Epoxy: Interactive Model Iteration with Weak Supervision

栏目: IT技术 · 发布时间: 3年前

内容简介:Epoxy uses weak supervision and pre-trained embeddings to create models that can train at programmatically-interactive speeds (less than 1/2 second), but that can retain the performance of training deep networks. This repository presents a simple proof-of-

Epoxy: Interactive Model Iteration with Weak Supervision and Pre-Trained Embeddings

Epoxy uses weak supervision and pre-trained embeddings to create models that can train at programmatically-interactive speeds (less than 1/2 second), but that can retain the performance of training deep networks. This repository presents a simple proof-of-concept implementation for Epoxy (our implementation is around 100 LOC, including docstrings).

Epoxy: Interactive Model Iteration with Weak Supervision

In weak supervision, users write noisy labeling functions that generate labels for the data. Historically, we have observed that these labeling functions are often high accuracy but low coverage (each labeling function only votes on a subset of points). The only ways to make up the gap in the past have been to write more labeling functions (which can get difficult as you start dealing with the long tail), or use the labeling functions to train an end model (see, e.g., FlyingSquid for more details).

In Epoxy, we use pre-trained embeddings to get some of the benefits of training an end model--without having to train one. We use the embeddings to create extended labeling functions through nearest-neighbors search (improving coverage), and then use FlyingSquid to aggregate the extended labeling functions. This helps get some of the benefits of training a deep network, but at a fraction of the cost. And if you do have time to train a deep network, Epoxy can be used to generate labels to train a downstream end model as well.

Check out our paper on arXiv for more details!

Getting Started

  • Check out the example tutorial for a simple Jupyter notebook showing the proof of concept in this repo.

Installation

This repository depends on FlyingSquid. We recommend using conda to install FlyingSquid, and then you can install Epoxy:

git clone https://github.com/HazyResearch/flyingsquid.git

cd flyingsquid

conda env create -f environment.yml
conda activate flyingsquid

pip install -e .

cd ..

git clone https://github.com/HazyResearch/epoxy.git

cd epoxy

pip install -e .

Alternatively, you can install FlyingSquid (and its dependencies) yourself, see the FlyingSquid repo for more details.

Citation

If you use our work or found it useful, please cite our arXiv paper for now:

@article{chen2020train,
  author = {Mayee F. Chen and Daniel Y. Fu and Frederic Sala and Sen Wu and Ravi Teja Mullapudi and Fait Poms and Kayvon Fatahalian and Christopher R\'e},
  title = {Train and You'll Miss It: Interactive Model Iteration with Weak Supervision and Pre-Trained Embeddings},
  journal = {arXiv preprint arXiv:2006.15168},
  year = {2020},
}

以上所述就是小编给大家介绍的《Epoxy: Interactive Model Iteration with Weak Supervision》,希望对大家有所帮助,如果大家有任何疑问请给我留言,小编会及时回复大家的。在此也非常感谢大家对 码农网 的支持!

查看所有标签

猜你喜欢:

本站部分资源来源于网络,本站转载出于传递更多信息之目的,版权归原作者或者来源机构所有,如转载稿涉及版权问题,请联系我们

人工智能+:AI与IA如何重塑未来

人工智能+:AI与IA如何重塑未来

[美]韩德尔·琼斯(Handel Jones) [中]张臣雄 / 机械工业出版社 / 2018-10 / 55.00

当深度学习模型引发了全世界对人工智能的再次关注时,人工智能迎来第三次高速增长,人工智能(AI)、增强现实(AR)和虚拟现实(VR)正把人类带向新的“智能增强时代”(IA),我们将在不知不觉中接纳机器智能。 针对人类社会长期存在的众多复杂的动态的难题,人机融合智能将会提供全新的解决方案,谷歌、Facebook、微软、亚马逊、腾讯、阿里巴巴、百度等平台巨头纷纷斥千亿巨资布局人工智能的尖端技术;智......一起来看看 《人工智能+:AI与IA如何重塑未来》 这本书的介绍吧!

HTML 压缩/解压工具
HTML 压缩/解压工具

在线压缩/解压 HTML 代码

Markdown 在线编辑器
Markdown 在线编辑器

Markdown 在线编辑器

RGB HSV 转换
RGB HSV 转换

RGB HSV 互转工具