Hyperbox | Open-Source AutoML Toolbox
A clean and scalable AutoML framework built on PyTorch + PyTorch Lightning. Supports NAS, HPO, and Auto Data Augmentation with a unified, modular API.
English
GitHub: marsggbo/hyperbox · Docs: hyperbox-doc.readthedocs.io
Hyperbox is an open-source AutoML framework I developed during my PhD, built on top of PyTorch and PyTorch Lightning. It provides a clean, modular API covering Neural Architecture Search (NAS), Hyperparameter Optimization (HPO), and Auto Data Augmentation (ADA).
Core Modules
hyperbox.mutables — Searchable building blocks
Three types of searchable modules cover the vast majority of NAS search spaces:
| Module | Purpose |
|---|---|
OperationSpace | Select one operation from a candidate list (e.g. conv 3×3 vs 5×5 vs 7×7) |
InputSpace | Select which previous layer outputs to use as skip connections |
ValueSpace | Search fine-grained parameters (kernel size, channel width, etc.) inside a single op |
from hyperbox.mutables.spaces import OperationSpace, InputSpace, ValueSpace
from hyperbox.mutables.ops import Conv2d
# Choose one conv kernel size
ks = ValueSpace([3, 5, 7], key='kernelSize')
cout = ValueSpace([16, 32, 64], key='channelOut')
conv = Conv2d(3, cout, ks, stride=1, padding=2, bias=False)
hyperbox.mutator — Search algorithms
Plug-and-play search algorithms that work with any mutables-based search space:
| Algorithm | File |
|---|---|
| Random Search | random_mutator.py |
| DARTS (gradient-based) | darts_mutator.py |
| ENAS (RL-based) | enas_mutator.py |
| Evolution | evolution_mutator.py |
| FairNAS | fairnas_mutator.py |
| Few-shot NAS | fewshot_mutator.py |
| ProxylessNAS | proxyless_mutator.py |
| RepNAS | repnas_mutator.py |
from hyperbox.mutator import RandomMutator
from hyperbox.networks.ofa import OFAMobileNetV3
net = OFAMobileNetV3() # supernet (mask=None → searchable)
mutator = RandomMutator(net)
mutator.reset() # sample one subnet
arch = mutator._cache # arch as dict
subnet = net.build_subnet(arch) # inherit supernet weights
hyperbox.networks — Pre-built search spaces
Implementations of 15+ published NAS search spaces, including: DARTS, ENAS, OFA (MobileNetV3), ProxylessNAS, SPOS, NAS-Bench-201, NAS-Bench-301, ViT, GPT, ResNet, RepNAS, …
Applications
Hyperbox is the experimental backbone of several published papers:
| Paper | Venue | Description |
|---|---|---|
| NAS-LID | AAAI 2023 | NAS via local intrinsic dimension |
| EMARS | MICCAI 2022 | Multi-objective evolutionary NAS for medical imaging |
| MedPipe | IEEE MedAI 2023 | Joint search of augmentation policy and architecture |
Install
pip install hyperbox
# or from source
git clone https://github.com/marsggbo/hyperbox && cd hyperbox
pip install -r requirements.txt && python setup.py develop
中文版本
GitHub： marsggbo/hyperbox · 文档： hyperbox-doc.readthedocs.io
Hyperbox 是我在读博期间开发的开源 AutoML 框架，基于 PyTorch 和 PyTorch Lightning 构建，提供一套模块化、可扩展的 API，覆盖神经架构搜索（NAS）、超参优化（HPO）和自动数据增强（ADA）三大方向。
核心模块
hyperbox.mutables — 可搜索模块
三种可搜索模块基本覆盖了现有 NAS 搜索空间的所有主流设计：
| 模块 | 用途 |
|---|---|
OperationSpace | 从候选操作列表中选一个（如 3×3 / 5×5 / 7×7 卷积） |
InputSpace | 从前面若干层输出中选一个或多个作为当前层输入（跳跃连接） |
ValueSpace | 在单个操作内部搜索细粒度参数（卷积核大小、通道数等） |
每个模块都有 key 和 mask 两个关键参数：key 是唯一标识符，mask 指定当前选择（None 表示待搜索）。
hyperbox.mutator — 搜索算法
Mutator 与搜索空间完全解耦，可以自由替换搜索算法而无需改动模型定义。已内置 Random、DARTS、ENAS、进化算法、FairNAS、Few-shot NAS、ProxylessNAS、RepNAS 等多种经典算法。
hyperbox.networks — 内置搜索空间
重新实现了 15+ 篇经典 NAS 论文的搜索空间，包括 DARTS、ENAS、OFA（MobileNetV3）、ProxylessNAS、SPOS、NAS-Bench-201/301、ViT、GPT 等。
应用论文
| 论文 | 会议 | 说明 |
|---|---|---|
| NAS-LID | AAAI 2023 | 基于局部本征维度的神经架构搜索 |
| EMARS | MICCAI 2022 | 多目标进化 NAS 用于医疗图像分类 |
| MedPipe | IEEE MedAI 2023 | 联合搜索数据增强策略与网络结构 |