feature: add TensorFlow 1.13 support (aws#860)

Author: icywang86rui

README.rst

@@ -173,7 +173,7 @@ TensorFlow SageMaker Estimators
 
 By using TensorFlow SageMaker Estimators, you can train and host TensorFlow models on Amazon SageMaker.
 
-Supported versions of TensorFlow: ``1.4.1``, ``1.5.0``, ``1.6.0``, ``1.7.0``, ``1.8.0``, ``1.9.0``, ``1.10.0``, ``1.11.0``, ``1.12.0``.
+Supported versions of TensorFlow: ``1.4.1``, ``1.5.0``, ``1.6.0``, ``1.7.0``, ``1.8.0``, ``1.9.0``, ``1.10.0``, ``1.11.0``, ``1.12.0``, ``1.13.1``.
 
 Supported versions of TensorFlow for Elastic Inference: ``1.11.0``, ``1.12.0``.
 

doc/using_tf.rst

@@ -443,20 +443,10 @@ After a TensorFlow estimator has been fit, it saves a TensorFlow SavedModel in
 the S3 location defined by ``output_path``. You can call ``deploy`` on a TensorFlow
 estimator to create a SageMaker Endpoint.
 
-SageMaker provides two different options for deploying TensorFlow models to a SageMaker
-Endpoint:
+Your model will be deployed to a TensorFlow Serving-based server. The server provides a super-set of the
+`TensorFlow Serving REST API <https://www.tensorflow.org/serving/api_rest>`_.
 
-- The first option uses a Python-based server that allows you to specify your own custom
-  input and output handling functions in a Python script. This is the default option.
-
-  See `Deploying to Python-based Endpoints <https://github.com/aws/sagemaker-python-sdk/blob/master/src/sagemaker/tensorflow/deploying_python.rst>`_ to learn how to use this option.
-
-
-- The second option uses a TensorFlow Serving-based server to provide a super-set of the
-  `TensorFlow Serving REST API <https://www.tensorflow.org/serving/api_rest>`_. This option
-  does not require (or allow) a custom python script.
-
-  See `Deploying to TensorFlow Serving Endpoints <https://github.com/aws/sagemaker-python-sdk/blob/master/src/sagemaker/tensorflow/deploying_tensorflow_serving.rst>`_ to learn how to use this option.
+See `Deploying to TensorFlow Serving Endpoints <https://github.com/aws/sagemaker-python-sdk/blob/master/src/sagemaker/tensorflow/deploying_tensorflow_serving.rst>`_ to learn how to deploy your model and make inference requests.
 
 
 SageMaker TensorFlow Docker containers

src/sagemaker/tensorflow/deploying_python.rst

@@ -26,7 +26,7 @@
 from sagemaker.tensorflow.defaults import TF_VERSION
 from sagemaker.tensorflow.model import TensorFlowModel
 from sagemaker.tensorflow.serving import Model
-from sagemaker.utils import get_config_value
+from sagemaker import utils
 from sagemaker.vpc_utils import VPC_CONFIG_DEFAULT
 
 logger = logging.getLogger("sagemaker")
@@ -190,9 +190,11 @@ class TensorFlow(Framework):
 
     __framework_name__ = "tensorflow"
 
-    LATEST_VERSION = "1.12"
+    LATEST_VERSION = "1.13"
     """The latest version of TensorFlow included in the SageMaker pre-built Docker images."""
 
+    _LOWEST_SCRIPT_MODE_ONLY_VERSION = [1, 13]
+
     def __init__(
         self,
         training_steps=None,
@@ -321,6 +323,17 @@ def _validate_args(
                     )
                 )
 
