/vega

AutoML tools chain

Primary LanguagePythonMIT LicenseMIT

Vega

English | 中文

Vega ver1.3.0 released:

  • Feature enhancement:
    • Ascend platform, search and training on the Ascend 910 (TensorFlow and MindSpore), and model evaluation on the Ascend 310.
    • Model evaluation is supported on the Kirin 990.
    • Allows user datasets to be FineTune on DNet pretrained models and surpass SOTA on Ascend 910/310.
    • Support the pruning capability of user datasets and ResNet models. For the Cifar100 data set, the precision changes slightly (+– 0.5), the latency decreases by 15%, and the model size decreases by 30%.
  • New algorithm:
    • ModularNAS: Towards Modularized and Reusable Neural Architecture Search, a code library for various neural architecture search methods including weight sharing and network morphism.
    • DNet: Network architecture search algorithms and Model Zoo that are affinity with Davinci chips.
    • MF-ASC: Multi-Fidelity neural Architecture Search with Co-kriging.

Introduction

Vega is an AutoML algorithm tool chain developed by Noah's Ark Laboratory, the main features are as follows:

  1. Full pipeline capailities: The AutoML capabilities cover key functions such as Hyperparameter Optimization, Data Augmentation, Network Architecture Search (NAS), Model Compression, and Fully Train. These functions are highly decoupled and can be configured as required, construct a complete pipeline.
  2. Industry-leading AutoML algorithms: Provides Noah's Ark Laboratory's self-developed industry-leading algorithm (Benchmark) and Model Zoo to download the state-of-the-art (SOTA) models.
  3. Fine-grained network search space: The network search space can be freely defined, and rich network architecture parameters are provided for use in the search space. The network architecture parameters and model training hyperparameters can be searched at the same time, and the search space can be applied to Pytorch, TensorFlow and MindSpore.
  4. High-concurrency neural network training capability: Provides high-performance trainers to accelerate model training and evaluation.
  5. Multi-Backend: PyTorch (GPU), TensorFlow (GPU and Ascend 910), MindSpore (Ascend 910).
  6. Ascend platform: Search and training on the Ascend 910 and model evaluation on the Ascend 310.

AutoML Tools Features

Supported Frameworks HPO Algorithms NAS Algorithms Device-Side Evaluation Model Filter Universal Network
AutoGluon mxnet, PyTorch Random Search, Bayesian, Hyper-Band Random Search, RL × × ×
AutoKeras Keras No Restrictions Network Morphism × × ×
Model Search TensorFlow No Restrictions Random Search, Beam Search × × ×
NNI No Restrictions Random Search and Grid Search, Bayesian, Annealing, Hyper-Band, Evolution, RL Random Search, Gradient-Based, One-Shot × × ×
Vega PyTorch, TensorFlow, MindSpore Random Search, Grid Search, Bayesian, Hyper-Band, Evolution Random Search, Gradient-Based, Evalution, One-Shot Ascend 310, Kirin 980/990 Quota (filter model based on parameters, flops, latency) provides networks compatibility with PyTorch, TensorFlow, and MindSpore

Algorithm List

Category Algorithm Description reference
NAS CARS: Continuous Evolution for Efficient Neural Architecture Search Structure Search Method of Multi-objective Efficient Neural Network Based on Continuous Evolution ref
NAS ModularNAS: Towards Modularized and Reusable Neural Architecture Search A code library for various neural architecture search methods including weight sharing and network morphism ref
NAS MF-ASC Multi-Fidelity neural Architecture Search with Co-kriging ref
NAS NAGO: Neural Architecture Generator Optimization An Hierarchical Graph-based Neural Architecture Search Space ref
NAS SR-EA An Automatic Network Architecture Search Method for Super Resolution ref
NAS ESR-EA: Efficient Residual Dense Block Search for Image Super-resolution Multi-objective image super-resolution based on network architecture search ref
NAS Adelaide-EA: SEGMENTATION-Adelaide-EA-NAS Network Architecture Search Algorithm for Image Segmentation ref
NAS SP-NAS: Serial-to-Parallel Backbone Search for Object Detection Serial-to-Parallel Backbone Search for Object Detection Efficient Search Algorithm for Object Detection and Semantic Segmentation in Trunk Network Architecture ref
NAS SM-NAS: Structural-to-Modular NAS Two-stage object detection architecture search algorithm Coming soon
NAS Auto-Lane: CurveLane-NAS An End-to-End Framework Search Algorithm for Lane Lines ref
NAS AutoFIS An automatic feature selection algorithm for recommender system scenes ref
NAS AutoGroup An automatically learn feature interaction for recommender system scenes ref
NAS MF-ASC Multi-Fidelity neural Architecture Search with Co-kriging ref
Model Compression Quant-EA: Quantization based on Evolutionary Algorithm Automatic mixed bit quantization algorithm, using evolutionary strategy to quantize each layer of the CNN network ref
Model Compression Prune-EA Automatic channel pruning algorithm using evolutionary strategies ref
HPO ASHA: Asynchronous Successive Halving Algorithm Dynamic continuous halving algorithm ref
HPO TPE: Tree-structured Parzen Estimator Approach A hyperparameter optimization Algorithm Based on Tree - Structured Parzen Estimation ref
HPO BO: Bayesian Optimization Bayesian optimization algorithm ref
HPO BOHB: Hyperband with Bayesian Optimization Hyperband with Bayesian Optimization ref
HPO BOSS: Bayesian Optimization via Sub-Sampling A universal hyperparameter optimization algorithm based on Bayesian optimization framework for resource-constraint hyper-parameters search ref
Data Augmentation PBA: Population Based Augmentation: Efficient Learning of Augmentation Policy Schedules Data augmentation based on PBT optimization ref
Data Augmentation CycleSR: Unsupervised Image Super-Resolution with an Indirect Supervised Path Unsupervised style migration algorithm for low-level vision problem. ref
Fully Train Beyond Dropout: Feature Map Distortion to Regularize Deep Neural Networks Neural network training (regularization) based on disturbance of feature map ref
Fully Train Circumventing Outliers of AutoAugment with Knowledge Distillation Joint knowledge distillation and data augmentation for high performance classication model training, achieved 85.8% Top-1 accuracy on ImageNet 1k Coming soon

Installation

Run the following commands to install Vega and related open-source software:

pip3 install --user --upgrade noah-vega

If you need to install the Ascend 910 training environment, please contact us.

Reference List

object refrence
User
(User Guide)
Install Guide, Deployment Guide, Configuration Guide, Examples, Evaluate Service
Developer
(Developer Guide)
Development Reference, Quick Start Guide, Dataset Guide, Algorithm Development Guide, Fine-Grained Search Space Guide

FAQ

For common problems and exception handling, please refer to FAQ.

Citation

@misc{wang2020vega,
      title={VEGA: Towards an End-to-End Configurable AutoML Pipeline},
      author={Bochao Wang and Hang Xu and Jiajin Zhang and Chen Chen and Xiaozhi Fang and Ning Kang and Lanqing Hong and Wei Zhang and Yong Li and Zhicheng Liu and Zhenguo Li and Wenzhi Liu and Tong Zhang},
      year={2020},
      eprint={2011.01507},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Cooperation and Contribution

Welcome to use Vega. If you have any questions or suggestions, need help, fix bugs, contribute new algorithms, or improve the documentation, submit an issue in the community. We will reply to and communicate with you in a timely manner.
Welcome to join our QQ chatroom (Chinese): 833345709.