Make FaceNet pickalable

In this MR FaceNet was made picklable. Two things were necessary.

1. Organize the initialization of the public variables of the facenet class
2. Add a semaphore to control two critical regions (call and getstate). This allows Dask to smoothly do multithreading.

bob/ip/tensorflow_extractor/FaceNet.py

@@ -9,6 +9,7 @@ from bob.io.image import to_matplotlib
 from bob.extension import rc
 import bob.extension.download
 import bob.io.base
+import multiprocessing
 
 logger = logging.getLogger(__name__)
 
@@ -45,7 +46,7 @@ def get_model_filenames(model_dir):
                 ckpt_file = step_str.groups()[0]
     return meta_file, ckpt_file
 
-
+_semaphore = multiprocessing.Semaphore()
 class FaceNet(object):
     """Wrapper for the free FaceNet variant:
     https://github.com/davidsandberg/facenet
@@ -85,9 +86,19 @@ class FaceNet(object):
         super(FaceNet, self).__init__()
         self.model_path = model_path
         self.image_size = image_size
+        self.layer_name = layer_name        
+        self._clean_unpicklables()
+
+
+    def _clean_unpicklables(self):
         self.session = None
         self.embeddings = None
-        self.layer_name = layer_name
+        self.graph = None
+        self.images_placeholder = None
+        self.embeddings = None
+        self.phase_train_placeholder = None
+        self.session = None        
+
 
     def _check_feature(self, img):
         img = numpy.ascontiguousarray(img)
@@ -142,17 +153,18 @@ class FaceNet(object):
         logger.info("Successfully loaded the model.")
 
     def __call__(self, img):
-        images = self._check_feature(img)
-        if self.session is None:
-            self.graph = tf.Graph()
-            self.session = tf.compat.v1.Session(graph=self.graph)
-        if self.embeddings is None:
-            self.load_model()
-        feed_dict = {
-            self.images_placeholder: images,
-            self.phase_train_placeholder: False,
-        }
-        features = self.session.run(self.embeddings, feed_dict=feed_dict)
+        with _semaphore:
+            images = self._check_feature(img)
+            if self.session is None:
+                self.graph = tf.Graph()
+                self.session = tf.compat.v1.Session(graph=self.graph)
+            if self.embeddings is None:
+                self.load_model()
+            feed_dict = {
+                self.images_placeholder: images,
+                self.phase_train_placeholder: False,
+            }
+            features = self.session.run(self.embeddings, feed_dict=feed_dict)
         return features.flatten()
 
     @staticmethod
@@ -176,3 +188,13 @@ class FaceNet(object):
             )
 
         return model_path
+
+    def __setstate__(self, d):
+        # Handling unpicklable objects
+        self.__dict__ = d
+
+    def __getstate__(self):
+        # Handling unpicklable objects
+        with _semaphore:
+            self._clean_unpicklables()
+        return self.__dict__