add config_helper.

doc/api/v2/config/layer.rst

@@ -372,6 +372,11 @@ cos_sim
 ..  autoclass:: paddle.v2.layer.cos_sim
     :noindex:
 
+l2_distance
+-----------
+..  autoclass:: paddle.v2.layer.l2_distance
+    :noindex:
+
 trans
 -----
 ..  autoclass:: paddle.v2.layer.trans

paddle/gserver/layers/L2DistanceLayer.cpp

@@ -25,9 +25,9 @@ bool L2DistanceLayer::init(const LayerMap& layerMap,
   /* Initialize the basic parent class */
   Layer::init(layerMap, parameterMap);
 
-  CHECK_EQ(inputLayers_.size(), 2UL) << "The L2 distance layer accepts two and "
+  CHECK_EQ(inputLayers_.size(), 2UL) << "The L2DistanceLayer accepts two and "
                                      << "only two inputs.";
-  CHECK_EQ(getSize(), 1UL) << "The output dimensionality of L2 distance"
+  CHECK_EQ(getSize(), 1UL) << "The output dimensionality of L2DistanceLayer "
                            << "is fixed to be 1.";
 
   return true;
@@ -41,9 +41,9 @@ void L2DistanceLayer::forward(PassType passType) {
 
   CHECK(inV1 && inV2);
   CHECK_EQ(inV1->getHeight(), inV2->getHeight())
-      << "The height of two inputs to this layer must be the same.";
+      << "The height of two inputs of this layer must be the same.";
   CHECK_EQ(inV1->getWidth(), inV2->getWidth())
-      << "The width of two inputs to this layer must be the same.";
+      << "The width of two inputs of this layer must be the same.";
 
   int batchSize = inV1->getHeight();
   int output_dim = getSize();
@@ -66,22 +66,21 @@ void L2DistanceLayer::forward(PassType passType) {
 void L2DistanceLayer::backward(const UpdateCallback& callback) {
   const auto outG = getOutputGrad();
   const auto outV = getOutputValue();
-  const auto inV1 = getInputValue(0);
-  const auto inV2 = getInputValue(1);
+  CHECK(outG && outV);
+
   auto inGrad1 = getInputGrad(0);
   auto inGrad2 = getInputGrad(1);
-  CHECK(outG && outV && inV1 && inV2 && inGrad1 && inGrad2);
 
   {
     REGISTER_TIMER_INFO("L2DistanceBpAtvTimer", getName().c_str());
 
-    outV->scalarDiv(*outV, 1.);
-    outV->dotMul(*outG, *outV);
-
-    if (inGrad1) {
-      inGrad1->addRowScale(0, *inputSub_, *outV);
+    if (inGrad1 || inGrad2) {
+      outV->scalarDiv(*outV, 1.);
+      outV->dotMul(*outG, *outV);
     }
 
+    if (inGrad1) inGrad1->addRowScale(0, *inputSub_, *outV);
+
     if (inGrad2) {
       inputSub_->mulScalar(-1.);
       inGrad2->addRowScale(0, *inputSub_, *outV);

paddle/gserver/layers/L2DistanceLayer.h

@@ -16,12 +16,11 @@ limitations under the License. */
 
 #include "Layer.h"
 #include "paddle/math/Matrix.h"
-#include "paddle/utils/ThreadLocal.h"
 
 namespace paddle {
 
 /**
- * @brief A layer for calculating l2 distance between the two input vectors.
+ * @brief The layer calculates the l2 distance between two input vectors.
  * \f[
  * f(\bf{x}, \bf{y}) = \sqrt{\sum_{i=1}^D(x_i - y_i)}
  * \f]
@@ -30,13 +29,12 @@ namespace paddle {
  * - Input2: A vector (batchSize * dataDim)
  * - Output: A vector (batchSize * 1)
  *
- * The config file api is l2_distance.
+ * The configuration api is: l2_distance_layer.
  */
 
 class L2DistanceLayer : public Layer {
 public:
   explicit L2DistanceLayer(const LayerConfig& config) : Layer(config) {}
-
   ~L2DistanceLayer() {}
 
   bool init(const LayerMap& layerMap,
@@ -46,7 +44,8 @@ public:
   void backward(const UpdateCallback& callback = nullptr) override;
 
 private:
-  // Store result of subtracting Input2 from Input1.
+  // Store the result of subtracting Input2 from Input1 in forward computation,
+  // which will be reused in backward computation.
   MatrixPtr inputSub_;
 };
 

python/paddle/trainer/config_parser.py

@@ -3330,6 +3330,18 @@ class RowL2NormLayer(LayerBase):
         self.set_layer_size(input_layer.size)
 
