Converted to TF2 using the upgrade script and tested on TF2 without manual intervention. Now we can migrate to idiomatic TF2 code.

Author: rigolepe

Dockerfile

@@ -1,4 +1,4 @@
-FROM tensorflow/tensorflow:1.15.0-gpu-py3-jupyter
+FROM tensorflow/tensorflow:2.0.0-gpu-py3-jupyter
 # see: https://www.tensorflow.org/install/docker
 # see: https://hub.docker.com/r/tensorflow/tensorflow/
 

benchmark.py

@@ -15,7 +15,7 @@
 
 def main():
 
-    with tf.Session() as sess:
+    with tf.compat.v1.Session() as sess:
         model_cfg, model_outputs = posenet.load_model(args.model, sess)
         output_stride = model_cfg['output_stride']
         num_images = args.num_images

docker_img_build.sh

@@ -1,3 +1,5 @@
 #!/usr/bin/env bash
 
+docker rmi -f posenet-python
+
 docker build -t posenet-python -f Dockerfile .

image_demo.py

@@ -18,7 +18,7 @@
 
 def main():
 
-    with tf.Session() as sess:
+    with tf.compat.v1.Session() as sess:
         model_cfg, model_outputs = posenet.load_model(args.model, sess)
         output_stride = model_cfg['output_stride']
 

posenet/converter/tfjs2python.py

@@ -26,6 +26,7 @@ def to_output_strided_layers(convolution_def, output_stride):
     variables (see the load_variables function).
     :return: An array containing an element for each layer with the detailed layer specs defined in each of them.
     """
+
     current_stride = 1
     rate = 1
     block_id = 0
@@ -70,6 +71,7 @@ def load_variables(chkpoint, base_dir=BASE_DIR):
     Note for refactoring: To make this function reusable for other networks, the weights downloader should be either
     1/ more generic, or 2/ extracted outside this function. Apart from this, this function is likely very reusable for other networks.
     """
+
     manifest_path = os.path.join(base_dir, chkpoint, "manifest.json")
     if not os.path.exists(manifest_path):
         print('Weights for checkpoint %s are not downloaded. Downloading to %s ...' % (chkpoint, base_dir))
@@ -102,6 +104,7 @@ def _read_imgfile(path, width, height):
     :param height: The requested image target height.
     :return: The resized image with normalized pixels as a 3D array (height, width, channels).
     """
+
     img = cv2.imread(path)
     # The cv2.resize shape definition is indeed (width, height), while the image shape from cv2.imread is (height, width, channels).
     img = cv2.resize(img, (width, height))
@@ -136,13 +139,13 @@ def _depthwise_weights(layer_name):
         return variables["MobilenetV1/" + layer_name + "/depthwise_weights"]['x']
 
     def _conv_to_output(mobile_net_output, output_layer_name):
-        w = tf.nn.conv2d(mobile_net_output, _weights(output_layer_name), [1, 1, 1, 1], padding='SAME')
+        w = tf.nn.conv2d(input=mobile_net_output, filters=_weights(output_layer_name), strides=[1, 1, 1, 1], padding='SAME')
         w = tf.nn.bias_add(w, _biases(output_layer_name), name=output_layer_name)
         return w
 
     def _conv(inputs, stride, block_id):
         return tf.nn.relu6(
-            tf.nn.conv2d(inputs, _weights("Conv2d_" + str(block_id)), stride, padding='SAME')
+            tf.nn.conv2d(input=inputs, filters=_weights("Conv2d_" + str(block_id)), strides=stride, padding='SAME')
             +
             _biases("Conv2d_" + str(block_id))
         )
@@ -159,19 +162,19 @@ def _separable_conv(inputs, stride, block_id, dilations):
         # A depthwise convolution uses a filter (kernel) with a depth of 1 instead of the channel depth to get fewer variables that
         # have to be learned, and so achieve a faster but less accurate network. When the rate (or dilation) is 1, then the strides
         # must all be 1, see: https://www.tensorflow.org/versions/r1.15/api_docs/python/tf/nn/depthwise_conv2d
-        w = tf.nn.depthwise_conv2d(inputs, _depthwise_weights(dw_layer), stride, 'SAME', rate=dilations, data_format='NHWC')
+        w = tf.nn.depthwise_conv2d(input=inputs, filter=_depthwise_weights(dw_layer), strides=stride, padding='SAME', dilations=dilations, data_format='NHWC')
         w = tf.nn.bias_add(w, _biases(dw_layer))
         w = tf.nn.relu6(w)
 
