无人驾驶 ai算法
Today, I continue my adventure in autoML tools. One of the leaders is H2O’s Driverless AI offering. It has some great features that impressed me. One drawback I had in this evaluation was that I didn’t have enough time to train the ‘Watson’ dataset properly. While the results are unavailable, I have been able to share enough of the experience for you to get a feel for the tool.
今天,我继续使用autoML工具进行冒险。 H2O的无人驾驶AI产品是领导者之一。 它的一些出色功能令我印象深刻。 我在此评估中的一个缺点是我没有足够的时间正确地训练“ Watson”数据集。 虽然无法获得结果,但我已经能够与您分享足够的经验,以使您对该工具有所了解。
为什么要使用无人驾驶AI? (Why Driverless AI?)
I’ve seen several demos of Driverless AI over the past couple of years. I’ve done an evaluation myself using a trial in my AWS account in 2018. I know from experience and reputation; this is one of the top autoML tools available. They tout their visualizations, and I was impressed with them two years ago.
在过去的几年中,我已经看过无人驾驶AI的多个演示。 我已经在2018年使用自己的AWS账户中的试用版进行了评估。 这是可用的顶级autoML工具之一。 他们吹捧他们的可视化效果,两年前,我对他们印象深刻。
设置和费用 (The setup and cost)
Like DataRobot, Driverless AI is a licensed product. Since last year, Driverless AI is available through IBM as well as other cloud platforms. I am unable to find the exact license cost for one named user for 2020. Based on my research, the price appears to be comparable to the DataRobot cost of $80k per year.
与DataRobot一样,无人驾驶AI也是许可产品 。 自去年以来,可通过IBM以及其他云平台使用无人驾驶AI。 我无法找到一位指定用户在2020年的确切许可成本。根据我的研究,该价格似乎与DataRobot每年8万美元的成本相当。
Lucky for us, we can do a free trial of Driverless AI. There are a couple of options. You can get 14 days on a cloud implementation or 2 hours of a hosted platform. In the past, I’ve used the AWS Marketplace offering. For this demo, I’m using the hosted platform. Two hours is a short period, so I need to be efficient.
幸运的是,我们可以免费试用无人驾驶AI。 有两种选择。 您可以在云实施上花费14天,在托管平台上花费2个小时。 过去,我曾使用过AWS Marketplace产品。 对于此演示,我正在使用托管平台。 两个小时很短,所以我需要提高效率。
I set up an account on Aquarium. There are several labs available to work through. As suggested, I started the test drive lab. I was not initially sure if I could load my dataset. I was happy to see I could.
我在水族馆开设了一个帐户。 有几个实验室可以解决。 按照建议,我开始了测试驱动器实验室。 最初我不确定是否可以加载我的数据集。 我很高兴看到自己能做到。
![Image for post](https://i-blog.csdnimg.cn/blog_migrate/b42ee7742912080ee0c3e27dbb360391.png)
Once the lab spun up, I got a URL that pointed to an AWS instance. Quick and easy.
实验室启动后,我得到了一个指向AWS实例的URL。 快捷方便。
数据 (The Data)
To keep parity across the tools in this series, I will stick to the Kaggle training file. Contradictory, My Dear Watson. Detecting contradiction and entailment in the multilingual text using TPUs. In this Getting Started Competition, we’re classifying pairs of sentences (consisting of a premise and a hypothesis) into three categories — entailment, contradiction, or neutral.
为了使本系列中的工具保持一致,我将坚持使用Kaggle培训文件。 矛盾的,亲爱的沃森。 使用TPU检测多语言文本中的矛盾和牵连 。 在本入门竞赛中,我们将成对的句子(由前提和假设组成)分为三类-蕴涵,矛盾或中立。
6 Columns x 13k+ rows — Stanford NLP documentation
6列x 13k +行— Stanford NLP 文档
- id ID
- premise 前提
- hypothesis 假设
- lang_abv lang_abv
- language 语言
- label 标签
加载数据 (Loading the data)
Loading the data is super easy, and some very nice visualizations are available before any training takes place. You can also add graphs to the list, so that’s a nice feature. Most autoML tools aren’t going to give you correlation graphs and radar plots. I thought the data heatmap was a bit weak, but that appears to have been due to the variety in the data (text). Interestingly when you download the visualizations, they arrive in .svg format. Nice if you have an Adobe license.
