pytorch-onn

pytorch-onn

基于PyTorch的光子神经网络仿真与优化框架

pytorch-onn是一个基于PyTorch的光子神经网络仿真框架。该框架支持GPU加速的相干和非相干光学神经网络训练与推理,可扩展至百万参数规模。它提供了高度优化的并行处理和多功能API,支持从器件到系统级的协同设计与优化。这一工具主要面向神经形态光子学、光学AI系统和光子集成电路优化等领域的研究人员。

PyTorch光子集成电路神经网络AI计算GPU加速Github开源项目
<p align="center"> <img src="torchonn_logo.jpg" alt="torchonn Logo" width="450"> </p> <h2><p align="center">A PyTorch Library for Photonic Integrated Circuit Simulation and Photonic AI Computing</p></h2>
@inproceedings{jiaqigu2021L2ight,
    title     = {L2ight: Enabling On-Chip Learning for Optical Neural Networks via Efficient in-situ Subspace Optimization},
    author    = {Jiaqi Gu and Hanqing Zhu and Chenghao Feng and Zixuan Jiang and Ray T. Chen and David Z. Pan},
    booktitle = {Conference on Neural Information Processing Systems (NeurIPS)},
    year      = {2021}
}
<h3><p align="center">Fast, Scalable, Easy Customization, Support Hardware-Aware Cross-Layer Co-Design</p></h3> <p align="center"> <a href="https://github.com/JeremieMelo/pytorch-onn/blob/release/LICENSEE"> <img alt="MIT License" src="https://img.shields.io/apm/l/atomic-design-ui.svg?"> </a> </p> <br />

👋 Welcome

What it is doing

Integrated neuromorphic photonics simulation framework based on PyTorch. It supports coherent and incoherent optical neural networks (ONNs) training/inference on GPUs. It can scale up to million-parameter ONNs with efficient implementation.

Who will benefit

Researchers on neuromorphic photonics, optical AI system design, photonic integrated circuit optimization, ONN training/inference.

Features

CUDA-backed fast GPU support, optimized highly-parallel tensorized processing, versatile APIs for device/circuit/architecture/algorithm co-optimization

Contents

<!-- toc --> <!-- tocstop -->

News

  • 09/17/2023: v0.0.6 available. Support add-drop MRR weight bank and initialize ONN layers from standard pytorch Conv2d/Linear!
  • 04/19/2022: v0.0.5 available. Automatic differentiable photonic tensor core search! Support customized coherent photonic SuperMesh construction from basic building blocks! (Gu+, ADEPT DAC 2022)
  • 04/18/2022: v0.0.4 available. Phase change material (PCM)-based photonic in-memory computing with endurance enhancement! (Zhu+, ELight ASP-DAC 2022)
  • 04/18/2022: v0.0.3 available. SqueezeLight architecture based on multi-operand microrings for ultra-compact optical neurocomputing! (Gu+, SqueezeLight DATE 2021)
  • 11/28/2021: v0.0.2 available. FFT-ONN-family is supported with trainable butterfly meshes for area-efficient frequency-domain optical neurocomputing! (Gu+, FFT-ONN ASP-DAC 2020) (Gu+, FFT-ONN-v2 IEEE TCAD 2021) (Feng+, PSNN arXiv 2021)
  • 06/10/2021: v0.0.1 available. MZI-ONN (Shen+, MZI-ONN) is supported. Feedbacks are highly welcomed!

Installation

From Source

Dependencies

  • Python >= 3.6
  • PyTorch >= 1.13.0
  • Tensorflow-gpu >= 2.5.0
  • pyutils >= 0.0.2
  • Others are listed in requirements.txt
  • GPU model training requires NVIDIA GPUs and compatible CUDA

Get the PyTorch-ONN Source

git clone https://github.com/JeremieMelo/pytorch-onn.git

Install PyTorch-ONN

cd pytorch-onn python3 setup.py install --user clean

or

./setup.sh

Usage

Construct optical NN models as simple as constructing a normal pytorch model.

import torch.nn as nn import torch.nn.functional as F import torchonn as onn from torchonn.models import ONNBaseModel class ONNModel(ONNBaseModel): def __init__(self, device=torch.device("cuda:0)): super().__init__(device=device) self.conv = onn.layers.MZIBlockConv2d( in_channels=1, out_channels=8, kernel_size=3, stride=1, padding=1, dilation=1, bias=True, miniblock=4, mode="usv", decompose_alg="clements", photodetect=True, device=device, ) self.pool = nn.AdaptiveAvgPool2d(5) self.linear = onn.layers.MZIBlockLinear( in_features=8*5*5, out_features=10, bias=True, miniblock=4, mode="usv", decompose_alg="clements", photodetect=True, device=device, ) self.conv.reset_parameters() self.linear.reset_parameters() def forward(self, x): x = torch.relu(self.conv(x)) x = self.pool(x) x = x.flatten(1) x = self.linear(x) return x