-        w = tf.nn.conv2d(w, _weights(pw_layer), [1, 1, 1, 1], padding='SAME')
+        w = tf.nn.conv2d(input=w, filters=_weights(pw_layer), strides=[1, 1, 1, 1], padding='SAME')
         w = tf.nn.bias_add(w, _biases(pw_layer))
         w = tf.nn.relu6(w)
 
         return w
 
     x = image
     buff = [] # remove this buffer, seems like it's not used
-    with tf.variable_scope(None, 'MobilenetV1'):
+    with tf.compat.v1.variable_scope(None, 'MobilenetV1'):
 
         for m in layers:
             stride = [1, m['stride'], m['stride'], 1]
@@ -196,6 +199,16 @@ def _separable_conv(inputs, stride, block_id, dilations):
 
 
 def convert(model_id, model_dir, check=False):
+    """
+    Download and read the weight and bias variables for MobileNetV1, create the network and instantiate it with those variables.
+    Then write the instantiated network to a model file and corresponding checkpoint files.
+
+    :param model_id: Refers to the model to load, as defined in the config.yaml file.
+    :param model_dir: Defines where the model and checkpoint files will be saved.
+    :param check: Indicates whether or not to verify the model by feeding it a sample image.
+    :return: Nothing, the model and checkpoint files are written to the filesystem.
+    """
+
     cfg = load_config()
     checkpoints = cfg['checkpoints']
     image_size = cfg['imageSize']
@@ -221,12 +234,12 @@ def convert(model_id, model_dir, check=False):
         layers = to_output_strided_layers(mobile_net_arch, output_stride)
         variables = load_variables(chkpoint)
 
-        init = tf.global_variables_initializer()
-        with tf.Session() as sess:
+        init = tf.compat.v1.global_variables_initializer()
+        with tf.compat.v1.Session() as sess:
             sess.run(init)
-            saver = tf.train.Saver()
+            saver = tf.compat.v1.train.Saver()
 
-            image_ph = tf.placeholder(tf.float32, shape=[1, None, None, 3], name='image')
+            image_ph = tf.compat.v1.placeholder(tf.float32, shape=[1, None, None, 3], name='image')
             outputs = build_network(image_ph, layers, variables)
 
             sess.run(
@@ -241,7 +254,7 @@ def convert(model_id, model_dir, check=False):
                 os.makedirs(os.path.dirname(save_path))
             checkpoint_path = saver.save(sess, save_path, write_state=False)
 
-            tf.train.write_graph(cg, model_dir, "model-%s.pbtxt" % chkpoint)
+            tf.io.write_graph(cg, model_dir, "model-%s.pbtxt" % chkpoint)
 
             # Freeze graph and write our final model file
             freeze_graph(

posenet/model.py

@@ -42,8 +42,8 @@ def load_model(model_id, sess, model_dir=MODEL_DIR):
         convert(model_ord, model_dir, check=False)
         assert os.path.exists(model_path)
 
-    with tf.gfile.GFile(model_path, 'rb') as f:
-        graph_def = tf.GraphDef()
+    with tf.io.gfile.GFile(model_path, 'rb') as f:
+        graph_def = tf.compat.v1.GraphDef()
     graph_def.ParseFromString(f.read())
     sess.graph.as_default()
     tf.import_graph_def(graph_def, name='')

upgrade-tf-v2.sh

@@ -0,0 +1,8 @@
+#!/usr/bin/env bash
+
+WORK=$(dirname $(pwd))
+
+docker run --gpus all -it -v $WORK:/work posenet-python tf_upgrade_v2 \
+  --intree posenet-python/ \
+  --outtree posenet-python_v2/ \
+  --reportfile posenet-python/report.txt

webcam_demo.py

@@ -16,7 +16,7 @@
 
 
 def main():
-    with tf.Session() as sess:
+    with tf.compat.v1.Session() as sess:
         model_cfg, model_outputs = posenet.load_model(args.model, sess)
         output_stride = model_cfg['output_stride']