Modifications to support benchmarking script

extract_exif.py

@@ -48,8 +48,6 @@ def main(config):
 
     assert len(metadata) == len(df_parsed)
 
-
-
     if config.save_files:
 
         EXIF_REPORT_PATH = join(config.repr.directory, 'exif_metadata.csv')

extract_features.py

@@ -41,7 +41,6 @@
     default=False,is_flag=True)
 
 
-
 def main(config, list_of_files, frame_sampling, save_frames):
     config = resolve_config(config_path=config, frame_sampling=frame_sampling, save_frames=save_frames)
     reps = ReprStorage(os.path.join(config.repr.directory))

winnow/feature_extraction/__init__.py

@@ -1,4 +1,4 @@
 
 from .intermediate_cnn import *
 from .frame_to_video import *
-from .similarity_model import *
+from .similarity_model import *

winnow/feature_extraction/loading_utils.py

@@ -4,6 +4,7 @@
 import pickle as pk
 import matplotlib.pylab as plt
 from sklearn.metrics import precision_recall_curve
+from scipy.spatial.distance import cdist
 
 
 def load_dataset(dataset):
@@ -137,6 +138,26 @@ def plot_pr_curve(pr_curve, title):
     plt.show()
 
 
+def calculate_similarities(queries, features):
+    """
+      Function that generates video triplets from CC_WEB_VIDEO.
+
+      Args:
+        queries: indexes of the query videos
+        features: global features of the videos in CC_WEB_VIDEO
+      Returns:
+        similarities: the similarities of each query with the videos in the dataset
+    """
+
+    features = features[0]
+    similarities = dict()
+    dist = np.nan_to_num(cdist(features[queries], features, metric='euclidean'))
+    for i, v in enumerate(queries):
+        sim = np.round(1 - dist[i] / dist.max(), decimals=6)
+        similarities[i + 1] = [(s, sim[s]) for s in sim.argsort()[::-1] if not np.isnan(sim[s])]
+    return similarities
+
+
 def evaluate(ground_truth,
              similarities,
              positive_labels='ESLMV',

winnow/feature_extraction/similarity_model.py

@@ -23,6 +23,11 @@ def predict(self, file_feature_dict):
         # values in the same order
         keys, features = zip(*file_feature_dict.items())
         features = np.array([tensor[0] for tensor in features])
+        embeddings = self.predict_from_features(features)
+
+        return dict(zip(keys, embeddings))
+
+    def predict_from_features(self, features):
 
         # Create model
         if self.model is None:
@@ -36,4 +41,4 @@ def predict(self, file_feature_dict):
 
         embeddings = self.model.embeddings(features)
         embeddings = np.nan_to_num(embeddings)
-        return dict(zip(keys, embeddings))
+        return embeddings