Including BEAM preprocessing graph in Keras models at serving

Question

Short Question:

Since Tensorflow is moving towards Keras and away from Estimators, how can we incorporate our preprocessing pipelines e.g. using tf.Transform and build_serving_input_fn() (which are used for estimators), into our tf.keras models?

From my understanding, the only way to incorporate this preprocessing graph is to first build the model using Keras. Train it. Then export it as an estimator using tf.keras.estimator.model_to_estimator. Then create a serving_input_fn and export the estimator as a saved model, along with this serving_input_fn to be used at serving time.

To me it seems tedious, and not the correct way of doing things. Instead, I would like to go directly from Keras to Saved Model.

Problem

I would like to be able to include APAHCE BEAM preprocessing graph in a Keras saved model.

I would like to serve the trained Keras model, hence I export it using SavedModel. Given a trained model, I would like to apply the following logic to predict the outcome.

raw_features = { 'feature_col_name': [100] } # features to transform
working_dir = 'gs://path_to_transform_fn' 

# transform features
transformed_features = tf_transform_output.transform_raw_features(raw_features)

model = load_model('model.h5')
model.predict(x=transformed_features)

When I define my model, I use Functional API and the model has the following inputs:

for i in numerical_features:
    num_inputs.append(tf.keras.layers.Input(shape=(1,), name=i))

This is the problem, because tensors are not inputed directly into keras from tf.Dataset, but instead, are linked using the Input() layer.

When I export the model using tf.contrib.saved_model.save_keras_model(model=model, saved_model_path=saved_model_path), I can readily serve predictions if I handle the preprocessing in a separate script. The output of this would look like

Is this what generally happens? For example, I would preprocess the features as part of some external script, and then send transformed_features to the model for prediction.

Ideally, it would all happen within the Keras model/part of a single graph. Currently it seems that I'm using the output of one graph as an input into another graph. Instead I would like to be able to use a single graph.

If using Estimators, we can build a serving_input_fn() which can be included as an argument to the estimator, which allows us to incorporate preprocessing logic into the graph.

I would like to also hear your Keras + SavedModel + Preprocessing ideas on serving models using Cloud ML

Are you using tf2.0? I haven't used Beam, but to include some simple preprocessing (normalizing and imputing) into a graph for serving I did this: 1. Use keras Inputs. 2. Use low level tf ops for preprocessing (the perhaps less than ideal part, but not too bad for my use case) `filled_nas_feature = tf.where(tf.math.is_nan(my_feature), tf.zeros_like(my_feature), my_feature)` 3. Use multiple Keras models to encapsulate preprocessing, training, and postprocessing as callable functions. 4. Serialize the model to serve using `tf.saved_model.save(exportModel, './path/to/model')`. — James McKeown, May 14 '19 at 21:00

score 1 · Answer 1 · answered Jun 12 '19 at 15:50

For your question on incorporating Apache Beam into the input function for a tf.transform pipeline, see this TF tutorial that explains how to do that:

"https://www.tensorflow.org/tfx/transform/get_started#apache_beam_implementation

On using TF 2.0 SavedModel with Keras, this notebook tutorial demonstrates how to do that:

https://www.tensorflow.org/beta/guide/keras/saving_and_serializing#export_to_savedmodel

score -2 · Answer 2 · answered Mar 24 '19 at 09:38

The Cloud ML is a way of Google Clouds's Machine learning.

It's quite simple to Get Started and train the UI using their documentation:

Develop and validate your training application locally

Before you run your training application in the cloud, get it running locally. Local environments provide an efficient development and validation workflow so that you can iterate quickly. You also won't incur charges for cloud resources when debugging your application locally. Get your training data

The relevant data files, adult.data and adult.test, are hosted in a public Cloud Storage bucket. For purposes of this sample, use the versions on Cloud Storage, which have undergone some trivial cleaning, instead of the original source data. See below for more information about the data.

You can read the data files directly from Cloud Storage or copy them to your local environment. For purposes of this sample you will download the samples for local training, and later upload them to your own Cloud Storage bucket for cloud training. Download the data to a folder:

mkdir data
gsutil -m cp gs://cloud-samples-data/ml-engine/census/data/* data/

Then, just set the TRAIN_DATA AND EVAL_DATA variables to your local file paths. For example, the following commands set the variables to local paths.