 
+@config_layer('cos')
+class CosSimLayer(LayerBase):
+    def __init__(self, name, inputs, cos_scale=1, device=None):
+        super(CosSimLayer, self).__init__(
+            name, 'cos', 1, inputs=inputs, device=device)
+        config_assert(len(self.inputs) == 2, 'CosSimLayer must have 2 inputs')
+        config_assert(
+            self.get_input_layer(0).size == self.get_input_layer(1).size,
+            'inputs of CosSimLayer must have same dim')
+        self.config.cos_scale = cos_scale
+
+
 @config_layer('cos_vm')
 class CosSimVecMatLayer(LayerBase):
     def __init__(self, name, size, inputs, cos_scale=1.0, device=None):
@@ -3343,6 +3355,20 @@ class CosSimVecMatLayer(LayerBase):
             'Wrong input size for CosSimVecMatLayer')
 
 
+@config_layer('l2_distance')
+class L2DistanceLayer(LayerBase):
+    def __init__(self, name, inputs, device=None):
+        super(L2DistanceLayer, self).__init__(
+            name, 'l2_distance', 1, inputs=inputs, device=device)
+        config_assert(
+            len(self.inputs) == 2, ('The L2DistanceLayer must have '
+                                    'and only have 2 inputs.'))
+        config_assert(
+            self.get_input_layer(0).size == self.get_input_layer(1).size,
+            ('Two inputs of the L2DistanceLayer must have '
+             'the same dimensionality.'))
+
+
 @config_layer('sampling_id')
 class SamplingIdLayer(LayerBase):
     def __init__(self, name, inputs, device=None):
@@ -3384,18 +3410,6 @@ class AverageLayer(LayerBase):
         self.create_bias_parameter(bias, self.config.size)
 
 
-@config_layer('cos')
-class CosSimLayer(LayerBase):
-    def __init__(self, name, inputs, cos_scale=1, device=None):
-        super(CosSimLayer, self).__init__(
-            name, 'cos', 1, inputs=inputs, device=device)
-        config_assert(len(self.inputs) == 2, 'CosSimLayer must have 2 inputs')
-        config_assert(
-            self.get_input_layer(0).size == self.get_input_layer(1).size,
-            'inputs of CosSimLayer must have same dim')
-        self.config.cos_scale = cos_scale
-
-
 @config_layer('tensor')
 class TensorLayer(LayerBase):
     def __init__(self, name, size, inputs, bias=True, **xargs):

python/paddle/trainer_config_helpers/layers.py

@@ -51,6 +51,7 @@ __all__ = [
     'last_seq',
     'first_seq',
     'cos_sim',
+    'l2_distance_layer',
     'hsigmoid',
     'conv_projection',
     'square_error_cost',
@@ -167,6 +168,7 @@ class LayerType(object):
     COST = 'cost'
     COSINE_SIM_VEC = 'cos_vm'
     COSINE_SIM = 'cos'
+    L2_DISTANCE = 'l2_distance'
     HSIGMOID = 'hsigmoid'
     CONV_LAYER = 'conv'
     CONVTRANS_LAYER = 'convt'
@@ -2332,6 +2334,51 @@ def cos_sim(a, b, scale=1, size=1, name=None, layer_attr=None):
     return LayerOutput(name, LayerType.COSINE_SIM, parents=[a, b], size=size)
 
