neuralfabric 0.1.0.dev2

A from-scratch machine learning and deep learning framework, from linear regression to transformers.

NeuralFabric

A from-scratch machine learning and deep learning framework, from linear regression to transformers.

Installation

pip install neuralfabric
# or, for local development:
pip install -e ".[dev]"

Quick start (target API — see roadmap)

from neuralfabric.linear_model import LinearRegression
from neuralfabric.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

model = LinearRegression()
model.fit(X_train, y_train)
preds = model.predict(X_test)

Project structure

src/neuralfabric/
├── core/             # Tensor + autograd engine (PyTorch-style backbone)
├── linear_model/      # Linear/Logistic/Ridge/Lasso/ElasticNet
├── tree/               # Decision trees
├── ensemble/           # RandomForest, GradientBoosting, AdaBoost, Bagging
├── svm/                # Support Vector Machines
├── naive_bayes/        # Gaussian / Multinomial Naive Bayes
├── cluster/            # KMeans, DBSCAN, hierarchical clustering
├── decomposition/      # PCA, SVD
├── nn/                 # Module, Parameter, layers, activations, losses
│   └── layers/         # Linear, Conv, Pooling, Normalization, RNN/LSTM/GRU...
├── transformer/         # Attention, positional encoding, encoder/decoder, full models
├── optim/              # SGD, Adam, RMSProp, LR schedulers
├── preprocessing/      # Scalers, encoders, imputers
├── model_selection/    # train_test_split, cross-validation, grid search
├── metrics/            # Regression / classification / clustering metrics
├── datasets/           # Toy dataset loaders
├── utils/              # Validation + math helpers
├── base.py             # BaseEstimator, RegressorMixin, ClassifierMixin, TransformerMixin
└── pipeline.py          # Pipeline / FeatureUnion

Roadmap

• core: Tensor + autograd engine
• linear_model: Linear & Logistic Regression
• tree / ensemble: Decision Tree, Random Forest, Gradient Boosting
• svm, naive_bayes, cluster, decomposition
• nn: Module system, core layers, losses, activations
• optim: SGD, Adam
• CNN layers (nn/layers/conv.py, pooling.py)
• RNN/LSTM/GRU (nn/layers/recurrent.py)
• transformer: attention → encoder/decoder → full model
• Docs site + tutorials
• First PyPI release (v0.1.0)

Development

make dev      # install with dev dependencies
make test     # run tests with coverage
make lint     # ruff + mypy
make format   # black
make build    # build sdist + wheel
make publish  # twine upload (requires PyPI credentials)

License

MIT — see LICENSE.