Make the image norm layer in appearance head configurable

deepcell/model_zoo/tracking.py

@@ -45,6 +45,8 @@
 
 from spektral.layers import GCSConv, GCNConv, GATConv
 
+from deepcell_tracking import TrackingConfig
+
 from deepcell.layers import ImageNormalization2D
 from deepcell.layers import Comparison, DeltaReshape, Unmerge, TemporalMerge
 
@@ -220,6 +222,7 @@ class GNNTrackingModel:
             ``<layer name>-kwarg:value-kwarg:value``
         appearance_shape (tuple): shape of each object's appearance tensor
         norm_layer (str): Must be one of {'layer', 'batch'}
+        appearance_norm (bool): Whether to apply an input normalization layer to the apperance head
     """
     def __init__(self,
                  max_cells=39,
@@ -230,14 +233,16 @@ def __init__(self,
                  n_layers=3,
                  graph_layer='gcs',
                  appearance_shape=(32, 32, 1),
-                 norm_layer='batch'):
+                 norm_layer='batch',
+                 appearance_norm=True):
 
         self.n_filters = n_filters
         self.encoder_dim = encoder_dim
         self.embedding_dim = embedding_dim
         self.n_layers = n_layers
         self.max_cells = max_cells
         self.track_length = track_length
+        self.appearance_norm = appearance_norm
 
         if len(appearance_shape) != 3:
             raise ValueError('appearanace_shape should be a '
@@ -309,8 +314,9 @@ def get_appearance_encoder(self):
         inputs = Input(shape=app_shape, name='encoder_app_input')
 
         x = inputs
-        x = TimeDistributed(ImageNormalization2D(norm_method='whole_image',
-                                                 name='imgnrm_ae'))(x)
+        if self.appearance_norm:
+            x = TimeDistributed(ImageNormalization2D(norm_method='whole_image',
+                                                    name='imgnrm_ae'))(x)
 
         for i in range(int(math.log(app_shape[1], 2))):
             x = Conv3D(self.n_filters,