TensorFlow is a popular library to train and use machine learning models. We integrate it with Jet to classify stream of events with the result of a TF model execution.
TensorFlow provides two ways of running models:
- In-process: load the model into memory and execute it
- In ModelServer: a request-response based service that executes the model and can also handle multiple versions of the models etc. The calls can be made using gRPC or through a REST API.
Each way has pros and cons and discussion of them is beyond the scope of this sample.
As a sample, we took the Large Movie Reviews Dataset as provided by the
TensorFlow Keras Datasets. We trained a model as described in a
tutorial
and save it. The result of the training is in the
data/model subdirectory.
The input of the model is a the text of a review, the output is probability of the review being positive or negative, a number in range of 0..1. The words in the input are encoded using the word index.
You can run the in-process classification using the following command:
$ mvn compile exec:java -Dexec.mainClass=InProcessClassification -Dexec.args="data"
To run the classification on a model server, you'll need Docker installed. It's possible to create a model server using the following command:
$ cd {hazelcast-jet-code-samples}
$ docker run -p 8500:8500 --mount \
type=bind,source=$(pwd)/tensorflow/data/model,target=/models/reviewSentiment \
-e MODEL_NAME=reviewSentiment -t tensorflow/serving
Then you can run the ModelServerClassification sample:
$ mvn compile exec:java -Dexec.mainClass=ModelServerClassification \
-Dexec.args="data 0.0.0.0:8500"
If you want to re-train the model yourself, you need to have python installed. Since TensorFlow 1.12 doesn't support the current python version, you can use python 3.4, 3.5 or 3.6 to install tensorflow using:
$ pip install tensorflow
Then, run the provided script to download the source dataset, train and save the model:
$ cd {hazelcast-jet-code-samples}/tensorflow
$ python bin/imdb_review_train.py
After this, data/model subdirectory should be newly written with the
following contents:
data/model/1/saved_model.pb
data/model/1/variables/variables.data-00000-of-00001
data/model/1/variables/variables.index
data/imdb_word_index.json