic', 'light', 'macarons', 'purple-passion', 'roma', 'romantic', 'shine', 'vintage', 'walden', 'westeros', 'white', 'wonderland'

---

9. 🔍 Partial Dependence Plots (PDP)

RetroFit includes full PDP support for both numeric and categorical features to help explain model behavior.

✅ Numeric PDP

• Automated binning (quantile or equal-width)
• Mean actual vs mean predicted across bins
• String-safe axis handling for QuickEcharts
• Clean Line-chart visualization
• Supports internal scored data or external df input
• Returns both the PDP table and plot object

✅ Categorical PDP

• Per-category partial dependence table
• Mean actual vs mean predicted for each category
• Optional sorting (feature order, actual mean, predicted mean)
• Line-based visualization using QuickEcharts
• Works with internal or external data
• Returns both table and plot object

Code Examples

1. Supervised Learning Examples

Below are runnable examples for all supported algorithms.
Each section includes regression, binary classification, and multiclass examples.

---

CatBoost Examples

Regression Training

# Setup environment
import os
import polars as pl
from PolarsFE import datasets
from QuickEcharts import Charts
from retrofit import MachineLearning as ml
from retrofit import utils


# Load some data
df = utils.make_retrofit_demo_data(
    n_rows=50_000,
    n_segments=5,
    seed=42,
)

# Get TrainData, ValidationData, and TestData
DataSets = datasets.partition_random(
    data=df,
    num_partitions=3,
    seed=42,
    percentages=[0.7, 0.2, 0.1]
)

# Initialize RetroFit
model = ml.RetroFit(TargetType="regression", Algorithm="catboost")

# Create algo-specific model data
model.create_model_data(
  TrainData=DataSets[0],
  ValidationData=DataSets[1],
  TestData=DataSets[2],
  TargetColumnName="Leads",
  NumericColumnNames=['XREGS1', 'XREGS2', 'XREGS3'],
  CategoricalColumnNames=['MarketingSegments', 'MarketingSegments2', 'MarketingSegments3'],
  TextColumnNames=None,
  WeightColumnName=None,
  Threads=-1
)

# Print default parameter settings
model.print_algo_args()

# Update algo args for GPU
model.update_model_parameters(
    task_type='GPU',
    sampling_frequency=None,
    rsm=1.0,
    iterations = 200
)

# Train Model
model.train()

# Score train, validation, and test; store internally
model.score()

# Inspect scored data
model.ScoredData["train"]
model.ScoredData["validation"]
model.ScoredData["test"]

# Evaluate scored data
global_eval = model.evaluate(
    DataName="test"
)

# Per segment
segment_eval = model.evaluate(
    DataName="test",
    ByVariables="MarketingSegments"
)

# Get variable importance
imp = model.compute_feature_importance()

# Get interaction importance
interact = model.compute_catboost_interaction_importance()

# Model Calibration Tables
cal = model.build_regression_calibration_table(
    DataName="test",
    binning="quantile"
)

# Store plot in working directory
model.plot_regression_calibration(
    DataName="test",
    n_bins=20,
    binning="quantile",
    plot_name=f"{os.getcwd()}/my_calibration_plot"
)

# Actual vs Predicted Scatterplot
model.plot_regression_scatter(
    DataName="test",
    SampleSize=15000,
    plot_name=f"{os.getcwd()}/my_scatter_plot",
    Theme="dark"
)

# Numeric Partial Dependence plot
model.plot_pdp_numeric(
    feature='XREGS1',
    DataName="test",
    plot_name=f"{os.getcwd()}/numeric_pdp",
)

# Categorical Partial Dependence plot
model.plot_pdp_categorical(
    feature='MarketingSegments',
    DataName="test",
    plot_name=f"{os.getcwd()}/categorical_pdp",
)

Classification Training

# Setup environment
import os
import polars as pl
from PolarsFE import datasets
from QuickEcharts import Charts
from retrofit import MachineLearning as ml
from retrofit import utils


