TensorFlow Evaluator

In this example, we'll look at how to configure the TensorFlow Evaluator processor to evaluate each record using the sample Iris classification model.

In this example, we'll create the following simple pipeline:

We just want to preview the TensorFlow Evaluator processor output in our example, so we'll simply use the Trash destination. Feel free to use another destination of your choice.

Complete the following steps to configure the TensorFlow Evaluator processor to predict the species of Iris flowers:

1. Download the sample Iris tensor model from the StreamSets Data Collector GitHub repository, and the store the model in the Data Collector resources directory, $SDC_RESOURCES.
   For example:

   /var/lib/sdc-resources/iris_saved_model

   Because the sample Iris classification model has already been built and trained, we can skip that prerequisite for this example. For more information about how the model was built and trained, see Premade Estimators in the TensorFlow documentation.

2. Create a Data Collector pipeline and add a Dev Raw Data Source origin.
   Add the following raw data to the origin, and then configure the origin to process the JSON data format:

   {
     "petalLength": 6.4,
     "petalWidth": 2.8,
     "sepalLength": 5.6,
     "sepalWidth": 2.2
   },
   {
     "petalLength": 5.0,
     "petalWidth": 2.3,
     "sepalLength": 3.3,
     "sepalWidth": 1.0
   },
   {
     "petalLength": 4.9,
     "petalWidth": 2.5,
     "sepalLength": 4.5,
     "sepalWidth": 1.7
   },
   {
     "petalLength": 4.9,
     "petalWidth": 3.1,
     "sepalLength": 1.5,
     "sepalWidth": 0.1
   }

3. The sample Iris model expects float data as input, so add a Field Type Converter processor to the pipeline.
   On the Conversions tab of the processor, set the conversion method to By Field Name, and then configure the conversion properties as follows to convert all fields to float:

   • Fields to Convert - /*
   • Conversion Type - FLOAT
4. Add a TensorFlow Evaluator processor to the pipeline.

   On the General tab, ensure that Produce Events is cleared because we need the processor to evaluate each record.

   On the TensorFlow tab, define the saved model path and model tags as follows:

   • Saved Model Path - Enter the absolute path where you saved the model on the Data Collector machine. For example: /var/lib/sdc-resources/iris_saved_model.
   • Model Tags - serve
   Configure four input configurations with the following values:

   Operation	Index	Fields to Convert	Shape	Tensor Type
   PetalLength	0	/petalLength	1	FLOAT
   PetalWidth	0	/petalWidth	1	FLOAT
   SepalLength	0	/sepalLength	1	FLOAT
   SepalWidth	0	/sepalWidth	1	FLOAT

   For example, the first input configuration should look like this:

   

   Configure two output configurations with the following values:

   Operation	Index	Tensor Type
   dnn/head/predictions/ExpandDims	0	FLOAT
   dnn/head/predictions/probabilities	0	FLOAT

   Ensure that Entire Batch is cleared, and then leave the default value of /output for the Output Field.

5. Finally, add a Trash destination to the pipeline.

When we run preview and view the output of the TensorFlow Evaluator processor, we can see that each field in the record is passed through the processor untouched. The processor generates an output map field for each record that contains the prediction results in the two fields that we defined in the output configurations. As shown below, the processor predicts the species of the first record to be 2, representing the Iris virginica species:

The TensorFlow Evaluator processor generates a tensorflow-event record when the processor completes processing all records in the batch. The event record is a Map field containing a field for each output configuration that you define for the processor.