Network: optional rescale image input domain

Past experiments have never shown this to improve results, but since
everyone else does it, optionally include it.

Do this in the network rather than the preprocessing for efficiency and
easy interoperability of trained networks.

diluvian/config.py

@@ -134,6 +134,8 @@ class NetworkConfig(BaseConfig):
         inputs and outputs. Data is assumed to be ZYX row-major, but old
         versions of diluvian used XYZ, so this is necessary to load old
         networks.
+    rescale_image : bool
+        If true, rescale the input image intensity from [0, 1) to [-1, 1).
     num_modules : int
         Number of convolution modules to use, each module consisting of a skip
         link in parallel with ``num_layers_per_module`` convolution layers.
@@ -176,6 +178,7 @@ class NetworkConfig(BaseConfig):
     def __init__(self, settings):
         self.factory = str(settings.get('factory'))
         self.transpose = bool(settings.get('transpose', False))
+        self.rescale_image = bool(settings.get('rescale_image', False))
         self.num_modules = int(settings.get('num_modules', 8))
         self.num_layers_per_module = int(settings.get('num_layers_per_module', 2))
         self.convolution_dim = np.array(settings.get('convolution_dim', [3, 3, 3]))

diluvian/network.py

@@ -16,6 +16,7 @@
         Cropping3D,
         Dropout,
         Input,
+        Lambda,
         Permute,
         )
 from keras.layers.merge import (
@@ -34,8 +35,12 @@ def make_flood_fill_network(input_fov_shape, output_fov_shape, network_config):
         raise ValueError('ResNet implementation only supports same padding.')
 
     image_input = Input(shape=tuple(input_fov_shape) + (1,), dtype='float32', name='image_input')
+    if network_config.rescale_image:
+        ffn = Lambda(lambda x: (x - 0.5) * 2.0)(image_input)
+    else:
+        ffn = image_input
     mask_input = Input(shape=tuple(input_fov_shape) + (1,), dtype='float32', name='mask_input')
-    ffn = concatenate([image_input, mask_input])
+    ffn = concatenate([ffn, mask_input])
 
     # Convolve and activate before beginning the skip connection modules,
     # as discussed in the Appendix of He et al 2016.
@@ -109,8 +114,12 @@ def make_flood_fill_unet(input_fov_shape, output_fov_shape, network_config):
     """Construct a U-net flood filling network.
     """
     image_input = Input(shape=tuple(input_fov_shape) + (1,), dtype='float32', name='image_input')
+    if network_config.rescale_image:
+        ffn = Lambda(lambda x: (x - 0.5) * 2.0)(image_input)
+    else:
+        ffn = image_input
     mask_input = Input(shape=tuple(input_fov_shape) + (1,), dtype='float32', name='mask_input')
-    ffn = concatenate([image_input, mask_input])
+    ffn = concatenate([ffn, mask_input])
 
     # Note that since the Keras 2 upgrade strangely models with depth > 3 are
     # rejected by TF.