# Load some data
df = utils.make_retrofit_demo_data(
    n_rows=50_000,
    n_segments=5,
    seed=42,
)

# Get TrainData, ValidationData, and TestData
DataSets = datasets.partition_random(
    data=df,
    num_partitions=3,
    seed=42,
    percentages=[0.7, 0.2, 0.1]
)

# Initialize RetroFit
model = ml.RetroFit(TargetType="classification", Algorithm="catboost")

# Create algo-specific model data
model.create_model_data(
  TrainData=DataSets[0],
  ValidationData=DataSets[1],
  TestData=DataSets[2],
  TargetColumnName="Label_binary",
  NumericColumnNames=['XREGS1','XREGS2','XREGS3'],
  CategoricalColumnNames=['MarketingSegments','MarketingSegments2','MarketingSegments3'],
  TextColumnNames=None,
  WeightColumnName=None,
  Threads=-1
)

# Print default parameter settings
model.print_algo_args()

# Update algo args for GPU
model.update_model_parameters(
    task_type='GPU',
    sampling_frequency=None,
    rsm=1.0,
    iterations=200
)

# Train Model
model.train()

# Score train, validation, and test; store internally
model.score()

# Inspect scored data
model.ScoredData["train"]
model.ScoredData["validation"]
model.ScoredData["test"]

# Evaluate scored data
global_eval = model.evaluate(
    DataName="test"
)

# Per segment
segment_eval = model.evaluate(
    DataName="test",
    ByVariables="MarketingSegments"
)

# Get variable importance
imp = model.compute_feature_importance()

# Get interaction importance
interact = model.compute_catboost_interaction_importance()

# Store plot in working directory
model.plot_classification_calibration(
    DataName="test",
    n_bins=20,
    binning="quantile",
    plot_name=f"{os.getcwd()}/my_calibration_plot"
)

# ROC Plot
model.plot_classification_roc(
    DataName="test",
    plot_name=f"{os.getcwd()}/my_roc_plot",
)

# PR Plot
model.plot_classification_pr(
    DataName="test",
    plot_name=f"{os.getcwd()}/my_pr_plot"
)

# Numeric Partial Dependence plot
model.plot_pdp_numeric(
    feature='XREGS1',
    DataName="test",
    plot_name=f"{os.getcwd()}/numeric_pdp",
)

# Categorical Partial Dependence plot
model.plot_pdp_categorical(
    feature='MarketingSegments',
    DataName="test",
    plot_name=f"{os.getcwd()}/categorical_pdp",
)

MultiClass Training

import os
import polars as pl
from PolarsFE import datasets
from QuickEcharts import Charts
from retrofit import MachineLearning as ml

# Load some data
# Generate demo data instead of reading packaged CSVs
df = utils.make_retrofit_demo_data(
    n_rows=50_000,
    n_segments=5,
    seed=42,
)

# Get TrainData, ValidationData, and TestData
DataSets = datasets.partition_random(
    data=df,
    num_partitions=3,
    seed=42,
    percentages=[0.7, 0.2, 0.1]
)

# Initialize RetroFit
model = ml.RetroFit(TargetType="multiclass", Algorithm="catboost")

# Create algo-specific model data
model.create_model_data(
  TrainData=DataSets[0],
  ValidationData=DataSets[1],
  TestData=DataSets[2],
  TargetColumnName="Label",
  NumericColumnNames=['XREGS1', 'XREGS2', 'XREGS3'],
  CategoricalColumnNames=['MarketingSegments', 'MarketingSegments2', 'MarketingSegments3'],
  TextColumnNames=None,
  WeightColumnName=None,
  Threads=-1
)

# Print default parameter settings
model.print_algo_args()

# Update algo args for GPU
model.update_model_parameters(
    task_type='GPU',
    bootstrap_type='Bayesian',
    rsm=1.0,
    iterations=200,
    subsample=None,
    sampling_frequency=None
)

# Train Model
model.train()

# Score train, validation, and test; store internally
model.score()

# Inspect scored data
model.ScoredData["train"]
model.ScoredData["validation"]
model.ScoredData["test"]

# Evaluate scored data
global_eval = model.evaluate(
    DataName="test"
)

# Per segment
segment_eval = model.evaluate(
    DataName="test",
    ByVariables="MarketingSegments"
)

# Get variable importance
imp = model.compute_feature_importance()

# Get interaction importance
interact = model.compute_catboost_interaction_importance()

XGBoost Examples

Regression Training

# Setup Environment
import os
import polars as pl
from PolarsFE import datasets, character
from QuickEcharts import Charts
from retrofit import MachineLearning as ml