TRAIN_DATA=$(pwd)/data/adult.data.csv
EVAL_DATA=$(pwd)/data/adult.test.csv

Then you have a TSV file like this:

39, State-gov, 77516, Bachelors, 13, Never-married, Adm-clerical, Not-in-family, White, Male, 2174, 0, 40, United-States, <=50K
50, Self-emp-not-inc, 83311, Bachelors, 13, Married-civ-spouse, Exec-managerial, Husband, White, Male, 0, 0, 13, United-States, <=50K
38, Private, 215646, HS-grad, 9, Divorced, Handlers-cleaners, Not-in-family, White, Male, 0, 0, 40, United-States, <=50K
53, Private, 234721, 11th, 7, Married-civ-spouse, Handlers-cleaners, Husband, Black, Male, 0, 0, 40, United-States, <=50K

To run it:

gcloud ml-engine local train \
    --module-name trainer.task \
    --package-path trainer/ \
    --job-dir $MODEL_DIR \
    -- \
    --train-files $TRAIN_DATA \
    --eval-files $EVAL_DATA \
    --train-steps 1000 \
    --eval-steps 100

For moreTraining considerations like your question says:

Running a Training Job

Cloud Machine Learning Engine provides model training as an asynchronous (batch) service. This page describes how to configure and submit a training job by running gcloud ml-engine jobs submit training from the command line or by sending a request to the API at projects.jobs.create. Before you begin

Before you can submit a training job, you must package your application and upload it and any unusual dependencies to a Cloud Storage bucket. Note: If you use the gcloud command-line tool to submit your job, you can package the application and submit the job in the same step. Configuring the job

You pass your parameters to the training service by setting the members of the Job resource, which includes the items in the TrainingInput resource.

If you use the gcloud command-line tool to submit your training jobs, you can:
Specify the most common training parameters as flags of the gcloud ml-engine jobs submit training command.
Pass the remaining parameters in a YAML configuration file, named config.yaml by convention. The configuration file mirrors the
structure of the JSON representation of the Job resource. You pass the path of your configuration file in the --config flag of the gcloud ml-engine jobs submit training command. So, if the path to your configuration file is config.yaml, you must set --config=config.yaml.

Gathering the job configuration data

The following properties are used to define your job.

Job name (jobId) A name to use for the job (mixed-case letters, numbers, and underscores only, starting with a letter). Cluster configuration (scaleTier) A scale tier specifying the type of processing cluster to run your job on. This can be the CUSTOM scale tier, in which case you also explicitly specify the number and type of machines to use. Training application package (packageUris) A packaged training application that is staged in a Cloud Storage location. If you are using the gcloud command-line tool, the application packaging step is largely automated. See the details in the guide to packaging your application. Module name (pythonModule) The name of the main module in your package. The main module is the Python file you call to start the application. If you use the gcloud command to submit your job, specify the main module name in the --module-name flag. See the guide to packaging your application. Region (region) The Compute Engine region where you want your job to run. You should run your training job in the same region as the Cloud Storage bucket that stores your training data. See the available regions for Cloud ML Engine services. Job directory (jobDir) The path to a Cloud Storage location to use for job output. Most training applications save checkpoints during training and save the trained model to a file at the end of the job. You need a Cloud Storage location to save them to. Your Google Cloud Platform project must have write access to this bucket. The training service automatically passes the path you set for the job directory to your training application as a command-line argument named job_dir. You can parse it along with your application's other arguments and use it in your code. The advantage to using the job directory is that the training service validates the directory before starting your application. Runtime version (runtimeVersion)
The Cloud ML Engine version to use for the job. If you don't specify a runtime version, the training service uses the default Cloud
ML Engine runtime version 1.0. Python version (pythonVersion)
The Python version to use for the job. Python 3.5 is available with Cloud ML Engine runtime version 1.4 or greater. If you don't
specify a Python version, the training service uses Python 2.7.

