Add support for non-trainable neural network layers.

This is useful for multi-task setups where a network is trained on one task and
then a subset of the layers are reused for a different (but related) task.
Change: 119310581

tensorflow/contrib/layers/python/layers/layers.py

@@ -51,6 +51,7 @@ def fully_connected(x,
                     weight_collections=(ops.GraphKeys.WEIGHTS,),
                     bias_collections=(ops.GraphKeys.BIASES,),
                     output_collections=(ops.GraphKeys.ACTIVATIONS,),
+                    trainable=True,
                     weight_regularizer=None,
                     bias_regularizer=None):
   # pylint: disable=anomalous-backslash-in-string
@@ -107,6 +108,8 @@ def fully_connected(x,
     weight_collections: List of graph collections to which weights are added.
     bias_collections: List of graph collections to which biases are added.
     output_collections: List of graph collections to which outputs are added.
+    trainable: If `True` also add variables to the graph collection
+      `GraphKeys.TRAINABLE_VARIABLES` (see tf.Variable).
     weight_regularizer: A regularizer like the result of
       `l1_regularizer` or `l2_regularizer`. Used for weights.
     bias_regularizer: A regularizer like the result of
@@ -137,7 +140,8 @@ def fully_connected(x,
                                     dtype=dtype,
                                     initializer=weight_init,
                                     collections=weight_collections,
-                                    regularizer=weight_regularizer)
+                                    regularizer=weight_regularizer,
+                                    trainable=trainable)
     x_2_dim = x if len(dims) <= 2 else array_ops.reshape(x,
                                                          [-1, num_input_units])
     y = standard_ops.matmul(x_2_dim, w)
@@ -150,7 +154,8 @@ def fully_connected(x,
                                       dtype=dtype,
                                       initializer=bias_init,
                                       collections=bias_collections,
-                                      regularizer=bias_regularizer)
+                                      regularizer=bias_regularizer,
+                                      trainable=trainable)
 
       y = nn.bias_add(y, b)
 
@@ -179,6 +184,7 @@ def convolution2d(x,
                   weight_collections=None,
                   bias_collections=None,
                   output_collections=None,
+                  trainable=True,
                   weight_regularizer=None,
                   bias_regularizer=None):
   """Adds the parameters for a conv2d layer and returns the output.
@@ -227,6 +233,8 @@ def convolution2d(x,
     weight_collections: List of graph collections to which weights are added.
     bias_collections: List of graph collections to which biases are added.
     output_collections: List of graph collections to which outputs are added.
+    trainable: If `True` also add variables to the graph collection
+      `GraphKeys.TRAINABLE_VARIABLES` (see tf.Variable).
     weight_regularizer: A regularizer like the result of
       `l1_regularizer` or `l2_regularizer`. Used for weights.
     bias_regularizer: A regularizer like the result of
@@ -258,7 +266,8 @@ def convolution2d(x,
                                     dtype=dtype,
                                     initializer=weight_init,
                                     collections=weight_collections,
-                                    regularizer=weight_regularizer)
+                                    regularizer=weight_regularizer,
+                                    trainable=trainable)
 
     y = nn.conv2d(x, w, stride, padding)
 
@@ -270,7 +279,8 @@ def convolution2d(x,
                                       dtype=dtype,
                                       initializer=bias_init,
                                       collections=bias_collections,
-                                      regularizer=bias_regularizer)
+                                      regularizer=bias_regularizer,
+                                      trainable=trainable)
 
       y = nn.bias_add(y, b)