Upgrade to TF 1.12 (#31)

* Upgrade to TF 1.12

Author: owen-t

create_integ_test_docker_images.py

@@ -9,7 +9,7 @@
 import glob
 import sys
 
-TF_VERSION = "1.11.0"
+TF_VERSION = "1.12.0"
 REGION = "us-west-2"
 
 if __name__ == '__main__':

setup.py

@@ -94,7 +94,7 @@ def build_extension(self, ext):
 
 setup(
     name='sagemaker_tensorflow',
-    version='1.11.0.1.0.0',
+    version='1.12.0.1.0.0',
     description='Amazon Sagemaker specific TensorFlow extensions.',
     packages=find_packages(where='src', exclude=('test',)),
     package_dir={'': 'src'},

src/sagemaker_tensorflow/pipemode.py

@@ -71,6 +71,10 @@ def __init__(self, channel, record_format='RecordIO',
     def _as_variant_tensor(self):
         return self._tf_plugin.pipe_mode_dataset(self.benchmark, self.record_format, self.state_dir, self.channel,
                                                  self.pipe_dir)
+
+    def _inputs(self):
+        return []
+
     def _validate_input_data_config(self):
         if self.channel not in self.input_data_config:
             raise PipeModeDatasetException("Channel {} not found in Training Job InputDataConfig".format(self.channel))

test/integ/Dockerfile

@@ -1,7 +1,7 @@
 FROM ubuntu:16.04
 
 ARG device=cpu
-ARG tensorflow_version=1.9.0
+ARG tensorflow_version=1.12.0
 ARG script
 ARG python
 

test/integ/scripts/estimator_script.py

@@ -6,8 +6,9 @@
 import tensorflow as tf
 from sagemaker_tensorflow import PipeModeDataset
 
-ds = PipeModeDataset("elizabeth")
+print("Starting estimator script")
 
+ds = PipeModeDataset("elizabeth")
 
 class BenchmarkConfig(object):
 
@@ -74,24 +75,37 @@ def parse(record):
 model_dir = tempfile.mkdtemp()
 estimator = tf.estimator.LinearClassifier(feature_columns=[column])
 
+print("About to call train")
 estimator.train(input_fn=input_fn)
 
 # Confirm that we have read the correct number of pipes
 assert os.path.exists('/opt/ml/input/data/{}_{}'.format(config.channel, config.epochs + 1))
 
+print("About to call evaluate")
+result = estimator.evaluate(input_fn=input_fn)
+for key,value in sorted(result.items()):
+    print('%s: %s' % (key, value))
+
+
 # Test that we can create a new PipeModeDataset after training has run
+print("Validate that new PipeModeDataset on existing channel can be created")
+
 ds = PipeModeDataset(config.channel)
 
 with tf.Session() as sess:
     it = ds.make_one_shot_iterator()
     next = it.get_next()
     sess.run(next)
 
+print("Validate create, read, discard, recreate")
+
 # Test that we can create a PipeModeDataset, discard it, and read from a new one
 ds = PipeModeDataset(config.channel)
 with tf.Session() as sess:
     it = ds.make_one_shot_iterator()
     next = it.get_next()
+
+
 ds = PipeModeDataset(config.channel)
 with tf.Session() as sess:
     it = ds.make_one_shot_iterator()

tox.ini

@@ -49,7 +49,7 @@ deps =
     mock
     contextlib2
     teamcity-messages
-    tensorflow==1.11
+    tensorflow==1.12
     awslogs
     docker
     cmake
@@ -60,5 +60,5 @@ deps =
     cmake
     flake8
     flake8-future-import
-    tensorflow==1.11
+    tensorflow==1.12
 commands = flake8