Formatting your configuration parameters

How you specify your configuration details depends on how you are starting your training job: Provide the job configuration details to the gcloud ml-engine jobs submit training command. You can do this in two ways:
With command-line flags.
In a YAML file representing the Job resource. You can name this file whatever you want. By convention the name is config.yaml.
Even if you use a YAML file, certain details must be supplied as command-line flags. For example, you must provide the --module-name flag and at least one of --package-path or --packages. If you use --package-path, you must also include --job-dir or --staging-bucket. Additionally, you must either provide the --region flag or set a default region for your gcloud client. These options—and any others you provide as command line flags—will override values for those options in your configuration file.

Example 1: In this example, you choose a preconfigured machine cluster and supply all the required details as command-line flags when submitting the job. No configuration file is necessary. See the guide to submitting the job in the next section.

Example 2: The following example shows the contents of the configuration file for a job with a custom processing cluster. The configuration file includes some but not all of the configuration details, assuming that you supply the other required details as command-line flags when submitting the job.

trainingInput: scaleTier: CUSTOM masterType: complex_model_m
workerType: complex_model_m parameterServerType: large_model
workerCount: 9 parameterServerCount: 3 runtimeVersion: '1.13'
pythonVersion: '3.5'

The above example specifies Python version 3.5, which is available when you use Cloud ML Engine runtime version 1.4 or greater.

Submitting the job

When submitting a training job, you specify two sets of flags:
Job configuration parameters. Cloud ML Engine needs these values to set up resources in the cloud and deploy your application on each
node in the processing cluster. User arguments, or application parameters. Cloud ML Engine passes the value of these flags through to your application.

Submit a training job using the gcloud ml-engine jobs submit training command.

First, it's useful to define some environment variables containing your configuration details. To create a job name, the following code appends the date and time to the model name:
TRAINER_PACKAGE_PATH="/path/to/your/application/sources"
now=$(date +"%Y%m%d_%H%M%S")
JOB_NAME="your_name_$now"
MAIN_TRAINER_MODULE="trainer.task"
JOB_DIR="gs://your/chosen/job/output/path"
PACKAGE_STAGING_PATH="gs://your/chosen/staging/path"
REGION="us-east1"
RUNTIME_VERSION="1.13"
The following job submission corresponds to configuration example 1 above, where you choose a preconfigured scale tier (basic) and you decide to supply all the configuration details via command-line flags. There is no need for a config.yaml file:
gcloud ml-engine jobs submit training $JOB_NAME \
        --scale-tier basic \
        --package-path $TRAINER_PACKAGE_PATH \
        --module-name $MAIN_TRAINER_MODULE \
        --job-dir $JOB_DIR \
        --region $REGION \
        -- \
        --user_first_arg=first_arg_value \
        --user_second_arg=second_arg_value
The following job submission corresponds to configuration example 2 above, where some of the configuration is in the file and you supply the other details via command-line flags:
gcloud ml-engine jobs submit training $JOB_NAME \
        --package-path $TRAINER_PACKAGE_PATH \
        --module-name $MAIN_TRAINER_MODULE \
        --job-dir $JOB_DIR \
        --region $REGION \
        --config config.yaml \
        -- \
        --user_first_arg=first_arg_value \
        --user_second_arg=second_arg_value
Notes:
If you specify an option both in your configuration file (config.yaml) and as a command-line flag, the value on the command
line overrides the value in the configuration file. The empty -- flag marks the end of the gcloud specific flags and the start of the USER_ARGS that you want to pass to your application. Flags specific to Cloud ML Engine, such as --module-name, --runtime-version, and --job-dir, must come before the empty -- flag. The Cloud ML Engine service interprets these flags. The --job-dir flag, if specified, must come before the empty -- flag, because Cloud ML Engine uses the --job-dir to validate the path. Your application must handle the --job-dir flag too, if specified. Even though the flag comes before the empty --, the --job-dir is also passed to your application as a command-line flag. You can define as many USER_ARGS as you need. Cloud ML Engine passes --user_first_arg, --user_second_arg, and so on, through to your application.

I'm sorry but the question is not on how to use CloudML, but how to incorporate Apache Beam preprocessing graphs into Keras instead of Estimators. — GRS, Mar 24 '19 at 16:20

Including BEAM preprocessing graph in Keras models at serving

Problem

2 Answers2