Features

  • Support pytorch training MZI-based ONNs. Support MZI-based Linear, Conv2d, BlockLinear, and BlockConv2d. Support weight, usv, phase modes and their conversion.
  • Support phase quantization and non-ideality injection, including phase shifter gamma error, phase variations, and crosstalk.
  • CUDA-accelerated batched MZI array decomposition and reconstruction for ultra-fast real/complex matrix mapping, which achieves 10-50X speedup over CPU-based unitary group parametrization. Francis (Triangle), Reck (Triangle), Clements (Rectangle) styles MZI meshes are supported. To see the efficiency of our CUDA implementation, try the following unittest command at root directory, python3 unitest/test_op.py , and check the runtime comparison.
  • Support pytorch training general frequency-domain ONNs (Gu+, FFT-ONN ASP-DAC 2020) (Gu+, FFT-ONN-v2 IEEE TCAD 2021) (Feng+, PSNN). Support FFT-ONN BlockLinear, and BlockConv2d. Support fft, hadamard, zero_bias, and trainable modes.
  • Support multi-operand ring-based ONN architecture (Gu+, SqueezeLight DATE 2021). Support AllpassMORRCirculantLinear, AllpassMORRCirculantConv2d with built-in MORR nonlinearity.
  • Support phase change material (PCM)-based ONN architecture (Zhu+, ELight ASP-DAC 2022). Support PCMLinear and PCMConv2d with logrithmic PCM wire quantization and PCM array assignment.

TODOs

  • Support micro-ring resonator (MRR)-based ONN. (Tait+, SciRep 2017)
<!-- - [x] Support general frequency-domain ONN. (Gu+, [FFT-ONN](https://doi.org/10.1109/ASP-DAC47756.2020.9045156) ASP-DAC 2020) (Gu+, [FFT-ONN-v2](https://doi.org/10.1109/TCAD.2020.3027649) IEEE TCAD 2021) --> <!-- - [ ] Support multi-operand micro-ring (MORR)-based ONN. (Gu+, [SqueezeLight](https://jeremiemelo.github.io/publications/papers/ONN_DATE2021_SqueezeLight_Gu.pdf) DATE 2021) --> <!-- - [ ] Support differentiable quantization-aware training. (Gu+, [ROQ](https://doi.org/10.23919/DATE48585.2020.9116521) DATE 2020) -->
  • Support ONN on-chip learning via zeroth-order optimization. (Gu+, FLOPS DAC 2020) (Gu+, MixedTrain AAAI 2021)
<!-- - [ ] Support automatic differentiable ONN architecture search with SuperMesh training. (Gu+, [ADEPT](https://arxiv.org/abs/2112.08703) DAC 2022) -->

Files

FileDescription
torchonn/Library source files with model, layer, and device definition
torchonn/opBasic operators and CUDA-accelerated operators
torchonn/layersOptical device-implemented layers
torchonn/modelsBase ONN model templete
torchonn/devicesOptical device parameters and configurations
examples/ONN model building and training examples
examples/configsYAML-based configuration files
examples/coreONN model definition and training utility
example/train.pytraining script

More Examples

The examples/ folder contains more examples to train the ONN models.

An example optical convolutional neural network MZI_CLASS_CNN is defined in examples/core/models/mzi_cnn.py.

Training facilities, e.g., optimizer, critetion, lr_scheduler, models are built in examples/core/builder.py. The training and validation logic is defined in examples/train.py. All training hyperparameters are hierarchically defined in the yaml configuration file examples/configs/mnist/mzi_onn/train.yml (The final config is the union of all default.yml from higher-level directories and this specific train.yml ).

By running the following commands,

# train the example MZI-based CNN model with 2 64-channel Conv layers and 1 Linear layer # training will happend in usv mode to optimize U, Sigma, and V* # projected gradient descent will be applied to guarantee the orthogonality of U and V* # the final step will convert unitary matrices into MZI phases and evaluate in the phase mode cd examples python3 train.py configs/mnist/mzi_cnn/train.yml # [followed by any command-line arguments that override the values in config file, e.g., --optimizer.lr=0.001]

Detailed documentations coming soon.

Contact

Jiaqi Gu (jqgu@utexas.edu)

Related Projects using PyTorch-ONN Library

  • Neural operator-enabled fast photonic device simulation: See NeurOLight, NeurIPS 2022.

  • Automatic photonic tensor core design: See ADEPT, DAC 2022.

  • Endurance-enhanced photonic in-memory computing: See ELight, ASP-DAC 2022.

  • Scalable ONN on-chip learning: See L2ight, NeurIPS 2021.

  • Memory-efficient ONN architecture: See Memory-Efficient-ONN, ICCV 2021.