# Generate demo data instead of reading packaged CSVs
df = utils.make_retrofit_demo_data(
    n_rows=50_000,
    n_segments=5,
    seed=42,
)

# Get TrainData, ValidationData, and TestData
DataSets = datasets.partition_random(
    data=df,
    num_partitions=3,
    seed=42,
    percentages=[0.7, 0.2, 0.1]
)

# Create target encodings for categorical variables
categorical_cols = ['MarketingSegments', 'MarketingSegments2', 'MarketingSegments3']
output = character.categorical_encoding(
    data=DataSets[0],
    ML_Type="regression",
    group_variables=categorical_cols,
    target_variable="Leads",
    method="target_encoding",
    scoring=False,
    keep_original_factors=False
)

# Collect data and encodings
DataSets[0] = output['data']
encodings = output['factor_components']

# Note parameter: scoring=True
DataSets[1] = character.categorical_encoding(
    data=DataSets[1],
    ML_Type="regression",
    group_variables=categorical_cols,
    target_variable="Leads",
    method="target_encoding",
    scoring=True,
    supply_factor_level_list=encodings,
    keep_original_factors=False
)

# Note parameter: scoring=True
DataSets[2] = character.categorical_encoding(
    data=DataSets[2],
    ML_Type="regression",
    group_variables=categorical_cols,
    target_variable="Leads",
    method="target_encoding",
    scoring=True,
    supply_factor_level_list=encodings,
    keep_original_factors=False
)

# Initialize RetroFit
model = ml.RetroFit(TargetType="regression", Algorithm="xgboost")

# Create algo-specific model data
model.create_model_data(
  TrainData=DataSets[0],
  ValidationData=DataSets[1],
  TestData=DataSets[2],
  TargetColumnName="Leads",
  NumericColumnNames=[
    'XREGS1',
    'XREGS2',
    'XREGS3',
    'MarketingSegments_TargetEncode',
    'MarketingSegments2_TargetEncode',
    'MarketingSegments3_TargetEncode'
  ],
  Threads=-1
)

# Print default parameter settings
model.print_algo_args()

# Update algo args
model.update_model_parameters(
    num_boost_round=200,
    num_parallel_tree=4,
    max_depth=4
)

# Train Model
model.train()

# Score train, validation, and test; store internally
model.score()

# Inspect scored data
model.ScoredData["train"]
model.ScoredData["validation"]
model.ScoredData["test"]

# Evaluate scored data
global_eval = model.evaluate(
    DataName="test"
)

# Per segment
segment_eval = model.evaluate(
    DataName="test",
    ByVariables="MarketingSegments_TargetEncode"
)

# Get variable importance
imp = model.compute_feature_importance()

# Model Calibration Tables
cal = model.build_regression_calibration_table(
    DataName="test",
    binning="quantile"
)

# Store plot in working directory
model.plot_regression_calibration(
    DataName="test",
    n_bins=20,
    binning="quantile",
    plot_name=f"{os.getcwd()}/my_calibration_plot"
)

# Numeric Partial Dependence plot
model.plot_pdp_numeric(
    feature='XREGS1',
    DataName="test",
    plot_name=f"{os.getcwd()}/numeric_pdp",
)

# Categorical Partial Dependence plot
model.plot_pdp_categorical(
    feature='MarketingSegments',
    DataName="test",
    plot_name=f"{os.getcwd()}/categorical_pdp",
)

Classification Training

# Setup Environment
import os
import polars as pl
from PolarsFE import datasets, character
from QuickEcharts import Charts
from retrofit import MachineLearning as ml

# Generate demo data instead of reading packaged CSVs
df = utils.make_retrofit_demo_data(
    n_rows=50_000,
    n_segments=5,
    seed=42,
)

