Added labels to data set

Padded them to make the tensors homogeneous. The zero tensors can be
removed later when evaluating. They are part of data set primarily
to ensure mapping between prediction result and corresponding labels.

Signed-off-by: Jim Martens <github@2martens.de>

src/twomartens/masterthesis/data.py

@@ -255,6 +255,7 @@ def load_scenenet_val(photo_paths: Sequence[Sequence[str]],
     
     annotation_file_train = f"{coco_path}/annotations/instances_train2014.json"
     cats_to_classes, _, _, _ = coco_utils.get_coco_category_maps(annotation_file_train)
+    max_nr_labels = -1
     
     for trajectory in trajectories:
         traj_image_paths, traj_instances = trajectory
@@ -269,16 +270,26 @@ def load_scenenet_val(photo_paths: Sequence[Sequence[str]],
                     bbox[2],
                     bbox[3]
                 ])
+            
+            len_labels = len(labels)
+            if len_labels > max_nr_labels:
+                max_nr_labels = len_labels
     
             final_image_paths.append(image_path)
             final_labels.append(labels)
         
+    empty_label = [0, 0, 0, 0, 0]
+    for labels in final_labels:
+        len_labels = len(labels)
+        if len_labels < max_nr_labels:
+            labels += empty_label * (max_nr_labels - len_labels)
+        
     length_dataset = len(final_image_paths)
     
     path_dataset = tf.data.Dataset.from_tensor_slices(final_image_paths)
-    # label_dataset = tf.data.Dataset.from_tensor_slices(final_labels)
+    label_dataset = tf.data.Dataset.from_tensor_slices(final_labels)
-    # dataset = tf.data.Dataset.zip((path_dataset, label_dataset))
+    dataset = tf.data.Dataset.zip((path_dataset, label_dataset))
-    dataset = path_dataset.apply(tf.data.experimental.shuffle_and_repeat(buffer_size=length_dataset, count=num_epochs))
+    dataset = dataset.repeat(num_epochs)
     dataset = dataset.batch(batch_size=batch_size)
     dataset = dataset.map(_load_images_ssd_callback(resized_shape))
     dataset = dataset.prefetch(1)

src/twomartens/masterthesis/main.py

@@ -196,7 +196,8 @@ def _ssd_val(args: argparse.Namespace) -> None:
         instances = pickle.load(file)
     
     scenenet_data, nr_digits = data.load_scenenet_val(file_names_photos, instances, args.coco_path,
-                                                      batch_size=batch_size, resized_shape=(image_size, image_size))
+                                                      batch_size=batch_size,
+                                                      resized_shape=(image_size, image_size))
     del file_names_photos, instances
     
     use_summary_writer = summary_ops_v2.create_file_writer(

src/twomartens/masterthesis/ssd.py

@@ -174,9 +174,11 @@ def _predict_one_epoch(dataset: tf.data.Dataset,
    
     # prepare filename
     filename = 'ssd_predictions'
+    label_filename = 'ssd_labels'
     if use_dropout:
         filename = f"dropout-{filename}"
     output_file = os.path.join(output_path, filename)
+    label_output_file = os.path.join(output_path, label_filename)
     
     # go through the data set
     counter = 0
@@ -184,7 +186,7 @@ def _predict_one_epoch(dataset: tf.data.Dataset,
     
     from tensorflow.python.eager import context
     
-    for inputs in dataset:
+    for inputs, labels in dataset:
         decoded_predictions_batch = []
         if use_dropout:
             for _ in range(forward_passes_per_image):
@@ -200,14 +202,17 @@ def _predict_one_epoch(dataset: tf.data.Dataset,
         if nr_digits is not None:
             counter_str = str(counter).zfill(nr_digits)
             filename = f"{output_file}-{counter_str}.npy"
+            label_filename = f"{label_output_file}-{counter_str}.npy"
         else:
             filename = f"{output_file}-{counter:d}.npy"
+            label_filename = f"{label_output_file}-{counter:d}.npy"
         
-        with open(filename, 'wb') as file:
+        with open(filename, 'wb') as file, open(label_filename, 'wb') as label_file:
             decoded_predictions_batch_np = np.array(decoded_predictions_batch)
             del decoded_predictions_batch
             np.save(file, decoded_predictions_batch_np, allow_pickle=False, fix_imports=False)
             del decoded_predictions_batch_np
+            np.save(label_file, labels, allow_pickle=False, fix_imports=False)
         
         counter += 1