The following steps show how to download a model training in H2O Driverless AI and push it to production in cnvrg.io:

1.                Create a Docker container based on Ubuntu installed with Python 3.8.
2.                Launch the Docker container using the docker run command.
3.                In your workstation, install the H2O Driverless AI library using the following command:

$ pip install driverlessai

4.                Download the Python Scoring Pipeline using the following Python code:

import driverlessai

dai = driverlessai.Client(address='<IP address of Driverless AI instance', username='<username>', password='<password>')

ex = dai.experiments.list()[<Experiment number>]

artifacts = ex.artifacts.download(overwrite=True)

5.                Extract the scoring pipeline ZIP file.

The ZIP file includes all the artifacts required for model inference. The scoring pipeline includes the run_example.sh and common-functions.sh files, which install all the required libraries in a Python virtual environment that is required for model serving.

6.                Because cnvrg model serving does not support containers with a virtual environment, we need to modify these files using the following command:

$ patch run_example.sh patchfile1

The following are the contents of the patchfile1 file:

21,22c21,22

<             virtualenv -p python3.8 --system-site-packages --never-download --copies --app-data env_app_data_dir --setuptools embed --pip embed --wheel embed env

<             source env/bin/activate

---

>             #virtualenv -p python3.8 --system-site-packages --never-download --copies --app-data env_app_data_dir --setuptools embed --pip embed --wheel embed env

>             #source env/bin/activate

30,31c30,31

<             virtualenv -p python3.8 --never-download --copies --app-data env_app_data_dir --setuptools embed --pip embed --wheel embed env || virtualenv -p python3.8 --never-download env

<             source env/bin/activate

---

>             #virtualenv -p python3.8 --never-download --copies --app-data env_app_data_dir --setuptools embed --pip embed --wheel embed env || virtualenv -p python3.8 --never-download env

>             #source env/bin/activate

45,47c45,47

<             mv $spackagespath/xgboost $spackagespath/xgboost_h2o4gpu

<             mv $spackagespath/lightgbm_gpu $spackagespath/lightgbm_gpu_h2o4gpu

<             mv $spackagespath/lightgbm_cpu $spackagespath/lightgbm_cpu_h2o4gpu

---

>             #mv $spackagespath/xgboost $spackagespath/xgboost_h2o4gpu

>             #mv $spackagespath/lightgbm_gpu $spackagespath/lightgbm_gpu_h2o4gpu

>             #mv $spackagespath/lightgbm_cpu $spackagespath/lightgbm_cpu_h2o4gpu

61c61,62

<        source env/bin/activate

---

>        #source env/bin/activate

>        echo "just pass"

114c115

<             set_PYTHON

---

>             #set_PYTHON

169c170

< source "./common-functions.sh"

---

> source "./modified-common-functions.sh"

172a174

>

7.                Update the common-functions.sh file using the following command:

$ patch common-functions.sh patchfile2

The following are the contents of the patchfile2 file:

<     PYTHON=`realpath env/bin/python`

---

>     #PYTHON=`realpath env/bin/python`

>     PYTHON=`realpath /usr/bin/python3`

76c77,78

<     spackagespath=`$PYTHON -c "from sysconfig import get_paths ; info = get_paths() ; print(info['purelib'])"`

---

>     #spackagespath=`$PYTHON -c "from sysconfig import get_paths ; info = get_paths() ; print(info['purelib'])"`

>     spackagespath="/usr/local/lib/python3.8/dist-packages"

133c135,136

<     export PYTHON=`realpath env/bin/python`

---

>     #export PYTHON=`realpath env/bin/python`

>     export PYTHON=`realpath /usr/bin/python3`

135c138,139

<     export spackagespath=`$PYTHON -c "from sysconfig import get_paths ; info = get_paths() ; print(info['purelib'])"`

---

>     #export spackagespath=`$PYTHON -c "from sysconfig import get_paths ; info = get_paths() ; print(info['purelib'])"`

>     export spackagespath="/usr/local/lib/python3.8/dist-packages"

140a145

>

8.                Run the following command to install the required libraries in the container:

$ apt-get update && apt install -y build-essential  libmagic-dev libopenblas-dev  git  locales  unzip  wget

$ /usr/bin/python3.8 -m pip install --upgrade pip

$ cd /driverless/scoring-pipeline

$ ./ run_example.sh

9.                Commit the Docker container and push it to a private repository.

This Docker container can be used in cnvrg.io for model serving.

10.           Implement the model prediction as a Python wrapper function.

To push a model to production, cnvrg requires a predict function to implement the model prediction. H2O Driverless AI implements model prediction using a scorer function. Use the following example Python predict wrapper function:

import os

import pandas as pd

import numpy as np

from numpy import nan

from scipy.special._ufuncs import expit

# import scoring_pipeline

from <scoring experiment module> import Scorer

os.environ['DRIVERLESS_AI_LICENSE_KEY']="<License key>"

def predict(*input_text):

    """

    Function to read input texts and predict the sentiment associated

    """

    # create scorer instance

    scorer = Scorer()

    model_input = input_text

    # predict using score using the instance

    prediction = scorer.score(model_input)

    scorer.finish()

return prediction

11.           Follow the instructions in cvnrg.io to publish a model for serving using the Docker container and the preceding Python code.