blublinsky commented on a change in pull request #8402: [FLINK-12473][ml] Add the interface of ML pipeline and ML lib URL: https://github.com/apache/flink/pull/8402#discussion_r285695555
########## File path: flink-ml/flink-ml-api/src/main/java/org/apache/flink/ml/api/core/PipelineStage.java ########## @@ -0,0 +1,70 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.flink.ml.api.core; + +import org.apache.flink.ml.api.misc.param.ParamInfo; +import org.apache.flink.ml.api.misc.param.WithParams; +import org.apache.flink.ml.api.misc.persist.Persistable; +import org.apache.flink.ml.util.param.ExtractParamInfosUtil; +import org.apache.flink.ml.util.persist.MLStageFactory; + +import com.google.gson.Gson; +import com.google.gson.JsonObject; + +import java.io.Serializable; +import java.util.HashMap; +import java.util.List; +import java.util.Map; + +/** + * Base class for a stage in a pipeline. The interface is only a concept, and does not have any + * actual functionality. Its subclasses must be either Estimator or Transformer. No other classes + * should inherit this interface directly. + * + * <p>Each pipeline stage is with parameters and meanwhile persistable. + * + * @param <T> The class type of the PipelineStage implementation itself, used by {@link + * org.apache.flink.ml.api.misc.param.WithParams} + * @see WithParams + */ +interface PipelineStage<T extends PipelineStage<T>> extends WithParams<T>, Serializable, + Persistable { + + default String toJson() { Review comment: There are a few questions here: 1. I think ML and serving will require 2 separate instances of Flink. ML is mostly batch, while Serving can be either batch or streaming. I think such separation should be very explicit in the overall design. Although theoretically it is possible to use ML to update the model directly, such a use case is not going to be a dominant one. 2. If we agree on this, there are 2 reasons to export the model - here model is really a pipeline. It is used for continue learning and serving. Continue learning, in this case means that I have new data and I want to start from where I left off parameters wise. For this, current persistence implementation might be fine. For serving, it is probably not. You want to to make sure that serving can be done either in Flink or in a standalone Java app, for example. Tensorflow is trying to use the same format for both, but I do not see a big problem separating the two. For serving you might want to try to stick with one of the existing standards to avoid writing your own evaluator. 3. For export, I like the separation of the exporter from pipelines (see https://openscoring.io/blog/2018/07/09/converting_sparkml_pipeline_pmml/). In this case you can roll out additional exporters without modifying the existing code. Actually using the same approach for persistence allows for more flexibility as well ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
