Ludwig v0.2 release, Uber open-source AI Kit

Ludwig 0.2 has been released, Ludwig based on open source toolkit on Google TensorFlow framework, users no longer need to write any code to develop the depth of learning. Toolbox version 0.2 of its depth study adds new features and other improvements, including a number of new tools, some improvements under the hood and more than 50 bug fixes. Including integration with Comet.ml, add Google's Bert natural language model for audio / voice, H3 (geospatial) and date (tense) features visualization API major improvements, and include services and so on.

Details are as follows:

Comet.ml

Comet.ml  This is a convenience AI code and experiment management utilities, especially DougBlank, providing a new Comet.ml integration. It can automatically monitor model from a unified dashboard. Through customizable panel, users can compare the experimental design, capture the model configuration changes and record test results and details, view real-time performance charts in the model training, and to establish better analyze ultra-parameter model.

Users simply add -comet parameters and configure their comet.config file.

BERT encoder

BERT is a text-based coding model data converter architecture, which is based on a self-supervised manner large amounts of text data for training. Bert added a encoder to Ludwig text encoder may list in order to allow the use of data in supervised small but very high performance requirements of the situation. This means that henceforth, anyone can get the most advanced text classifier, without writing a single line of code.

Audio / speech characteristics

Audio features added not only opens the door to more applications, such as automatic speech recognition and speaker recognition, but also provides additional signal is a multi-modal task. They work in the pretreatment aspect similar to the image feature, because in both cases need to specify a file path, so that they can load data from Ludwig, and the coding sequence and characteristics similar to the time series.

在默认情况下,它们被映射为一维原始信号,但短时间傅里叶变换和群延迟特征提取也可以通过改变预处理参数来实现,并获得许多语音模型中使用的经典特征。下面是如何在 YAML 模型定义中添加音频/语音功能的示例:

H3 特性

通过 H3 集成,用户可以向 Ludwig 模型提供六角形数据,而无需任何额外的操作。此外,还提供了三个新的编码器,通过编码 H3 整数的分量(模式、基本六边形和低分辨率单元),将其编码成一个潜在的表示形式。

第一个编码器,即 Embedded,嵌入每个组件,并通过对嵌入的求和来聚合它们。第二种是 weighted_sum,在第一个基础上加了学习权重来组合嵌入。第三种是 RNN,嵌入组件,使用递归神经网络将它们组合起来,遵循层次结构中从更细粒度的六边形到最细粒度的六角的顺序依赖关系。

下面是如何将H3特性添加到 YAML 模型定义文件中:

Date 特性

0.2 版本添加了一种简单的方法来支持 Ludwig 中的日期和时间戳。此功能允许用户将发生在某一天或某一特定时间的事件输入到 Ludwig,以获得关于它们的预测。另外还为日期提供了两个编码器:第一个,embed,独立嵌入每个组件,然后连接得到的表示;第二个,wave,用一个周期函数对每个分量进行编码,并将结果连在一起。

下面是如何向 YAML 模型定义文件中添加日期特性的示例:

服务器

0.2 版允许用户直接将经过训练的模型服务到核心的 Ludwig 库中,官方采用了 FastAPI 库来生成一个 REST 服务器,该服务器可以被查询以获得预测。

为了直接向 Ludwig 提供经过培训的模型,用户只需运行:

然后,他们将能够查询模型以获得如下预测:

可视化 API

解决可视化问题,提供一个选项来指定一个输出路径,在其中保存图幅,而不是在一个新窗口中显示它们。这种功能可以这样使用:

解决可视化问题的第二种方法是在 API 中提供函数,这些函数可以在笔记本中以编程方式调用。以前通过命令行可用的每个函数现在也可以通过 API 获得。以下是一个例子:

为了添加此功能,官方重构了所有的可视化代码,使其更加清晰、封装更好、更易于测试。

spaCy 预处理

在这个新版本中,还引入了向平台核心库提供经过培训的人工智能模型的能力,它还添加了意大利语、西班牙语、德语、法语、葡萄牙语、荷兰语、希腊语和多语言标记(开放源码 spacy nlp 库的最新版本)。

以上是主要的更新内容,详情见发布说明

Guess you like

Origin www.oschina.net/news/108596/ludwig-v-0-2-released