Added class SpatialSELayer

niftynet/layer/squeeze_excitation_layer.py

@@ -5,6 +5,7 @@
 
 from niftynet.layer.base_layer import Layer
 from niftynet.layer.fully_connected import FullyConnectedLayer
+from niftynet.layer.convolution import ConvolutionalLayer
 from niftynet.utilities.util_common import look_up_operations
 
 SUPPORTED_OP = set(['AVG', 'MAX'])
@@ -14,7 +15,7 @@ class ChannelSELayer(Layer):
     def __init__(self,
                  func='AVG',
                  reduction_ratio=16,
-                 name='squeeze_excitation'):
+                 name='channel_squeeze_excitation'):
         self.func = func.upper()
         self.reduction_ratio = reduction_ratio
         self.layer_name = '{}_{}'.format(self.func.lower(), name)
@@ -23,7 +24,7 @@ def __init__(self,
         look_up_operations(self.func, SUPPORTED_OP)
 
     def layer_op(self, input_tensor):
-        # squeeze: global information embedding
+        # spatial squeeze
         input_rank = len(input_tensor.shape)
         reduce_indices = list(range(input_rank))[1:-1]
         if self.func == 'AVG':
@@ -33,7 +34,7 @@ def layer_op(self, input_tensor):
         else:
             raise NotImplementedError("pooling function not supported")
 
-        # excitation: adaptive recalibration
+        # channel excitation
         num_channels = int(squeeze_tensor.shape[-1])
         reduction_ratio = self.reduction_ratio
         if num_channels % reduction_ratio != 0:
@@ -62,3 +63,24 @@ def layer_op(self, input_tensor):
         output_tensor = tf.multiply(input_tensor, fc_out_2)
 
         return output_tensor
+
+class SpatialSELayer(Layer):
+    def __init__(self,
+                 name='spatial_squeeze_excitation'):
+        self.layer_name = '{}_{}'.format(self.func.lower(), name)
+        super(SpatialSELayer, self).__init__(name=self.layer_name)
+
+    def layer_op(self, input_tensor):
+        # channel squeeze
+        conv = ConvolutionalLayer(n_output_chns=1,
+                                  kernel_size=1,
+                                  with_bn=False,
+                                  acti_func='sigmoid',
+                                  name="se_conv")
+
+        squeeze_tensor = conv(input_tensor)
+
+        # spatial excitation
+        output_tensor = tf.multiply(input_tensor, squeeze_tensor)
+
+        return output_tensor