Optorch v0.1.0 - Initial Release

Optorch v0.1.0

The first public release of Optorch, a NumPy-based differentiable photonic computing framework.

Highlights

• Pure NumPy implementation with zero PyTorch dependency
• GPU acceleration via CuPy (auto-detect, zero code change)
• Physics-based adjoint autodifferentiation
• MZI optical neural network layers (Clements decomposition)
• Diffractive optical layers (D2NN)
• Phase mask layers
• Physical noise modeling for noise-aware training
• Arbitrary matrix auto-decomposition (SVD + Clements)
• Trainer framework with checkpointing and LR schedulers
• Custom gradient registry API

Installation

pip install -e .

Quick Start

import optorch
field = optorch.OpticalField.uniform_plane(shape=(1024, 1024))
field_gpu = field.to_gpu()
model = optorch.OpticalNetwork([
    optorch.MZILayer(n_modes=1024),
    optorch.PhaseMaskLayer(shape=(1024,))
])
output = model.forward(field_gpu)

See README.md for full documentation.

---

中文： Optorch 首个公开发布版本，基于 NumPy 的可微分光计算模拟框架。

核心特性

• 纯 NumPy 实现，零 PyTorch 依赖
• CuPy GPU 加速（自动检测，无需改代码）
• 物理伴随法自动微分
• MZI 光神经网络层（Clements 分解）
• 衍射光学层（D2NN）
• 相位掩模层
• 物理噪声建模（噪声感知训练）
• 任意矩阵自动分解（SVD + Clements）
• Trainer 训练框架（含 checkpoint 和学习率调度）
• 自定义梯度注册 API

完整文档见 README.md。