# Turn Label into a binary target variable
df = df.with_columns(
    pl.when(pl.col("XREGS1") > 200).then(1)
      .otherwise(0)
      .alias("Label")
)

# Get TrainData, ValidationData, and TestData
DataSets = datasets.partition_random(
    data=df,
    num_partitions=3,
    seed=42,
    percentages=[0.7, 0.2, 0.1]
)

# Create target encodings for categorical variables
categorical_cols = ['MarketingSegments', 'MarketingSegments2', 'MarketingSegments3']
output = character.categorical_encoding(
    data=DataSets[0],
    ML_Type="classification",
    group_variables=categorical_cols,
    target_variable="Leads",
    method="target_encoding",
    scoring=False,
    keep_original_factors=False
)

# Collect data and encodings
DataSets[0] = output['data']
encodings = output['factor_components']

# Note parameter: scoring=True
DataSets[1] = character.categorical_encoding(
    data=DataSets[1],
    ML_Type="classification",
    group_variables=categorical_cols,
    target_variable="Leads",
    method="target_encoding",
    scoring=True,
    supply_factor_level_list=encodings,
    keep_original_factors=False
)

# Note parameter: scoring=True
DataSets[2] = character.categorical_encoding(
    data=DataSets[2],
    ML_Type="classification",
    group_variables=categorical_cols,
    target_variable="Label",
    method="target_encoding",
    scoring=True,
    supply_factor_level_list=encodings,
    keep_original_factors=False
)

# Initialize RetroFit
model = ml.RetroFit(TargetType="classification", Algorithm="xgboost")

# Create algo-specific model data
model.create_model_data(
  TrainData=DataSets[0],
  ValidationData=DataSets[1],
  TestData=DataSets[2],
  TargetColumnName="Label",
  NumericColumnNames=[
    'XREGS1',
    'XREGS2',
    'XREGS3',
    'MarketingSegments_TargetEncode',
    'MarketingSegments2_TargetEncode',
    'MarketingSegments3_TargetEncode'
  ],
  Threads=-1
)

# Print default parameter settings
model.print_algo_args()

# Update algo args
model.update_model_parameters(
    num_boost_round=200,
    num_parallel_tree=4,
    max_depth=4
)

# Train Model
model.train()

# Score train, validation, and test; store internally
model.score()

# Inspect scored data
model.ScoredData["train"]
model.ScoredData["validation"]
model.ScoredData["test"]

# Evaluate scored data
global_eval = model.evaluate(
    DataName="test"
)

# Per segment
segment_eval = model.evaluate(
    DataName="test",
    ByVariables="MarketingSegments_TargetEncode"
)

# Get variable importance
imp = model.compute_feature_importance()

# Store plot in working directory
model.plot_classification_calibration(
    DataName="test",
    n_bins=20,
    binning="quantile",
    plot_name=f"{os.getcwd()}/my_calibration_plot"
)

# ROC Plot
model.plot_classification_roc(
    DataName="test",
    plot_name=f"{os.getcwd()}/my_roc_plot",
)

# PR Plot
model.plot_classification_pr(
    DataName="test",
    plot_name=f"{os.getcwd()}/my_pr_plot"
)

# Numeric Partial Dependence plot
model.plot_pdp_numeric(
    feature='XREGS1',
    DataName="test",
    plot_name=f"{os.getcwd()}/numeric_pdp",
)

# Categorical Partial Dependence plot
model.plot_pdp_categorical(
    feature='MarketingSegments',
    DataName="test",
    plot_name=f"{os.getcwd()}/categorical_pdp",
)

MultiClass Training

# Setup Environment
import os
import polars as pl
from PolarsFE import datasets, character
from QuickEcharts import Charts
from retrofit import MachineLearning as ml

# Generate demo data instead of reading packaged CSVs
df = utils.make_retrofit_demo_data(
    n_rows=50_000,
    n_segments=5,
    seed=42,
)

# Get TrainData, ValidationData, and TestData
DataSets = datasets.partition_random(
    data=df,
    num_partitions=3,
    seed=42,
    percentages=[0.7, 0.2, 0.1]
)

