Upgrade to TensorFlow 2 (#40)

Changes our uses of TensorFlow APIs to be compatible with TensorFlow 2, mostly automatically through the tf_upgrade_v2 command. This change has been validated to work with all capsules in the Capsule Zoo as well as our internally developed capsules at Aotu.ai.

This aims to make OpenVisionCapsules TensorFlow 2 compatible without any features disabled, but does not attempt to upgrade our use of old v1 APIs.

While this upgrade requires no changes to the API of vcap or vcap-utils, it is still a breaking change for any capsule that has to use TensorFlow directly.

Fixes #39

vcap/setup.py

@@ -24,7 +24,7 @@
         "scipy==1.4.1",
         "scikit-learn==0.22.2",
         "numpy>=1.16,<2",
-        "tensorflow-gpu==1.15.4",
+        "tensorflow~=2.5.0",
     ],
     extras_require={
         "tests": test_packages,

vcap/vcap/device_mapping.py

@@ -25,7 +25,7 @@ def get_all_devices() -> List[str]:
             #
             # TODO: Use tf.config.list_physical_devices in TF 2.1
 
-            with tf.Session():
+            with tf.compat.v1.Session():
                 all_devices = device_lib.list_local_devices()
 
             # Get the device names and remove duplicates, just in case...

vcap_utils/vcap_utils/backends/crowd_density.py

@@ -14,7 +14,7 @@ class CrowdDensityCounter(BaseTFBackend):
 
     def __init__(self, model_bytes,
                  device: str=None,
-                 session_config: tf.ConfigProto=None):
+                 session_config: tf.compat.v1.ConfigProto=None):
         """
         :param model_bytes: Model file data, likely a loaded *.pb file
         :param device: The device to run the model on

vcap_utils/vcap_utils/backends/depth.py

@@ -19,7 +19,7 @@ class DepthPredictor(BaseTFBackend):
 
     def __init__(self, model_bytes,
                  device: str=None,
-                 session_config: tf.ConfigProto=None):
+                 session_config: tf.compat.v1.ConfigProto=None):
         """
         :param model_bytes: Model file data, likely a loaded *.pb file
         :param device: The device to run the model on

vcap_utils/vcap_utils/backends/load_utils.py

@@ -5,7 +5,7 @@
 
 def parse_tf_model_bytes(model_bytes,
                          device: str = None,
-                         session_config: tf.ConfigProto = None):
+                         session_config: tf.compat.v1.ConfigProto = None):
     """
 
     :param model_bytes: The bytes of the model to load
@@ -18,7 +18,7 @@ def parse_tf_model_bytes(model_bytes,
     detection_graph = tf.Graph()
     with detection_graph.as_default():
         # Load a (frozen) Tensorflow model from memory
-        graph_def = tf.GraphDef()
+        graph_def = tf.compat.v1.GraphDef()
         graph_def.ParseFromString(model_bytes)
 
         with tf.device(device):
@@ -29,16 +29,16 @@ def parse_tf_model_bytes(model_bytes,
                                 name='')
 
     if session_config is None:
-        session_config = tf.ConfigProto()
+        session_config = tf.compat.v1.ConfigProto()
 
     if device is not None:
         # allow_soft_placement lets us remap GPU only ops to GPU, and doesn't
         # crash for non-gpu only ops (it will place those on CPU, instead)
         session_config.allow_soft_placement = True
 
     # Create a session for later use
-    persistent_sess = tf.Session(graph=detection_graph,
-                                 config=session_config)
+    persistent_sess = tf.compat.v1.Session(graph=detection_graph,
+                                           config=session_config)
 
     return detection_graph, persistent_sess
 

vcap_utils/vcap_utils/backends/openface_encoder.py

@@ -22,7 +22,7 @@ class OpenFaceEncoder(BaseEncoderBackend, BaseTFBackend):
 
     def __init__(self, model_bytes, model_name,
                  device: str = None,
-                 session_config: tf.ConfigProto = None):
+                 session_config: tf.compat.v1.ConfigProto = None):
         """
         :param model_bytes: Model file bytes, a loaded *.pb file
         :param model_name: The name of the model in order to load correct

vcap_utils/vcap_utils/backends/segmentation.py

@@ -16,7 +16,7 @@ class Segmenter(BaseTFBackend):
 
     def __init__(self, model_bytes, metadata_bytes,
                  device: str = None,
-                 session_config: tf.ConfigProto = None):
+                 session_config: tf.compat.v1.ConfigProto = None):
         """
         :param model_bytes: Model file data, likely a loaded *.pb file
         :param metadata_bytes: The dataset metadata file data, likely named

vcap_utils/vcap_utils/backends/tf_image_classification.py

@@ -15,7 +15,7 @@
 class TFImageClassifier(BaseTFBackend):
     def __init__(self, model_bytes, metadata_bytes, model_name,
                  device: str = None,
-                 session_config: tf.ConfigProto = None):
+                 session_config: tf.compat.v1.ConfigProto = None):
         """
         :param model_bytes: Loaded model data, likely from a *.pb file
         :param metadata_bytes: Loaded dataset metadata, likely from a file
@@ -41,7 +41,7 @@ def __init__(self, model_bytes, metadata_bytes, model_name,
         # Create the input node to the graph, with preprocessing built-in
         with self.graph.as_default():
             # Create a new input node for images of various sizes
-            self.input_node = tf.placeholder(
+            self.input_node = tf.compat.v1.placeholder(
                 dtype=tf.float32,
                 shape=[None, self.config.img_size, self.config.img_size, 3])
 

vcap_utils/vcap_utils/backends/tf_object_detection.py

@@ -18,7 +18,7 @@ class TFObjectDetector(BaseTFBackend):
     def __init__(self, model_bytes, metadata_bytes,
                  confidence_thresh=0.05,
                  device: str = None,
-                 session_config: tf.ConfigProto = None):
+                 session_config: tf.compat.v1.ConfigProto = None):
         """
         :param model_bytes: Model file data, likely a loaded *.pb file
         :param metadata_bytes: The dataset metadata file data, likely named