Fix ElasticTransform bug, manager repeat

src/data/augment_layers.py

@@ -67,7 +67,6 @@ def call(self, inputs, training=None):
         )
 
         # Apply the shear transformation
-        # Fill value is set to 0 to ensure that binarization is not affected
         sheared_image = tf.raw_ops.ImageProjectiveTransformV3(
             images=inputs,
             transforms=shear_matrix_tf,
@@ -136,6 +135,10 @@ def call(self, inputs, training=None):
         x_deformed = tf.clip_by_value(x_deformed,
                                       clip_value_min=0.0,
                                       clip_value_max=1.0)
+
+        # HACK: Reshape to original shape to force a shape
+        x_deformed = tf.reshape(x_deformed, tf.shape(inputs))
+
         return x_deformed
 
 

src/data/manager.py

@@ -159,7 +159,7 @@ def _fill_datasets_dict(self,
                 partition_list = self.config[f"{partition}_list"]
             if partition_list:
                 # Create dataset for the current partition
-                datasets[partition] = self.create_dataset(
+                datasets[partition] = self._create_dataset(
                     files=partitions[partition],
                     labels=labels[partition],
                     sample_weights=sample_weights[partition],
@@ -410,11 +410,11 @@ def get_train_batches(self):
         return int(np.ceil(len(self.raw_data['train'])
                            / self.config['batch_size']))
 
-    def create_dataset(self,
-                       files: List[str],
-                       labels: List[str],
-                       sample_weights: List[str],
-                       partition_name: str) -> tf.data.Dataset:
+    def _create_dataset(self,
+                        files: List[str],
+                        labels: List[str],
+                        sample_weights: List[str],
+                        partition_name: str) -> tf.data.Dataset:
         """
         Create a dataset for a specific partition.
 
@@ -453,8 +453,7 @@ def create_dataset(self,
         dataset = tf.data.Dataset.from_tensor_slices(data)
         if is_training:
             # Add additional repeat and shuffle for training
-            dataset = dataset.repeat(self.config["aug_multiply"])\
-                .shuffle(len(files))
+            dataset = dataset.repeat().shuffle(len(files))
 
         dataset = (dataset
                    .map(data_loader.load_images,