Merge pull request google#69 from ducklingll/master

use tf rewrite pca operation

Author: samihaija

feature_extractor/feature_extractor.py

@@ -64,6 +64,13 @@ def __init__(self, model_dir=MODEL_DIR):
     if not os.path.exists(model_dir):
       os.makedirs(model_dir)
 
+    # Load PCA Matrix.
+    download_path = self._maybe_download(YT8M_PCA_MAT)
+    pca_mean = os.path.join(self._model_dir, 'mean.npy')
+    if not os.path.exists(pca_mean):
+      tarfile.open(download_path, 'r:gz').extractall(model_dir)
+    self._load_pca()
+
     # Load Inception Network
     download_path = self._maybe_download(INCEPTION_TF_GRAPH)
     inception_proto_file = os.path.join(
@@ -72,12 +79,7 @@ def __init__(self, model_dir=MODEL_DIR):
       tarfile.open(download_path, 'r:gz').extractall(model_dir)
     self._load_inception(inception_proto_file)
 
-    # Load PCA Matrix.
-    download_path = self._maybe_download(YT8M_PCA_MAT)
-    pca_mean = os.path.join(self._model_dir, 'mean.npy')
-    if not os.path.exists(pca_mean):
-      tarfile.open(download_path, 'r:gz').extractall(model_dir)
-    self._load_pca()
+
 
   def extract_rgb_frame_features(self, frame_rgb, apply_pca=True):
     """Applies the YouTube8M feature extraction over an RGB frame.
@@ -98,13 +100,8 @@ def extract_rgb_frame_features(self, frame_rgb, apply_pca=True):
     assert len(frame_rgb.shape) == 3
     assert frame_rgb.shape[2] == 3  # 3 channels (R, G, B)
     with self._inception_graph.as_default():
-      frame_features = self.session.run('pool_3/_reshape:0',
+      frame_features = self.session.run('pca_final_feature:0',
                                         feed_dict={'DecodeJpeg:0': frame_rgb})
-      frame_features = frame_features[0]  # Unbatch.
-
-    if apply_pca:
-      frame_features = self.apply_pca(frame_features)
-
     return frame_features
 
   def apply_pca(self, frame_features):
@@ -148,6 +145,13 @@ def _load_inception(self, proto_file):
     with self._inception_graph.as_default():
       _ = tf.import_graph_def(graph_def, name='')
       self.session = tf.Session()
+      Frame_Features = self.session.graph.get_tensor_by_name('pool_3/_reshape:0')
+      Pca_Mean = tf.constant(value=self.pca_mean, dtype=tf.float32)
+      Pca_Eigenvecs = tf.constant(value=self.pca_eigenvecs, dtype=tf.float32)
+      Pca_Eigenvals = tf.constant(value=self.pca_eigenvals, dtype=tf.float32)
+      Feats = Frame_Features[0] - Pca_Mean
+      Feats = tf.reshape(tf.matmul(tf.reshape(Feats, [1, 2048]), Pca_Eigenvecs), [1024, ])
+      tf.divide(Feats, tf.sqrt(Pca_Eigenvals + 1e-4), name='pca_final_feature')
 
   def _load_pca(self):
     self.pca_mean = numpy.load(