Exporting a TensorFlow model

Last updated 7th February, 2020.

Objective

Tensorflow is a popular machine learning library and SavedModel is a serialization format that is supported by OVHcloud ML Serving. This tutorial will cover how to export a Tensorflow trained model into an SavedModel file.

Requirements

• A python environment with Tensorflow installed

Convert a simple model to SavedModel

ML Serving supports TensorFlow models thanks to the SavedModel serialization format of TensorFlow.

Let's take a simple example of a TensorFlow model to illustrate:

from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split

import tensorflow as tf

# Loading data
iris = load_iris()
X, y = iris.data, iris.target
X_train, X_test, y_train, y_test = train_test_split(X, y)

# Creating classifier
classifier = tf.estimator.DNNClassifier(
    feature_columns=[tf.feature_column.numeric_column('x', shape=[4])],
    hidden_units=[10, 20, 10],
    n_classes=3
)
# Training classifier
classifier.train(
    input_fn=tf.estimator.inputs.numpy_input_fn(
        x={'x': X_train},
        y=y_train,
        # batch_size=5,
        num_epochs=1000,
        shuffle=True
    ),
    steps=100
)

Define transformation function you need

If you want to export your data transformation along with your trained model, you should describe the mapping between them using a ServingInputReceiver.

INPUT_FEATURES = ['sepal_length', 'sepal_width', 'petal_length', 'petal_width']

def serving_input_receiver_fn():
    """
    This is used to define inputs to serve the model.
   :return: ServingInputReceiver
    """
    input_features = {feature: tf.placeholder(tf.float32, [None, 1]) for feature in INPUT_FEATURES}
    model_features = {'x', tf.concat([receiver_tensors[feature] for feature in INPUT_FEATURES]), axis=1)

    return tf.estimator.export.ServingInputReceiver(
        receiver_tensors=input_features,
        features=model_features
    )

In the previous example, four tensors are transformed:

• sepal_length of shape [None, 1]
• sepal_width of shape [None, 1]
• petal_length of shape [None, 1]
• petal_width of shape [None, 1]

Into a single tensor x of shape [None, 4] before feeding the classifier. Note: The None value in a tensor represents a dimension of arbitrary length. For example, if you are using a classification algorithm, None can simply be the number of records you wish to classify.

Launch the conversion and save it into a file

The trained model conversion is made by calling the export_saved_model function on your classifier. This function takes two parameters:

• export_dir_base: The path where to save the serialized model
• serving_input_receiver_fn: The transformation function used at serving time to feed your input to the trained classifier.

classifier.export_saved_model(
    export_dir_base='<path/where/to/saved/serialized/model>',
    serving_input_receiver_fn=serving_input_receiver_fn
)

Going further

• You can check the OVHcloud documentation on how to deploy custom models.
• You can check the supported compatibilities for Tensorflow SavedModel

These guides might also interest you...