# Create target encodings for categorical variables
categorical_cols = ['MarketingSegments', 'MarketingSegments2', 'MarketingSegments3']
output = character.categorical_encoding(
    data=DataSets[0],
    ML_Type="multiclass",
    group_variables=categorical_cols,
    target_variable="Leads",
    method="target_encoding",
    scoring=False,
    keep_original_factors=False
)

# Collect data and encodings
DataSets[0] = output['data']
encodings = output['factor_components']

# Note parameter: scoring=True
DataSets[1] = character.categorical_encoding(
    data=DataSets[1],
    ML_Type="multiclass",
    group_variables=categorical_cols,
    target_variable="Leads",
    method="target_encoding",
    scoring=True,
    supply_factor_level_list=encodings,
    keep_original_factors=False
)

# Note parameter: scoring=True
DataSets[2] = character.categorical_encoding(
    data=DataSets[2],
    ML_Type="multiclass",
    group_variables=categorical_cols,
    target_variable="Label",
    method="target_encoding",
    scoring=True,
    supply_factor_level_list=encodings,
    keep_original_factors=False
)

# Initialize RetroFit
model = ml.RetroFit(TargetType="multiclass", Algorithm="xgboost")

# Model Variables
drop_cols = ['CalendarDateColumn', 'Label']
features = [c for c in DataSets[2].columns if c not in drop_cols]

# Create algo-specific model data
model.create_model_data(
  TrainData=DataSets[0],
  ValidationData=DataSets[1],
  TestData=DataSets[2],
  TargetColumnName="Label",
  NumericColumnNames=features,
  Threads=-1
)

# Print default parameter settings
model.print_algo_args()

# Update algo args
model.update_model_parameters(
    num_boost_round=200,
    num_parallel_tree=4,
    max_depth=4
)

# Train Model
model.train()

# Score train, validation, and test; store internally
model.score()

# Inspect scored data
model.ScoredData["train"]
model.ScoredData["validation"]
model.ScoredData["test"]

# Evaluate scored data
global_eval = model.evaluate(
    DataName="test"
)

# Get variable importance
imp = model.compute_feature_importance()

LightGBM Examples

Regression Training

# Setup Environment
import os
import polars as pl
from PolarsFE import datasets, character
from QuickEcharts import Charts
from retrofit import MachineLearning as ml

# Generate demo data instead of reading packaged CSVs
df = utils.make_retrofit_demo_data(
    n_rows=50_000,
    n_segments=5,
    seed=42,
)

# Get TrainData, ValidationData, and TestData
DataSets = datasets.partition_random(
    data=df,
    num_partitions=3,
    seed=42,
    percentages=[0.7, 0.2, 0.1]
)

# Create target encodings for categorical variables
categorical_cols = ['MarketingSegments', 'MarketingSegments2', 'MarketingSegments3']
output = character.categorical_encoding(
    data=DataSets[0],
    ML_Type="regression",
    group_variables=categorical_cols,
    target_variable="Leads",
    method="target_encoding",
    scoring=False,
    keep_original_factors=False
)

# Collect data and encodings
DataSets[0] = output['data']
encodings = output['factor_components']

# Note parameter: scoring=True
DataSets[1] = character.categorical_encoding(
    data=DataSets[1],
    ML_Type="regression",
    group_variables=categorical_cols,
    target_variable="Leads",
    method="target_encoding",
    scoring=True,
    supply_factor_level_list=encodings,
    keep_original_factors=False
)

# Note parameter: scoring=True
DataSets[2] = character.categorical_encoding(
    data=DataSets[2],
    ML_Type="regression",
    group_variables=categorical_cols,
    target_variable="Leads",
    method="target_encoding",
    scoring=True,
    supply_factor_level_list=encodings,
    keep_original_factors=False
)

# Initialize RetroFit
model = ml.RetroFit(TargetType="regression", Algorithm="lightgbm")

# Create algo-specific model data
model.create_model_data(
  TrainData=DataSets[0],
  ValidationData=DataSets[1],
  TestData=DataSets[2],
  TargetColumnName="Leads",
  NumericColumnNames=[
    'XREGS1',
    'XREGS2',
    'XREGS3',
    'MarketingSegments_TargetEncode',
    'MarketingSegments2_TargetEncode',
    'MarketingSegments3_TargetEncode'
  ],
  Threads=-1
)

