CODEWITHSUNDEEP
pythontensorflowtensorflow-serving
palazzo train
palazzo train

Reputation: 3409

How to save a tensorflow estimator by export_savedmodel, then load and use it locally later?

I have trained a tf linear regression estimator as below:

sample_size = train_x.shape[0]
feature_size = train_x.shape[1]

feature_columns = [tf.feature_column.numeric_column("x", shape=[feature_size])]

lr_estimator = tf.estimator.LinearRegressor(feature_columns=feature_columns )

train_x_mat = train_x.as_matrix()
test_x_mat = test_x.as_matrix()


# Define the training inputs
train_input_fn = tf.estimator.inputs.numpy_input_fn(
    x={"x": train_x_mat},
    y=np.array(train_y_mat),
    num_epochs=None,
    shuffle=True)

# Train model.
lr_estimator.train(input_fn=train_input_fn, steps=2000)

where train_x and train_y are pandas dataframe. the lr_estimator does work and I can call .predict successfully.

How I can save it to a file, and then load it back for prediction later? I am only trying to build a small python program. The prediction program will be run on the same desktop. I don't need complicated server serving, yet.

Upvotes: 2

Views: 650

Answers (1)

Ilredirana
Ilredirana

Reputation: 21

def serving_input_receiver_fn():
    """
    input placeholder
    """
    inputs = {"x": tf.placeholder(shape=[feature_size], dtype=tf.float32)}
    return tf.estimator.export.ServingInputReceiver(inputs, inputs)

# export model and weights
export_dir = est_inception_v3.export_savedmodel(export_dir_base="/export_dir", 
    serving_input_receiver_fn=serving_input_receiver_fn)

# restore from disk
with tf.Session() as sess:
    tf.saved_model.loader.load(sess, [tf.saved_model.tag_constants.SERVING], export_dir)
    predictor = SavedModelPredictor(export_dir)
    print(predictor({"x": test_x_mat}))

Upvotes: 2

Related Questions