jaxflow 0.1.0.dev0

A lightweight neural-network library built on JAX

JAXFlow

A lightweight neural-network library built on JAX
– fast imports, pure-functional APIs, and batteries-included for research and production.

---

🚀 Features

• Module API
  Define layers with familiar setup/__call__ style or pure-function transforms.
• PyTree compatibility
  Everything is a JAX PyTree; seamless jit, vmap, pmap, and pjit.
• Rich layer collection
  Dense, Conv, BatchNorm, Dropout, and more in jaxflow.layers.
• Optimizers & Schedulers
  Thin wrappers around Optax in jaxflow.optimizers.
• Activations & Initializers
  relu, gelu, swish, he_normal, glorot_uniform, …
• Losses & Metrics
  Standard losses (mse, cross_entropy) and metrics (accuracy, precision, …).
• Callbacks & Checkpointing
  Training hooks and Orbax-powered jaxflow.checkpt utilities.
• Pre-built Models
  ResNet, Transformer, MLP, and easy to extend in jaxflow.models.
• Modular & Lazy
  Top-level import is lightning-fast; submodules load on demand.

---

📦 Installation

bash pip install jaxflow

Note:

Requires JAX (CPU/GPU/TPU), e.g.

pip install "jax[cuda]>=0.4.0" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

Python ≥3.8

---

🎉 Quickstart

---

📖 Documentation

API Reference: https://mthd98.github.io/JAXFlow/

Cookbook: Layer recipes, advanced transforms, multi-host training

---

🛠️ Structure

jaxflow/ ├── core/ # Variable, RNG contexts ├── gradient/ # Gradient utilities ├── activations/ # relu, gelu, swish, … ├── initializers/ # weight initializers ├── layers/ # Dense, Conv, BatchNorm, … ├── losses/ # mse, cross_entropy, … ├── optimizers/ # Optax wrappers, schedulers ├── callbacks/ # EarlyStopping, Logging, … ├── metrics/ # accuracy, precision, … ├── models/ # ResNet, Transformer, … └── regularizers/ # Dropout, weight decay, …

---

🤝 Contributing

We welcome contributions! Please see our CONTRIBUTING.md for:

1. Setting up a dev environment

2. Code style & linting

3. Testing & CI guidelines

4. How to file issues & propose features

---

📄 License

This project is licensed under the Apache-2.0 License. See the LICENSE file for details.

---

“Simplicity is the ultimate sophistication.” – Leonardo da Vinci With JAXFlow, keep your research code clean, fast, and reproducible.