Yonsei CVLAB website ^HOT      CVLAB Open Sources ^{TRY IT OUT}

 

📘Documentation | 🛠️Installation | 🚀Research | 🤔Reporting Issues

---

Introduction

YSAutoML is an open-source library for the automated construction and optimization of AI systems, developed by Yonsei CVLab.
It integrates key technologies across data, network architectures, and optimization to build task-specific, efficient deep learning solutions with minimal manual intervention.

The library is designed as part of a multi-year project that aims to:

• Develop a next-generation automated AI platform applicable to image recognition, segmentation, and object detection.
• Provide a unified solution that integrates dataset handling, model search, and efficient training.
• Support users in rapidly generating customized AI models according to their requirements (task, dataset, resource budget, etc.).

---

✨ Major Features

• Integrated Automation
  Combines dataset utilities, network architecture search (NAS), and optimization into a single pipeline.

• Data Utilities
  Tools for dataset condensation, augmentation-aware BN (DSBN), and memory-efficient training.

• Network Utilities (NAS)
  Support for few-shot, zero-shot, and one-shot NAS frameworks to automatically discover architectures.

• Optimization Utilities
  Fixed-point quantization, loss function search, and multi-task loss integration for efficient deployment.

• User-Centered Interface
  Accepts user requirements such as task type, dataset, computational budget, and memory constraints, and generates optimized models automatically.

---

What's New

💎 Initial release of YSAutoML: a unified library integrating dataset, network, and optimization utilities.

Highlight

• Automated condensation of large datasets into compact synthetic sets.
• DSBN (Dual BatchNorm) support for domain-adaptive training.
• Unified NAS framework supporting multiple search paradigms.
• Built-in support for quantization and multi-task loss training.

---

Installation

YSAutoML requires Python 3.8+ and PyTorch 1.8+.

You can install ysautoml using two primary methods: via the Python Package Index (PyPI) for stable releases or directly from the source code via Git for the latest developments.

---

1. Installation via PyPI (Recommended)

Since ysautoml is a deep learning library with specific CUDA dependencies, you must use a custom index URL provided by PyTorch to ensure compatibility.

Prerequisites

To successfully install and run the package, you must first build and run a Docker container based on the provided Dockerfile in the root directory. This ensures that all necessary system dependencies and the correct CUDA environment (cu111) are available.

Installation Command

Once inside the prepared Docker container, execute the following command to install the package, noting the use of multiple index URLs to find both ysautoml (on TestPyPI) and PyTorch's CUDA-specific binaries:

pip install \
    --index-url [https://test.pypi.org/simple/](https://test.pypi.org/simple/) \
    --extra-index-url [https://pypi.org/simple/](https://pypi.org/simple/) \
    --extra-index-url [https://download.pytorch.org/whl/cu111](https://download.pytorch.org/whl/cu111) \
    ysautoml==0.1.1

2. Installation via Git Clone

If you prefer to work with the latest, unreleased version of the source code, you can clone the repository directly and install it in "editable" mode.

# Clone the repository
git clone [https://github.com/YourOrganization/ysautoml.git](https://github.com/YourOrganization/ysautoml.git)
cd ysautoml

# Install the package in editable mode within your environment
pip install -e .

---

Getting Started

Dataset Condensation with DSA

from ysautoml.data.fyi import run_dsa

run_dsa(
    dataset="CIFAR10",
    model="ConvNet",
    ipc=10,
    dsa_strategy="color_crop_cutout_flip_scale_rotate",
    init="real", lr_img=1.0, num_exp=5, num_eval=5,
    run_name="DSAFYI", run_tags="CIFAR10_10IPC", device="0", eval_mode="M",
)

Domain-Specific BatchNorm (DSBN)

from ysautoml.data.dsbn import convert_and_wrap, train_with_dsbn

model = convert_and_wrap("resnet18_cifar", dataset="CIFAR10", num_classes=10)
result = train_with_dsbn(model, source_loader, target_loader, epochs=5, lr=0.01)
print(result["final_acc"])

---

Module Overview

ysautoml
│
├── data
│   ├── fyi      # Dataset condensation (Flip Your Images)
│   ├── dsbn     # Domain-Specific BatchNorm
│
├── network
│   ├── fewshot  # Few-shot NAS
│   ├── zeroshot # Zero-shot NAS
│   ├── oneshot  # One-shot NAS
│
└── optimization
    ├── fxp        # Fixed-point quantization
    ├── losssearch # Loss search
    └── mtlloss    # Multi-task learning loss

---

Roadmap

This project is part of a 4-year research initiative.
In the final phase, we focus on integrating and optimizing exploration spaces for data, network, and loss jointly.

• Unified Search Space Optimization: jointly optimize network structures, data augmentation, and objectives.
• Automated Library Development: provide a unified library that can generate AI models from user inputs.
• Efficiency Enhancements: integrate dataset compression and fast optimization techniques to reduce memory and training cost.

---

FAQ

Please refer to the documentation for common usage questions.

---

Acknowledgement

YSAutoML is developed by researchers at Yonsei CVLab.
We thank all contributors and collaborators who helped design, implement, and test its components.

---

Citation

If you use this toolbox or benchmark in your research, please cite:

@misc{ysautoml2025,
  title   = {YSAutoML: Automated AI System Construction Library},
  author  = {Yonsei CVLab},
  year    = {2025},
  note    = {https://github.com/cvlab-yonsei/YonseiAutoML}
}