fitted

Machine learning tools

To install: pip install fitted

Documentation here

ml_experiments

Lightweight experiment runner for ML model evaluation with minimal boilerplate.

Features

• Minimal boilerplate: Just provide data and models, get results
• Flexible storage: Pluggable MutableMapping backends for result persistence
• Multiple datasets/models: Automatically runs all combinations
• Comprehensive metrics: Built-in cross-validation with multiple scoring metrics
• AI-friendly: Simple, predictable API for programmatic use
• Functional design: Generator-based, dependency injection, SSOT principles

Quick Start

from sklearn.datasets import load_iris
from sklearn.ensemble import RandomForestClassifier
from sklearn.svm import SVC
from fitted.ml_experiments import ExperimentRunner

# Load data
iris = load_iris()
X, y = iris.data, iris.target

# Run experiments
runner = ExperimentRunner()
results = runner.run(
    datasets={'iris': (X, y)},
    models={
        'rf': RandomForestClassifier(random_state=42),
        'svm': SVC(random_state=42),
    }
)

# Access results
print(results['iris/rf']['accuracy_test'])
# {'mean': 0.96, 'std': 0.02, 'min': 0.93, 'max': 0.98}

Core Concepts

ExperimentRunner

The main entry point for running experiments. Initialize once, run many times.

runner = ExperimentRunner(
    project_store='my_experiments',  # str, MutableMapping, or None
    config=ExperimentConfig(n_splits=5, verbose=True)
)

Project Store Options:

• None: Uses temporary directory (printed at start)
• str: Project name → creates persistent store in your default projects directory
• MutableMapping: Any custom key-value store

To see what your default projects directory is, see

from fitted.ml_experiments import DFLT_PROJECTS_DIR
print(DFLT_PROJECTS_DIR)

It lives in fitted's app folder (not in the current directory). You can change this default by specifying a different one in the FITTED_DFLT_EXP_PROJECTS_DIR environment variable (use an absolute path; a relative one is resolved against the current directory).

ExperimentConfig

Configuration for experiment runs:

from fitted.ml_experiments import ExperimentConfig

config = ExperimentConfig(
    n_splits=5,              # Cross-validation splits
    test_size=0.2,           # Train/test split ratio
    random_state=42,         # For reproducibility
    scoring='accuracy',      # Single metric or list
    verbose=True,            # Print progress
    n_jobs=-1               # Parallel CV folds
)

Result Structure

Each experiment result is a dict with:

{
    'config': {...},                    # Experiment configuration
    'timestamp': '2025-10-10T14:30:00', # When run
    'duration_seconds': 1.23,           # Runtime
    'n_samples': 150,                   # Dataset size
    'n_features': 4,                    # Feature count
    'accuracy_test': {                  # Stats for each metric
        'mean': 0.96,
        'std': 0.02,
        'min': 0.93,
        'max': 0.98
    },
    'accuracy_train': {...},            # Training scores
    'cv_results_raw': {...}             # Raw sklearn output
}

Usage Patterns

Multiple Datasets

datasets = {
    'iris': (X_iris, y_iris),
    'wine': (X_wine, y_wine),
    'cancer': (X_cancer, y_cancer),
}

results = runner.run(datasets=datasets, models=models)

Multiple Metrics

runner = ExperimentRunner(
    config=ExperimentConfig(
        scoring=['accuracy', 'precision', 'recall', 'f1', 'roc_auc']
    )
)

Config Overrides

# Default: 5 splits
runner = ExperimentRunner(config=ExperimentConfig(n_splits=5))

# Override for specific run
results = runner.run(datasets, models, n_splits=10, verbose=False)

Iterable Inputs (Alternative to Dicts)

datasets = [
    ('dataset1', (X1, y1)),
    ('dataset2', (X2, y2)),
]

models = [
    ('model1', RandomForestClassifier()),
    ('model2', SVC()),
]

results = runner.run(datasets=datasets, models=models)

Persistence and Reloading

# First session: run and save
runner = ExperimentRunner(project_store='my_project')
runner.run(datasets, models)

# Later session: reload results
from fitted.ml_experiments import load_experiment_results

results = load_experiment_results('my_project')
for key in results:
    print(f"{key}: {results[key]['accuracy_test']['mean']:.3f}")

Result Analysis

from fitted.ml_experiments import summarize_results

# Get sorted summary
for key, stats in summarize_results(results, metric='accuracy'):
    dataset, model = key.split('/')
    print(f"{dataset} × {model}: {stats['mean']:.4f} ± {stats['std']:.4f}")

Custom Storage Backends

You can provide any MutableMapping as the project store:

class S3Store(MutableMapping):
    """Store results in S3."""
    def __init__(self, bucket_name):
        self.bucket = boto3.resource('s3').Bucket(bucket_name)
    
    def __setitem__(self, key, value):
        self.bucket.put_object(
            Key=key,
            Body=dill.dumps(value)
        )
    
    def __getitem__(self, key):
        obj = self.bucket.Object(key).get()
        return dill.loads(obj['Body'].read())
    
    # ... implement __delitem__, __iter__, __len__

# Use it
runner = ExperimentRunner(project_store=S3Store('my-bucket'))

AI Assistant Usage

This tool is designed to be easily used by AI assistants. Example prompt:

I have this dataset [description]. Please:
1. Suggest appropriate models
2. Run experiments using ml_experiments
3. Analyze results and give recommendations

The AI can programmatically:

• Determine appropriate models based on data characteristics
• Construct the models dict
• Run experiments
• Parse and analyze results
• Generate reports

Architecture

The library follows these principles:

• Functional over OOP: Prefer generators and pure functions
• SOLID principles: Single responsibility, dependency injection
• Mapping interface: Results are MutableMapping instances
• Generator-based: Lazy evaluation where appropriate
• Minimal dependencies: Only sklearn, dol, dill, numpy

API Reference

Core Classes

ExperimentRunner

• __init__(*, project_store, config)
• run(datasets, models, **config_overrides) -> MutableMapping
• project_store property

ExperimentConfig (dataclass)

• n_splits: int = 5
• test_size: float = 0.2
• random_state: int | None = None
• scoring: str | list[str] = 'accuracy'
• verbose: bool = True
• n_jobs: int = -1

Utility Functions

load_experiment_results(project_store)

• Load results from a saved project

summarize_results(results, *, metric, sort_by)

• Generate sorted summary statistics

uri_to_project_store(name, *, projects_dir)

• Convert project name to MutableMapping store

Examples

See examples.py for comprehensive usage examples including:

• Minimal usage
• Multiple datasets
• Multiple metrics
• Config overrides
• Persistence
• AI-friendly programmatic generation
• Custom storage backends

Testing

python test_ml_experiments.py