NLP is one of the key use cases supported by H2O Driverless AI. The H2O Driverless AI platform can support both stand-alone text and text with other column types as predictive features. TensorFlow-based and PyTorch Transformer Architectures (for example, BERT) are used for feature engineering and model building. H2O.ai has also validated this use case and made results available in their documentation.

We used a classic example of sentiment analysis of tweets using the US Airline Sentiment dataset. The sentiment of each tweet was labeled in advance and our model was used to label new tweets. 

The following table shows some samples from the dataset:

Table 2.      Example rows in dataset

The following table describes the hardware and software of the two configurations used for validating this design. The software versions were the latest during the time of validation. Newer versions might be available after the publication of this white paper and are fully supported:

Table 3.      Validation setup

[1] The validated use case does not fully utilize the cluster. The listed servers are the minimum number of servers recommended in a cluster.

Data ingestion—The US Airline Sentiment dataset is available in Kaggle. We uploaded the dataset to the H2O Driverless AI local repository.

Data preparation—Using H2O Driverless AI, we split the dataset into training, validation, and test sets. We used the tweets in the ‘text’ column and the sentiment (positive, negative, or neutral) in the ‘airline_sentiment’ column for this demonstration. Because we did not want to use other columns in the dataset, we clicked Dropped Cols, and then excluded everything. We also used the visualization feature for a better understanding of the dataset, as shown in the following figure:

Figure 6.   Data visualization options in H2O Driverless AI 

Model Building—We configured the experiment using the settings shown in the following figure and then ran the experiment. Text features were automatically generated and evaluated during the feature engineering process. Some features such as TextCNN rely on TensorFlow models can be accelerated through GPUs. We verified the use of GPUs using the nvidia-smi command.

Figure 7.   Completed experiment for sentiment analysis

Productization—When the experiment is completed, you can make new predictions and download the scoring pipeline as with any other H2O Driverless AI experiments. The model is then pushed to production and made available for prediction through a REST end point. We downloaded the MOJO Scoring Pipeline and built a container for it. The instructions for building the container are available in H2O’s GitHub. 

We used the following Dockerfile to build the container: 

FROM ubuntu:18.04

# Install necessary dependencies ls -al

RUN apt-get update && \

apt-get -y install --no-install-recommends \

build-essential \

ca-certificates \

openjdk-11-jdk \

&& rm -rf /var/lib/apt/lists/*

# Install H2O.ai Inference server

RUN mkdir -p /opt/H2O

WORKDIR /opt/H2O

COPY /opt/H2O/DAI-Mojo2-RestServer-1.0.jar ./

COPY H2OaiRestServer.properties ./

COPY license.sig ./

COPY pipeline.mojo ./

COPY goRestServerStart.sh ./

Run /bin/sh -c 'chmod a+x goRestServerStart.sh'

ENTRYPOINT ["/bin/sh","./goRestServerStart.sh"]

We then built a Kubernetes service to deploy the container. Using the external IP address configured for the service, we can access the REST end point. We tested the end point with the following curl command and received the following output:

$ curl -X POST -d '{"fields": ["text"], "rows": [["@Airlines You are horrible. Really bad food"]]}' http://<external-ip>/model/score -H "Content-Type: application/json”

{"id":"92fcc278-e8e5-11ec-9e4d-06f524030811","fields":["airline_sentiment.negative","airline_sentiment.neutral","airline_sentiment.positive"],"score":[["0.14810233","0.13797826","0.7139194"]]}