Model Importing

Gridgain 8.8/Ignite 2.8 supports importing Machine Learning models from external platforms including Apache Spark ML and XGBoost. Working with imported models, you can:

store imported models in Ignite for further inference,
use imported models as part of pipelines,
apply ensembling methods such as boosting, bagging, or stacking to those models.

Also, imported pre-trained models can be updated inside GridGain.

GridGain provides an API for distributed inference for models trained in Apache Spark ML, XGBoost, and TensorFlow. These models cannot be modified via GridGain. GridGain provides a common API and you just need to use the right parser implementation to work with your specific external lib.

Model Import from Apache Spark

Model import from Apache Spark via Parquet files

GridGain 8.8/Ignite 2.8 supports the following models for import from Apache Spark ML:

Logistic regression (org.apache.spark.ml.classification.LogisticRegressionModel)
Linear regression (org.apache.spark.ml.classification.LogisticRegressionModel)
Decision tree (org.apache.spark.ml.classification.DecisionTreeClassificationModel)
Support Vector Machine (org.apache.spark.ml.classification.LinearSVCModel)
Random forest (org.apache.spark.ml.classification.RandomForestClassificationModel)
K-Means (org.apache.spark.ml.clustering.KMeansModel)
Decision tree regression (org.apache.spark.ml.regression.DecisionTreeRegressionModel)
Random forest regression (org.apache.spark.ml.regression.RandomForestRegressionModel)
Gradient boosted trees regression (org.apache.spark.ml.regression.GBTRegressionModel)
Gradient boosted trees (org.apache.spark.ml.classification.GBTClassificationModel)

This feature works with models saved in snappy.parquet files.

Here is an example of importing a model from Apache Spark ML via Apache Parquet:

DecisionTreeNode mdl = (DecisionTreeNode)SparkModelParser.parse(
   SPARK_MDL_PATH,
   SupportedSparkModels.DECISION_TREE
);

You can see more examples of using this API in the examples module in the package: org.apache.ignite.examples.ml.inference.spark.modelparser

Model import from Apache Spark via MLeap

In this mode you cannot update models or pipelines saved with MLeap, but you can import them and perform distributed inference.

For distributed inference, use the AsyncModelBuilder interface. Using this interface you can import a model or pipeline for local or distributed inference. For local inference (inference on the same node where you perform import) you can use ThreadedModelBuilder. For distributed inference, use the IgniteDistributedModelBuilder implementation of the AsyncModelBuilder interface.

This is an example of model import in a distributed manner:

File mdlRsrc = IgniteUtils.resolveIgnitePath(TEST_MODEL_RES);

ModelReader reader = new FileSystemModelReader(mdlRsrc.getPath());

MLeapModelParser parser = new MLeapModelParser();

AsyncModelBuilder mdlBuilder = new IgniteDistributedModelBuilder(ignite, 4, 4);

Model<NamedVector, Future<Double>> mdl = mdlBuilder.build(reader, parser);

In this example, we imported a model and deployed 4 instances in our cluster with the limit of maximum model instances per node.

Detailed examples are located in the example module in the package: org.apache.ignite.examples.ml.mleap

Model Import from GXBoost

You can import pre-trained models from XGBoost for local or distributed inference.

File mdlRsrc = IgniteUtils.resolveIgnitePath(TEST_MODEL_RES);

ModelReader reader = new FileSystemModelReader(mdlRsrc.getPath());

XGModelParser parser = new XGModelParser();

AsyncModelBuilder mdlBuilder = new IgniteDistributedModelBuilder(ignite, 4, 4);

Model<NamedVector, Future<Double>> mdl = mdlBuilder.build(reader, parser);

Model Import and Inference of Pre-Trained Models from TensorFlow

For more information about integration with TensorFlow, see this.

Just like with Apache Spark and XGBoost, you can import models from TensorFlow. In order to do that, you need to provide information about the model structure to a model parser and set inputs and outputs for your imported model.

File mdlRsrc = IgniteUtils.resolveIgnitePath(MODEL_PATH);

ModelReader reader = new FileSystemModelReader(mdlRsrc.getPath());

ModelParser<double[], Long, ?> parser = new TensorFlowSavedModelModelParser<double[], Long>("serve")

   .withInput("Placeholder", doubles -> {
       float[][][] reshaped = new float[1][28][28];
       for (int i = 0; i < doubles.length; i++)
           reshaped[0][i / 28][i % 28] = (float)doubles[i];
       return Tensor.create(reshaped);
   })

   .withOutput(Collections.singletonList("ArgMax"), collectedTensors -> collectedTensors.get("ArgMax")
       .copyTo(new long[1])[0]);

And after that, you will be able to import and deploy the TensorFlow model in your cluster:

Model<double[], Future<Long>> threadedMdl = new IgniteDistributedModelBuilder(ignite, 4, 4)

You can find more examples of importing models from TensorFlow in the related package in the example module: org.apache.ignite.examples.ml.inference

© 2024 GridGain Systems, Inc. All Rights Reserved. Privacy Policy | Legal Notices. GridGain® is a registered trademark of GridGain Systems, Inc.
Apache, Apache Ignite, the Apache feather and the Apache Ignite logo are either registered trademarks or trademarks of The Apache Software Foundation.

Last updated on Dec 04, 2024