CHG: Add L2 normalization layer

Author: cecabert

kaffe/layers.py

@@ -35,6 +35,7 @@
     'MemoryData': shape_mem_data,
     'MultinomialLogisticLoss': shape_scalar,
     'MVN': shape_not_implemented,
+    'Normalize': shape_identity,
     'Pooling': shape_pool,
     'Power': shape_identity,
     'ReLU': shape_identity,

kaffe/tensorflow/network.py

@@ -253,3 +253,13 @@ def batch_normalization(self, input, name, scale_offset=True, relu=False):
     def dropout(self, input, keep_prob, name):
         keep = 1 - self.use_dropout + (self.use_dropout * keep_prob)
         return tf.nn.dropout(input, keep, name=name)
+        
+    @layer
+    def l2_normalize(self, input):
+        # NOTE: Currently, only inference is supported
+        with tf.variable_scope(name) as scope:
+          shp = input.get_shape().as_list()
+          outputs = tf.nn.l2_normalize(x=input, axis=-1)
+          alpha = self.make_var('alpha', shape=[-1:])
+          outputs = tf.multiply(outputs, alpha)
+          return outputs

kaffe/tensorflow/transformer.py

@@ -160,6 +160,9 @@ def map_eltwise(self, node):
             return TensorFlowNode(operations[op_code])
         except KeyError:
             raise KaffeError('Unknown elementwise operation: {}'.format(op_code))
+            
+    def map_normalize(self, node):
+        return TensorFlowNode('l2_normalize')
 
     def commit(self, chains):
         return chains

kaffe/transformers.py

@@ -282,7 +282,7 @@ def __call__(self, graph):
                 names = ('mean', 'variance')
                 if len(node.data) == 4:
                     names += ('scale', 'offset')
-            elif node.kind == NodeKind.PReLU:
+            elif node.kind == NodeKind.PReLU or node.kind == NodeKind.Normalize:
                 names = ('alpha',)
             else:
                 print_stderr('WARNING: Unhandled parameters: {}'.format(node.kind))