 
+@wrap_name_default()
+@layer_support()
+def l2_distance_layer(x, y, name=None, layer_attr=None):
+    """
+    This layer calculate and return the Euclidean distance between two input
+    vectors a and b. The equation is as follows:
+
+    ..  math::
+        l2_distance(\\mathbf{x}, \\mathbf{y}) = \\sqrt{\\sum_{i=1}^D(x_i - y_i)}
+
+    The output size of this layer is fixed to be 1. Note that the above
+    computation is for one sample. Multiple samples are processed in one batch.
+
+    The example usage is:
+
+    .. code-block:: python
+
+       l2_sim = l2_distance(x=layer1, y=layer2)
+
+    :param name: The name of this layer. It is optional.
+    :type name: basestring
+    :param x: The first input x for this layer, whose output is a matrix with
+              dimensionality N x D. N is the sample number in a mini-batch.
+              D is the dimensionality of x's output.
+    :type x: LayerOutput
+    :param y: The second input y for this layer, whose output is a matrix with
+              dimensionality N x D. N is the sample number in a mini-batch.
+              D is the dimensionality of y's output.
+    :type y: LayerOutput
+    :param layer_attr: The extra layer attributes, for example, drop rate.
+                       See ExtraLayerAttribute for more details.
+    :type layer_attr: ExtraLayerAttribute
+    :return: The returned LayerOutput object.
+    :rtype: LayerOutput
+    """
+
+    assert isinstance(x, LayerOutput) and isinstance(x, LayerOutput)
+    Layer(
+        name=name,
+        type=LayerType.L2_DISTANCE,
+        inputs=[x.name, x.name],
+        **ExtraLayerAttribute.to_kwargs(layer_attr))
+    return LayerOutput(name, LayerType.L2_DISTANCE, parents=[x, y], size=1)
+
+
 @wrap_name_default()
 @wrap_bias_attr_default(has_bias=True)
 @wrap_param_attr_default()
@@ -3867,7 +3914,7 @@ def recurrent_layer(input,
     :type input: LayerOutput
     :param act: Activation type. TanhActivation is the default activation.
     :type act: BaseActivation
-    :param bias_attr: The parameter attribute for bias. If this parameter is set to 
+    :param bias_attr: The parameter attribute for bias. If this parameter is set to
                       False or an object whose type is not ParameterAttribute,
                       no bias is defined. If the parameter is set to True,
                       the bias is initialized to zero.

python/paddle/trainer_config_helpers/tests/configs/file_list.sh

@@ -10,6 +10,7 @@ test_prelu_layer test_row_conv test_detection_output_layer test_multibox_loss_la
 test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_layer
 test_kmax_seq_socre_layer test_sub_nested_seq_select_layer test_scale_shift_layer
 test_seq_slice_layer test_cross_entropy_over_beam test_roi_pool_layer test_pooling3D_layer
-test_conv3d_layer test_deconv3d_layer test_BatchNorm3D test_resize_layer test_scale_sub_region_layer)
+test_conv3d_layer test_deconv3d_layer test_BatchNorm3D test_resize_layer
+test_scale_sub_region_layer test_l2_distance_layer)
 
 export whole_configs=(test_split_datasource)

python/paddle/trainer_config_helpers/tests/configs/protostr/test_l2_distance_layer.protostr

+type: "nn"
+layers {
+  name: "x"
+  type: "data"
+  size: 128
+  active_type: ""
+}
+layers {
+  name: "y"
+  type: "data"
+  size: 128
+  active_type: ""
+}
+layers {
+  name: "__l2_distance_layer_0__"
+  type: "l2_distance"
+  size: 1
+  active_type: ""
+  inputs {
+    input_layer_name: "x"
+  }
+  inputs {
+    input_layer_name: "x"
+  }
+}
+input_layer_names: "x"
+input_layer_names: "y"
+output_layer_names: "__l2_distance_layer_0__"
+sub_models {
+  name: "root"
+  layer_names: "x"
+  layer_names: "y"
+  layer_names: "__l2_distance_layer_0__"
+  input_layer_names: "x"
+  input_layer_names: "y"
+  output_layer_names: "__l2_distance_layer_0__"
+  is_recurrent_layer_group: false
+}
+

python/paddle/trainer_config_helpers/tests/configs/test_l2_distance_layer.py

+from paddle.trainer_config_helpers import *
+
+outputs(
+    l2_distance_layer(
+        x=data_layer(
+            name='x', size=128), y=data_layer(
+                name='y', size=128)))