For a detailed explanation of this inference example, visit the documentation.
Make sure to set up a virtual environment for Python with all the required dependencies. More details on how to do this can be found here.
Make sure you have Maven installed and added to PATH. Also make sure that JAVA_HOME points to the correct Java version.
First we need to download the Maven archetype for Beam. Run the following command:
export BEAM_VERSION=<Beam version>
mvn archetype:generate \
-DarchetypeGroupId=org.apache.beam \
-DarchetypeArtifactId=beam-sdks-java-maven-archetypes-examples \
-DarchetypeVersion=$BEAM_VERSION \
-DgroupId=org.example \
-DartifactId=multi-language-beam \
-Dversion="0.1" \
-Dpackage=org.apache.beam.examples \
-DinteractiveMode=false
This will set up all the required dependencies for the Java pipeline. Next the pipeline needs to be
implemented. The logic of this pipeline is written in the MultiLangRunInference.java
file. After that,
run the following command to start the Java pipeline:
export GCP_PROJECT=<your gcp project>
export GCP_BUCKET=<your gcp bucker>
export GCP_REGION=<region of bucket>
export MODEL_NAME=bert-base-uncased
export LOCAL_PACKAGE=<path to tarball>
cd last_word_prediction
mvn compile exec:java -Dexec.mainClass=org.apache.beam.examples.MultiLangRunInference \
-Dexec.args="--runner=DataflowRunner \
--project=$GCP_PROJECT\
--region=$GCP_REGION \
--gcpTempLocation=gs://$GCP_BUCKET/temp/ \
--inputFile=gs://$GCP_BUCKET/input/imdb_reviews.csv \
--outputFile=gs://$GCP_BUCKET/output/ouput.txt \
--modelPath=gs://$GCP_BUCKET/input/bert-model/bert-base-uncased.pth \
--modelName=$MODEL_NAME \
--localPackage=$LOCAL_PACKAGE" \
-Pdataflow-runner
The localPackage
argument is the path to a locally available package compiled as a tarball. This package must be created by the user and contain the python transforms used in the pipeline.
Make sure to run this in the last_word_prediction
directory. This will start the Java pipeline.