加载数据非常容易,在进行任何培训之前,可以使用一些非常漂亮的可视化文件。 您还可以将图形添加到列表中,因此这是一个不错的功能。 大多数autoML工具不会为您提供相关图和雷达图。 我以为数据热图有些薄弱,但这似乎是由于数据(文本)的多样性所致。 有趣的是,当您下载可视化文件时,它们以.svg格式到达。 如果您拥有Adobe许可证,那就很好。
![Image for post](https://i-blog.csdnimg.cn/blog_migrate/d200947666ecca0c18ed8cab6eb368e9.png)
![Image for post](https://i-blog.csdnimg.cn/blog_migrate/7f63f6e530efd1094333a63b9cdff014.png)
![Image for post](https://i-blog.csdnimg.cn/blog_migrate/06c8308cc5a45d7e9dc3ea5b5117205b.png)
训练模型 (Training your model)
It’s pretty simple to get the models training. Launch Experiment. You can tune the Accuracy, Time, and Interpretability dials to what your preferences are. There are also MANY expert settings you can review.
进行模型训练非常简单。 启动实验。 您可以将“精度”,“时间”和“可解释性”拨盘调整到您的首选项。 您还可以查看许多专家设置。
![Image for post](https://i-blog.csdnimg.cn/blog_migrate/4f8d1789c91a3874f8cb9f7d25df0d11.png)
Between this main visualization and the logs, you can get an excellent idea of the progress your training job is making. I appreciate this! I may have been too aggressive with my dials, and I was not able to fully make it through an entire experiment during my time allotment.
在此主要可视化和日志之间,您可以很好地了解培训工作的进度。 我很欣赏这个! 我的表盘可能太激进了,在分配时间的过程中,我无法完全通过整个实验来做到这一点。
![Image for post](https://i-blog.csdnimg.cn/blog_migrate/767a7c8293107765a4a9635b9e6f6905.png)
评估培训结果 (Evaluate Training Results)
Well, I didn’t have enough time to train the models with the time I had. I tried to start a new lab and try again, but I just got a ‘waiting for worker’ message for 45 minutes. I would have been interested in taking a look at the generated features. Feature generation is one of the noted differentiators of this tool.
好吧,我没有足够的时间来训练模型。 我尝试开始一个新实验室,然后再试一次,但是我收到了45分钟的“等待工人”消息。 我本来想对生成的功能感兴趣。 特征生成是该工具的显着差异之一。
The good news is that there are pre-run projects that you can poke through on your own.
好消息是,您可以自己浏览一些预运行项目。
![Image for post](https://miro.medium.com/max/9999/1*iQB7clnCn0Ky02vAlp9a5w.png)
I did see H2O added an AutoReport feature! Very nice addition. The report isn’t quite as extensive as DataRobot’s, but it’s good work in the right direction.
我确实看到H2O添加了AutoReport功能! 非常好。 该报告并不像DataRobot的报告那样广泛,但是在正确的方向上是一项很好的工作。
![Image for post](https://i-blog.csdnimg.cn/blog_migrate/532a2f6353514e677cb9ebdf54f6ef31.png)
H2O has a free booklet on model interpretability that I highly recommend. They are a leader in this area.
我强烈建议H2O提供有关模型可解释性的免费手册。 他们是这一领域的领导者。
When you click Interpret this Model, you have to wait while the process runs. Don’t stare at the screen or you’ll find your eyes circling with the status’.
单击“解释此模型”时,必须等待过程运行。 不要凝视屏幕,否则您会发现眼睛在盘旋状态。
![Image for post](https://i-blog.csdnimg.cn/blog_migrate/82888846ce2fa1fbe799dcd964f932b1.png)
Slowly the explanations become available. The result is worth the wait.
慢慢地可以得到解释。 结果值得期待。
![Image for post](https://i-blog.csdnimg.cn/blog_migrate/3fdcf3d77df5d8a9fa2fe2ef8a85cc5e.png)
![Image for post](https://i-blog.csdnimg.cn/blog_migrate/9ecb6630b804c0b724a9339acbda4915.png)
结论 (Conclusions)
Driverless AI is a great tool. They provide interesting visualizations and allow you to add additional ones. The model interpretability metrics and graphs are terrific. As with DataRobot, you pay for greatness.
无人驾驶AI是一个很棒的工具。 它们提供了有趣的可视化效果,并允许您添加其他可视化效果。 模型的可解释性指标和图表非常棒。 与DataRobot一样,您需要付出巨大的代价。
The labs are free, so I encourage you to try them today. The pre-trained projects give you a good jump start so that you can take a look for yourself.
实验室是免费的,所以我建议您立即尝试。 经过预培训的项目为您提供了一个良好的入门指南,以便您可以自己看看。
翻译自: https://towardsdatascience.com/h2o-driverless-ai-71414b441425
无人驾驶 ai算法