AutoML Trainer

A no-code machine learning web application to upload spreadsheet data, train multiple models, compare performance, and generate predictions through UI and API.

---

---

✨ Features

• 📤 Upload CSV/XLSX/XLS datasets with instant data-quality analysis
• 🧹 Automatic preprocessing (missing values, duplicates, type handling)
• 🤖 Multi-model training for both classification and regression
• 🎯 Smart model recommendations with comparison charts
• 💾 Download trained .joblib models and cleaned datasets
• 🔮 Single-record and batch-file prediction workflows
• 📚 Built-in OpenAPI + Swagger docs for prediction endpoints
• 🧾 Prediction history tracking and clear/reset actions
• 🧼 Artifact cleanup endpoint + helper script

📦 Tech Stack

• Backend: Flask
• ML: scikit-learn, NumPy, pandas
• Visualization: Matplotlib, Chart.js (CDN)
• API Docs: Swagger UI (CDN)
• Optional async deps included: Celery + Redis

🧠 Supported Models

Classification

• Random Forest
• Gradient Boosting
• Logistic Regression
• Support Vector Machine (SVM)
• K-Nearest Neighbors (KNN)
• Decision Tree

Regression

• Random Forest Regressor
• Gradient Boosting Regressor
• Linear Regression
• Support Vector Regressor (SVR)
• Decision Tree Regressor
• Elastic Net

🏗️ Architecture

---

flowchart TD
    A[Browser UI\nJinja + CSS + JS] --> B[Flask App\napp.py]
    B --> C[Upload + Quality Analysis\n/api/data-quality]
    B --> D[Preprocessing + Model Selection]
    D --> E[Training Pipeline\nscikit-learn + RandomizedSearchCV]
    E --> F[(models/*.joblib)]
    D --> G[(cleaned_data/*)]

    F --> H[Prediction UI\n/predict/<model>]
    F --> I[Prediction API\n/api/predict/<model>]
    H --> J[(prediction_history/*.jsonl)]
    I --> J

    B --> K[Model Insights\n/api/model-insights/<model>]
    B --> L[OpenAPI Spec\n/api/openapi.json]

Loading

⚙️ Model Training Pipeline

Data preparation

• Upload accepts csv, xlsx, xls
• Data-quality report computes missing values, duplicates, data types, and per-column stats
• Cleaning removes duplicate rows and rows with missing target values
• Problem type is auto-detected from target dtype/distribution (classification vs regression)
• Training is capped by MAX_TRAIN_ROWS to avoid oversized jobs

Feature preprocessing pipeline

• Numeric features: median imputation
• Scaled numeric models (svm, svr, knn, logistic_regression, linear_regression, elastic_net) add StandardScaler
• Categorical features: most-frequent imputation + OneHotEncoder(handle_unknown="ignore")
• Combined with ColumnTransformer and wrapped in a single scikit-learn Pipeline

Model training and selection

• Train/validation split is automatic
• Classification uses stratified split when feasible and class-aware handling
• Optional hyperparameter tuning uses RandomizedSearchCV (ENABLE_HYPERPARAM_TUNING=True)
• Cross-validation folds/iterations are bounded by app config:
• HYPERPARAM_SEARCH_MAX_ITER
• HYPERPARAM_SEARCH_CV_MAX_FOLDS
• Best estimator is persisted as .joblib

Metrics and artifacts

• Classification metrics: accuracy, precision, recall, F1
• Regression metrics: R2, MAE, RMSE
• Saved model payload includes:
• preprocessing + estimator pipeline
• feature columns and feature dtypes
• inferred feature value hints
• target classes/encoder metadata (classification)
• training metrics and configuration

Prediction pipeline

• UI and API inference both load the same saved .joblib pipeline
• Input validation is dtype-aware and checks required feature columns
• Batch prediction endpoint supports uploaded files
• Prediction history is stored in prediction_history/*.jsonl

📁 Project Structure

ML-web/
|-- app.py
|-- requirements.txt
|-- scripts/
|   `-- cleanup_artifacts.py
|-- templates/
|   |-- base.html
|   |-- index.html
|   |-- upload.html
|   |-- train.html
|   |-- select_models.html
|   |-- results_comparison.html
|   |-- models.html
|   |-- predict.html
|   `-- api_docs.html
|-- static/
|   |-- css/style.css
|   |-- js/script.js
|   `-- images/
|-- uploads/
|-- cleaned_data/
|-- models/
`-- prediction_history/

🚀 Quick Start

1. Clone and enter project

git clone <your-repo-url>
cd ML-web

2. Create virtual environment

python -m venv .venv

Windows (PowerShell):

.\.venv\Scripts\Activate.ps1

macOS/Linux:

source .venv/bin/activate

3. Install dependencies

pip install -r requirements.txt

4. Run the app

python app.py

Open: http://127.0.0.1:5000

🔄 End-to-End Workflow

1. Upload a dataset on /upload
2. Inspect quality report and choose target column
3. Select recommended models
4. Train and compare metrics
5. Download best model and cleaned data
6. Run predictions from /predict/<model_filename> or API

🔌 Key Endpoints

• GET / - Landing page
• GET|POST /upload - Upload dataset + training configuration
• POST /select-models - Model recommendation page
• POST /train-models - Train selected models
• GET /models - Saved model library
• GET|POST /predict/<filename> - Interactive prediction workspace
• POST /api/predict/<filename> - JSON prediction API
• POST /predict-file/<filename> - Batch predictions from uploaded file
• GET /api/model-insights/<filename> - Model metrics + feature importance
• GET /api/openapi.json - OpenAPI document
• GET /api/docs - Swagger UI docs
• GET /api/predict-history/<filename> - Prediction history
• POST /api/predict-history/<filename>/clear - Clear history
• POST /admin/cleanup-artifacts - Cleanup generated artifacts

🖼️ Screenshots

⚙️ Configuration Notes

Default runtime settings in app.py include:

• MAX_TRAIN_ROWS = 50000
• ENABLE_HYPERPARAM_TUNING = True
• HYPERPARAM_SEARCH_MAX_ITER = 12
• HYPERPARAM_SEARCH_CV_MAX_FOLDS = 5
• Auto-cleanup enabled after training with retention limits

Optional env var:

• CLEANUP_TOKEN for securing /admin/cleanup-artifacts

🧼 Cleanup Utility

Dry run:

python scripts/cleanup_artifacts.py

Apply deletion:

python scripts/cleanup_artifacts.py --apply

Include uploads cleanup:

python scripts/cleanup_artifacts.py --include-uploads --apply

📌 Notes

• Generated artifacts (uploads/, models/, cleaned_data/) are git-ignored.
• App currently runs with debug=True in app.py; disable for production.