+        if (not self._script_mode_enabled()) and self._only_script_mode_supported():
+            logger.warning(
+                "Legacy mode is deprecated in versions 1.13 and higher. Using script mode instead."
+            )
+            self.script_mode = True
+
+    def _only_script_mode_supported(self):
+        return [
+            int(s) for s in self.framework_version.split(".")
+        ] >= self._LOWEST_SCRIPT_MODE_ONLY_VERSION
+
     def _validate_requirements_file(self, requirements_file):
         if not requirements_file:
             return
@@ -489,7 +502,7 @@ def _create_tfs_model(self, role=None, vpc_config_override=VPC_CONFIG_DEFAULT):
             image=self.image_name,
             name=self._current_job_name,
             container_log_level=self.container_log_level,
-            framework_version=self.framework_version,
+            framework_version=utils.get_short_version(self.framework_version),
             sagemaker_session=self.sagemaker_session,
             vpc_config=self.get_vpc_config(vpc_config_override),
         )
@@ -553,7 +566,7 @@ def hyperparameters(self):
         return hyperparameters
 
     def _default_s3_path(self, directory, mpi=False):
-        local_code = get_config_value("local.local_code", self.sagemaker_session.config)
+        local_code = utils.get_config_value("local.local_code", self.sagemaker_session.config)
         if self.sagemaker_session.local_mode and local_code:
             return "/opt/ml/shared/{}".format(directory)
         elif mpi:

src/sagemaker/tensorflow/estimator.py

@@ -123,6 +123,18 @@ def get_config_value(key_path, config):
     return current_section
 
 
+def get_short_version(framework_version):
+    """Return short version in the format of x.x
+
+    Args:
+        framework_version: The version string to be shortened.
+
+    Returns:
+        str: The short version string
+    """
+    return ".".join(framework_version.split(".")[:2])
+
+
 def to_str(value):
     """Convert the input to a string, unless it is a unicode string in Python 2.
 

src/sagemaker/utils.py

@@ -14,17 +14,11 @@
 
 import argparse
 import json
-import logging as _logging
 import numpy as np
 import os
-import sys as _sys
 import tensorflow as tf
-from tensorflow.python.platform import tf_logging
 
 tf.logging.set_verbosity(tf.logging.DEBUG)
-_handler = _logging.StreamHandler(_sys.stdout)
-tf_logger = tf_logging._get_logger()
-tf_logger.handlers = [_handler]
 
 
 def cnn_model_fn(features, labels, mode):
@@ -179,5 +173,3 @@ def serving_input_fn():
 
     if args.current_host == args.hosts[0]:
         mnist_classifier.export_savedmodel("/opt/ml/model", serving_input_fn)
-
-    tf_logger.info("====== Training finished =========")

tests/data/tensorflow_mnist/mnist.py

@@ -85,14 +85,14 @@ def _create_model(output_path):
 
 @pytest.mark.local_mode
 @pytest.mark.skipif(PYTHON_VERSION != "py2", reason="TensorFlow image supports only python 2.")
-def test_tf_local_mode(tf_full_version, sagemaker_local_session):
+def test_tf_local_mode(sagemaker_local_session):
     with timeout(minutes=5):
         script_path = os.path.join(DATA_DIR, "iris", "iris-dnn-classifier.py")
 
         estimator = TensorFlow(
             entry_point=script_path,
             role="SageMakerRole",
-            framework_version=tf_full_version,
+            framework_version="1.12",
             training_steps=1,
             evaluation_steps=1,
             hyperparameters={"input_tensor_name": "inputs"},
@@ -135,6 +135,7 @@ def test_tf_distributed_local_mode(sagemaker_local_session):
         estimator = TensorFlow(
             entry_point=script_path,
             role="SageMakerRole",
+            framework_version="1.12",
             training_steps=1,
             evaluation_steps=1,
             hyperparameters={"input_tensor_name": "inputs"},
@@ -176,6 +177,7 @@ def test_tf_local_data(sagemaker_local_session):
         estimator = TensorFlow(
             entry_point=script_path,
             role="SageMakerRole",
+            framework_version="1.12",
             training_steps=1,
             evaluation_steps=1,
             hyperparameters={"input_tensor_name": "inputs"},
@@ -216,6 +218,7 @@ def test_tf_local_data_local_script():
         estimator = TensorFlow(
             entry_point=script_path,
             role="SageMakerRole",
+            framework_version="1.12",
             training_steps=1,
             evaluation_steps=1,
             hyperparameters={"input_tensor_name": "inputs"},