# Print default parameter settings
model.print_algo_args()

# Update algo args
model.update_model_parameters(
    num_iterations=200,
    max_depth=6,
    min_data_in_leaf=2
)

# Train Model
model.train()

# Score train, validation, and test; store internally
model.score()

# Inspect scored data
model.ScoredData["train"]
model.ScoredData["validation"]
model.ScoredData["test"]

# Evaluate scored data
global_eval = model.evaluate(
    DataName="test"
)

# Per segment
segment_eval = model.evaluate(
    DataName="test",
    ByVariables="MarketingSegments_TargetEncode"
)

# Get variable importance
imp = model.compute_feature_importance()

# Model Calibration Tables
cal = model.build_regression_calibration_table(
    DataName="test",
    binning="quantile"
)

# Store plot in working directory
model.plot_regression_calibration(
    DataName="test",
    n_bins=20,
    binning="quantile",
    plot_name=f"{os.getcwd()}/my_calibration_plot"
)

# Numeric Partial Dependence plot
model.plot_pdp_numeric(
    feature='XREGS1',
    DataName="test",
    plot_name=f"{os.getcwd()}/numeric_pdp",
)

# Categorical Partial Dependence plot
model.plot_pdp_categorical(
    feature='MarketingSegments',
    DataName="test",
    plot_name=f"{os.getcwd()}/categorical_pdp",
)

Classification Training

# Setup Environment
import os
import polars as pl
from PolarsFE import datasets, character
from QuickEcharts import Charts
from retrofit import MachineLearning as ml

# Generate demo data instead of reading packaged CSVs
df = utils.make_retrofit_demo_data(
    n_rows=50_000,
    n_segments=5,
    seed=42,
)

# Turn Label into a binary target variable
df = df.with_columns(
    pl.when(pl.col("XREGS1") > 200).then(1)
      .otherwise(0)
      .alias("Label")
)

# Get TrainData, ValidationData, and TestData
DataSets = datasets.partition_random(
    data=df,
    num_partitions=3,
    seed=42,
    percentages=[0.7, 0.2, 0.1]
)

# Create target encodings for categorical variables
categorical_cols = ['MarketingSegments', 'MarketingSegments2', 'MarketingSegments3']
output = character.categorical_encoding(
    data=DataSets[0],
    ML_Type="classification",
    group_variables=categorical_cols,
    target_variable="Leads",
    method="target_encoding",
    scoring=False,
    keep_original_factors=False
)

# Collect data and encodings
DataSets[0] = output['data']
encodings = output['factor_components']

# Note parameter: scoring=True
DataSets[1] = character.categorical_encoding(
    data=DataSets[1],
    ML_Type="classification",
    group_variables=categorical_cols,
    target_variable="Leads",
    method="target_encoding",
    scoring=True,
    supply_factor_level_list=encodings,
    keep_original_factors=False
)

# Note parameter: scoring=True
DataSets[2] = character.categorical_encoding(
    data=DataSets[2],
    ML_Type="classification",
    group_variables=categorical_cols,
    target_variable="Label",
    method="target_encoding",
    scoring=True,
    supply_factor_level_list=encodings,
    keep_original_factors=False
)

# Initialize RetroFit
model = ml.RetroFit(TargetType="classification", Algorithm="lightgbm")

# Create algo-specific model data
model.create_model_data(
  TrainData=DataSets[0],
  ValidationData=DataSets[1],
  TestData=DataSets[2],
  TargetColumnName="Label",
  NumericColumnNames=[
    'XREGS1',
    'XREGS2',
    'XREGS3',
    'MarketingSegments_TargetEncode',
    'MarketingSegments2_TargetEncode',
    'MarketingSegments3_TargetEncode'
  ],
  Threads=-1
)

