ghostjat/pml

High-performance Machine Learning and Deep Learning framework for PHP. Powered by FFI, OpenBLAS, LibSVM, and XGBoost with a scikit-learn compatible API.

dev-main 2026-03-25 11:39 UTC

Requires

• php: ^8.1
• ext-ffi: *

MIT

• Shubham Chaudhary

• php
• svm
• machine-learning
• tensor
• deep-learning
• scikit-learn
• data-science
• ffi
• neural-networks
• openblas
• xgboost

README

Pml is a pure-PHP, enterprise-grade Machine Learning and Deep Learning framework. It bypasses PHP's native array limitations by using the Foreign Function Interface (FFI) to orchestrate raw C-memory (Float32 Tensors) directly through OpenBLAS, LibSVM, and XGBoost.

Whether you want to build a scikit-learn style Pipeline for classic machine learning, or train a custom Multi-Layer Perceptron using a dynamic Autograd engine, Pml brings true, C-speed AI to the PHP backend.

⚡ Key Features

• Scikit-Learn Parity: Pml\Classic provides an identical API to Python's scikit-learn (fit, predict, fit_transform).
• Blazing Fast Math: Zero-copy 1D Float32 flat-buffers routed directly into libopenblas.so for hardware-accelerated matrix multiplication (sgemm).
• Deep Learning Autograd: A fully-featured dynamic computation graph, supporting backpropagation, Cross-Entropy Loss, and AdamW optimization.
• Native Integrations: Direct FFI bindings to industry-standard C libraries: libsvm and libxgboost.
• Memory Safe: Rigorously tested with Resident Set Size (RSS) trackers. Guaranteed no C-pointer memory leaks during garbage collection.
• FFI-Safe Serialization: Save and load models instantly using Joblib with bit-identical float preservation.

🚀 Installation

Pml requires PHP 8.1+ with the FFI extension enabled, and underlying OS-level C libraries.

1. Install System Dependencies (Ubuntu/Debian):

sudo apt-get update
sudo apt-get install libopenblas-dev libsvm-dev libxgboost-dev

2. Enable FFI in your php.ini:

ffi.enable=true

3. Install via Composer:

composer require ghostjat/pml

📖 Quick Start

Classic ML: The Pipeline API

Clean data, scale features, and train a Support Vector Machine in 5 lines of code.

use Pml\Classic\Pipeline\make_pipeline;
use Pml\Classic\Preprocess\StandardScaler;
use Pml\Classic\Preprocess\SimpleImputer;
use Pml\Classic\SVM\SVC;

// 1. Build the pipeline
$pipeline = make_pipeline(
    new SimpleImputer(strategy: 'mean'),
    new StandardScaler(),
    new SVC(kernel: 'rbf', C: 1.0)
);

// 2. Train on Tensors
$pipeline->fit($X_train, $y_train);

// 3. Predict
$predictions = $pipeline->predict($X_test);

Deep Learning: The Autograd Engine

Train a Neural Network using AdamW.

use Pml\Classic\NeuralNetwork\MLPClassifier;

$model = new MLPClassifier(
    hidden_layer_sizes: [64, 32],
    activation: 'relu',
    solver: 'adamw',
    learning_rate: 1e-3,
    max_iter: 200
);

$model->fit($X_train, $y_train);

🧰 Supported Algorithms

Linear Models

• LogisticRegression (Binary & Multinomial Auto-routing)
• Ridge, Lasso, ElasticNet, LinearRegression

Ensembles & Trees

• RandomForestClassifier, DecisionTreeClassifier
• AdaBoostClassifier, BaggingClassifier
• XGBClassifier (via libxgboost)

Support Vector Machines

• SVC, SVR (via libsvm)

Clustering & Distance

• KMeans
• DBSCAN
• KNeighborsClassifier

Preprocessing & Tuning

• StandardScaler, MinMaxScaler, SimpleImputer, OneHotEncoder
• PCA, VarianceThreshold
• GridSearchCV, cross_val_score, Pipeline

🚧 Limitations

• Float32 Only: To maximize BLAS throughput and keep FFI footprints small, all internal tensors operate strictly on 32-bit floats.
• CPU Bound: Currently, matrix operations are handled by CPU-based OpenBLAS.
• Data Size: While highly optimized, datasets larger than available RAM will require chunked batching, as Pml currently loads Tensors entirely into memory.

🔮 Roadmap / Future To-Do

We are constantly pushing the boundaries of what PHP can do. Upcoming features include:

• GPU Acceleration: Implement libcudart and cublas bindings for NVIDIA GPU support.
• DataFrames: Integrate Apache Arrow or Polars FFI bindings to replace standard PHP arrays for data loading.
• Advanced Metrics: Expand the Metrics module to include ROC-AUC, F1-Score, and Confusion Matrix generation.
• Model Checkpointing: Epoch-based save states for the Deep Learning Autograd engine.

🤝 Contributing

Contributions are welcome! Please ensure that all pull requests pass the rigorous memory-leak test suite.

php tests/run_e2e.php

📄 License

Pml is open-sourced software licensed under the MIT license.