menonpg/h2o_mlengine

Running batch predictions on Google ml-Engine

Making h2o.ai batch predictions on Google ML Engine

This little package allows to make batch predictions for h2o.ai models, the output is the original csv file with new columns for the prediction and each class probabilities.

The main goal of the package is making predictions for big files and heavy models, using google machine learning engine in a serverless way.

Requirements

After training your h2o model, export the MOJO (.zip file) and h2o_genmodel.jar and save them in Google Cloud Storage.

Usage

For running on Google ML Engine, launch the following command

gcloud ml-engine jobs submit training JOB_NAME \
        --module-name=h2o_mlengine.main \
        --package-path h2o_mlengine/ \
        --region=us-east1 \
        --staging-bucket=gs://project-bucket \
        --scale-tier=BASIC \
        --runtime-version=1.9 \
        -- \
        --mojo_path gs://path/to/mojo.zip \
        --genmodel_path gs://path/to/h2o-genmodel.jar \
        --input_file gs://path/to/input.csv \
        --output_dir gs://bucket/ \
        --output_name output_name.csv

For running on your local computer, go to the h2o_mlengine folder and lauch:

python main.py --mojo_path gs://path/to/mojo.zip --genmodel_path gs://path/to/h2o-genmodel.jar --input_file gs://path/to/input.csv --output_dir gs://bucket/ --output_name output_name.csv

You can track what is happening using the command:

gcloud ml-engine jobs describe JOB_NAME