# Print default parameter settings
model.print_algo_args()

# Update algo args
model.update_model_parameters(
    num_iterations=200,
    max_depth=6,
    min_data_in_leaf=2
)

# Train Model
model.train()

# Score train, validation, and test; store internally
model.score()

# Inspect scored data
model.ScoredData["train"]
model.ScoredData["validation"]
model.ScoredData["test"]

# Evaluate scored data
global_eval = model.evaluate(
    DataName="test"
)

# Per segment
segment_eval = model.evaluate(
    DataName="test",
    ByVariables="MarketingSegments_TargetEncode"
)

# Get variable importance
imp = model.compute_feature_importance()

# Store plot in working directory
model.plot_classification_calibration(
    DataName="test",
    n_bins=20,
    binning="quantile",
    plot_name=f"{os.getcwd()}/my_calibration_plot"
)

# ROC Plot
model.plot_classification_roc(
    DataName="test",
    plot_name=f"{os.getcwd()}/my_roc_plot",
)

# PR Plot
model.plot_classification_pr(
    DataName="test",
    plot_name=f"{os.getcwd()}/my_pr_plot"
)

# Numeric Partial Dependence plot
model.plot_pdp_numeric(
    feature='XREGS1',
    DataName="test",
    plot_name=f"{os.getcwd()}/numeric_pdp",
)

# Categorical Partial Dependence plot
model.plot_pdp_categorical(
    feature='MarketingSegments',
    DataName="test",
    plot_name=f"{os.getcwd()}/categorical_pdp",
)

MultiClass Training

# Setup Environment
import os
import polars as pl
from PolarsFE import datasets, character
from QuickEcharts import Charts
from retrofit import MachineLearning as ml

# Generate demo data instead of reading packaged CSVs
df = utils.make_retrofit_demo_data(
    n_rows=50_000,
    n_segments=5,
    seed=42,
)

# Get TrainData, ValidationData, and TestData
DataSets = datasets.partition_random(
    data=df,
    num_partitions=3,
    seed=42,
    percentages=[0.7, 0.2, 0.1]
)

# Create target encodings for categorical variables
categorical_cols = ['MarketingSegments', 'MarketingSegments2', 'MarketingSegments3']
output = character.categorical_encoding(
    data=DataSets[0],
    ML_Type="multiclass",
    group_variables=categorical_cols,
    target_variable="Leads",
    method="target_encoding",
    scoring=False,
    keep_original_factors=False
)

# Collect data and encodings
DataSets[0] = output['data']
encodings = output['factor_components']

# Note parameter: scoring=True
DataSets[1] = character.categorical_encoding(
    data=DataSets[1],
    ML_Type="multiclass",
    group_variables=categorical_cols,
    target_variable="Leads",
    method="target_encoding",
    scoring=True,
    supply_factor_level_list=encodings,
    keep_original_factors=False
)

# Note parameter: scoring=True
DataSets[2] = character.categorical_encoding(
    data=DataSets[2],
    ML_Type="multiclass",
    group_variables=categorical_cols,
    target_variable="Label",
    method="target_encoding",
    scoring=True,
    supply_factor_level_list=encodings,
    keep_original_factors=False
)

# Initialize RetroFit
model = ml.RetroFit(TargetType="multiclass", Algorithm="lightgbm")

# Model Variables
drop_cols = ['CalendarDateColumn', 'Label']
features = [c for c in DataSets[2].columns if c not in drop_cols]

# Create algo-specific model data
model.create_model_data(
  TrainData=DataSets[0],
  ValidationData=DataSets[1],
  TestData=DataSets[2],
  TargetColumnName="Label",
  NumericColumnNames=features,
  Threads=-1
)

# Print default parameter settings
model.print_algo_args()

# Update algo args
model.update_model_parameters(
    num_iterations=200,
    max_depth=6,
    min_data_in_leaf=2
)

# Train Model
model.train()

# Score train, validation, and test; store internally
model.score()

# Inspect scored data
model.ScoredData["train"]
model.ScoredData["validation"]
model.ScoredData["test"]

# Evaluate scored data
global_eval = model.evaluate(
    DataName="test"
)

# Get variable importance
imp = model.compute_feature_importance()

---

2. Model Evaluation Visuals

Below is a gallery of example evaluation plots produced by RetroFit.

Click to expand gallery