【PYTHON OPENCV】Saving a linear regression model using SavedModelBuilder in TensorFlow

 """

Saving a linear regression model using SavedModelBuilder in TensorFlow

"""

# Import required packages:

import numpy as np

import tensorflow.compat.v1 as tf

tf.disable_v2_behavior()

from tensorflow.python.saved_model import signature_constants

from tensorflow.python.saved_model import signature_def_utils

def export_model():

  """Exports the model"""

  trained_checkpoint_prefix = 'linear_regression'

  loaded_graph = tf.Graph()

  with tf.Session(graph=loaded_graph) as sess:

    sess.run(tf.global_variables_initializer())

    # Restore from checkpoint:

    loader = tf.train.import_meta_graph(trained_checkpoint_prefix + '.meta')

    loader.restore(sess, trained_checkpoint_prefix)

    # Add signature:

    graph = tf.get_default_graph()

    inputs = tf.saved_model.utils.build_tensor_info(graph.get_tensor_by_name('X:0'))

    outputs = tf.saved_model.utils.build_tensor_info(graph.get_tensor_by_name('y_model:0'))

    signature = signature_def_utils.build_signature_def(inputs={'X': inputs},

                                                        outputs={'y_model': outputs},

                                                        method_name=signature_constants.PREDICT_METHOD_NAME)

    

    signature_map = {signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: signature}

    

    # Export model:

    builder = tf.saved_model.builder.SavedModelBuilder('./model')

    builder.add_meta_graph_and_variables(sess, signature_def_map=signature_map,tags=[tf.saved_model.tag_constants.SERVING])

    builder.save()

                                         

                                         

# Export the model:

export_model()

# Define 'M' more points to get the predictions using the trained model:

new_x = np.linspace(50 + 1, 50 + 10, 3)

with tf.Session(graph=tf.Graph()) as sess:

  tf.saved_model.loader.load(sess, [tf.saved_model.tag_constants.SERVING], './my_model')

  graph = tf.get_default_graph()

  x = graph.get_tensor_by_name('X:0')

  model = graph.get_tensor_by_name('y_model:0')

  print(sess.run(model, {x: new_x}))