  • SqueezeLight: Scalable ONNs with Multi-Operand Ring Resonators: See SqueezeLight, DATE 2021.

编辑推荐精选

Trae

Trae

字节跳动发布的AI编程神器IDE

Trae是一种自适应的集成开发环境(IDE),通过自动化和多元协作改变开发流程。利用Trae,团队能够更快速、精确地编写和部署代码,从而提高编程效率和项目交付速度。Trae具备上下文感知和代码自动完成功能,是提升开发效率的理想工具。

热门AI工具生产力协作转型TraeAI IDE
问小白

问小白

全能AI智能助手,随时解答生活与工作的多样问题

问小白,由元石科技研发的AI智能助手,快速准确地解答各种生活和工作问题,包括但不限于搜索、规划和社交互动,帮助用户在日常生活中提高效率,轻松管理个人事务。

聊天机器人AI助手热门AI工具AI对话
Transly

Transly

实时语音翻译/同声传译工具

Transly是一个多场景的AI大语言模型驱动的同声传译、专业翻译助手,它拥有超精准的音频识别翻译能力,几乎零延迟的使用体验和支持多国语言可以让你带它走遍全球,无论你是留学生、商务人士、韩剧美剧爱好者,还是出国游玩、多国会议、跨国追星等等,都可以满足你所有需要同传的场景需求,线上线下通用,扫除语言障碍,让全世界的语言交流不再有国界。

讯飞智文

讯飞智文

一键生成PPT和Word,让学习生活更轻松

讯飞智文是一个利用 AI 技术的项目,能够帮助用户生成 PPT 以及各类文档。无论是商业领域的市场分析报告、年度目标制定,还是学生群体的职业生涯规划、实习避坑指南,亦或是活动策划、旅游攻略等内容,它都能提供支持,帮助用户精准表达,轻松呈现各种信息。

热门AI工具AI办公办公工具讯飞智文AI在线生成PPTAI撰写助手多语种文档生成AI自动配图
讯飞星火

讯飞星火

深度推理能力全新升级,全面对标OpenAI o1

科大讯飞的星火大模型,支持语言理解、知识问答和文本创作等多功能,适用于多种文件和业务场景,提升办公和日常生活的效率。讯飞星火是一个提供丰富智能服务的平台,涵盖科技资讯、图像创作、写作辅助、编程解答、科研文献解读等功能,能为不同需求的用户提供便捷高效的帮助,助力用户轻松获取信息、解决问题,满足多样化使用场景。

模型训练热门AI工具内容创作智能问答AI开发讯飞星火大模型多语种支持智慧生活
Spark-TTS

Spark-TTS

一种基于大语言模型的高效单流解耦语音令牌文本到语音合成模型

Spark-TTS 是一个基于 PyTorch 的开源文本到语音合成项目,由多个知名机构联合参与。该项目提供了高效的 LLM(大语言模型)驱动的语音合成方案,支持语音克隆和语音创建功能,可通过命令行界面(CLI)和 Web UI 两种方式使用。用户可以根据需求调整语音的性别、音高、速度等参数,生成高质量的语音。该项目适用于多种场景,如有声读物制作、智能语音助手开发等。

咔片PPT

咔片PPT

AI助力,做PPT更简单!

咔片是一款轻量化在线演示设计工具,借助 AI 技术,实现从内容生成到智能设计的一站式 PPT 制作服务。支持多种文档格式导入生成 PPT,提供海量模板、智能美化、素材替换等功能,适用于销售、教师、学生等各类人群,能高效制作出高品质 PPT,满足不同场景演示需求。

讯飞绘文

讯飞绘文

选题、配图、成文,一站式创作,让内容运营更高效

讯飞绘文,一个AI集成平台,支持写作、选题、配图、排版和发布。高效生成适用于各类媒体的定制内容,加速品牌传播,提升内容营销效果。

AI助手热门AI工具AI创作AI辅助写作讯飞绘文内容运营个性化文章多平台分发
材料星

材料星

专业的AI公文写作平台,公文写作神器

AI 材料星,专业的 AI 公文写作辅助平台,为体制内工作人员提供高效的公文写作解决方案。拥有海量公文文库、9 大核心 AI 功能,支持 30 + 文稿类型生成,助力快速完成领导讲话、工作总结、述职报告等材料,提升办公效率,是体制打工人的得力写作神器。

openai-agents-python

openai-agents-python

OpenAI Agents SDK,助力开发者便捷使用 OpenAI 相关功能。

openai-agents-python 是 OpenAI 推出的一款强大 Python SDK,它为开发者提供了与 OpenAI 模型交互的高效工具,支持工具调用、结果处理、追踪等功能,涵盖多种应用场景,如研究助手、财务研究等,能显著提升开发效率,让开发者更轻松地利用 OpenAI 的技术优势。

下拉加载更多