Merge pull request #5736 from lcy-seso/l2_distance

Add the L2 distance layer.

doc/api/v2/config/layer.rst

@@ -382,6 +382,11 @@ 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

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "L2DistanceLayer.h"
+#include "paddle/utils/Logging.h"
+#include "paddle/utils/Stat.h"
+
+namespace paddle {
+
+REGISTER_LAYER(l2_distance, L2DistanceLayer);
+
+bool L2DistanceLayer::init(const LayerMap& layerMap,
+                           const ParameterMap& parameterMap) {
+  /* Initialize the basic parent class */
+  Layer::init(layerMap, parameterMap);
+
+  CHECK_EQ(inputLayers_.size(), 2UL) << "The L2DistanceLayer accepts two and "
+                                     << "only two inputs.";
+  CHECK_EQ(getSize(), 1UL) << "The output dimensionality of L2DistanceLayer "
+                           << "is fixed to be 1.";
+
+  return true;
+}
+
+void L2DistanceLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  const auto inV1 = getInputValue(0);
+  const auto inV2 = getInputValue(1);
+
+  CHECK(inV1 && inV2);
+  CHECK_EQ(inV1->getHeight(), inV2->getHeight())
+      << "The height of two inputs of this layer must be the same.";
+  CHECK_EQ(inV1->getWidth(), inV2->getWidth())
+      << "The width of two inputs of this layer must be the same.";
+
+  int batchSize = inV1->getHeight();
+  int output_dim = getSize();
+  {
+    REGISTER_TIMER_INFO("L2DistanceBpAtvTimer", getName().c_str());
+    reserveOutput(batchSize, output_dim);
+    auto outV = getOutputValue();
+    CHECK(outV) << "The output matrix should not be null.";
+
+    Matrix::resizeOrCreate(
+        inputSub_, inV1->getHeight(), inV1->getWidth(), false, useGpu_);
+
+    inputSub_->assign(*inV1);
+    inputSub_->sub(*inV2);
+    outV->sumOfProducts(*inputSub_, *inputSub_, 1, 0);
+    outV->sqrt2(*outV);
+  }
+}
+
+void L2DistanceLayer::backward(const UpdateCallback& callback) {
+  const auto outG = getOutputGrad();
+  const auto outV = getOutputValue();
+  CHECK(outG && outV);
+
+  auto inGrad1 = getInputGrad(0);
+  auto inGrad2 = getInputGrad(1);
+
+  {
+    REGISTER_TIMER_INFO("L2DistanceBpAtvTimer", getName().c_str());
+
+    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);
+    }
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/L2DistanceLayer.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "Layer.h"
+#include "paddle/math/Matrix.h"
+
+namespace paddle {
+
+/**
+ * @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]
+ *
+ * - Input1: A vector (batchSize * dataDim)
+ * - Input2: A vector (batchSize * dataDim)
+ * - Output: A vector (batchSize * 1)
+ *
+ * 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,
+            const ParameterMap& parameterMap) override;
+
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+
+private:
+  // Store the result of subtracting Input2 from Input1 in forward computation,
+  // which will be reused in backward computation.
+  MatrixPtr inputSub_;
+};
+
+}  // namespace paddle

paddle/gserver/tests/test_LayerGrad.cpp

@@ -583,6 +583,7 @@ TEST(Layer, maxoutLayer) {
     testLayerGrad(config, "maxout", 10, false, useGpu);
   }
 }
+
 void testFcLayer(string format, size_t nnz) {
   TestConfig config;
   config.biasSize = 1024;
@@ -2444,6 +2445,25 @@ TEST(Layer, ScaleSubRegionLayer) {
   }
 }
 
+TEST(Layer, L2DistanceLayer) {
+  TestConfig config;
+  config.layerConfig.set_type("l2_distance");
+  config.layerConfig.set_size(1);
+  config.biasSize = 0;
+
+  const size_t input_dim = 27;
+  const size_t batch_size = 11;
+
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", input_dim, 0});
+  config.inputDefs.push_back({INPUT_DATA, "layer_1", input_dim, 0});
+  config.layerConfig.add_inputs();
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "l2_distance", batch_size, false, useGpu);
+  }
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

python/paddle/trainer/config_parser.py

@@ -3342,6 +3342,20 @@ 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,
+            'The CosSimLayer expects two and only two inputs.')
+        config_assert(
+            self.get_input_layer(0).size == self.get_input_layer(1).size,
+            'The two inputs of CosSimLayer must have the same dimensionality.')
+        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):
@@ -3349,10 +3363,24 @@ class CosSimVecMatLayer(LayerBase):
             name, 'cos_vm', size, inputs=inputs, device=device)
         self.config.cos_scale = cos_scale
         config_assert(
-            len(self.inputs) == 2, 'CosSimVecMatLayer must have 2 inputs')
+            len(self.inputs) == 2, 'The CosSimVecMatLayer must have 2 inputs.')
         config_assert(
             size * self.get_input_layer(0).size == self.get_input_layer(1).size,
-            'Wrong input size for CosSimVecMatLayer')
+            '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')
@@ -3396,18 +3424,6 @@ 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',
@@ -168,6 +169,7 @@ 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'
@@ -2334,6 +2336,51 @@ 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 calculates and returns the Euclidean distance between two input
+    vectors x and y. 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(y, LayerOutput)
+    Layer(
+        name=name,
+        type=LayerType.L2_DISTANCE,
+        inputs=[x.name, y.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()
@@ -3873,7 +3920,7 @@ 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

@@ -10,7 +10,8 @@ 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_dot_prod_layer)
+test_conv3d_layer test_deconv3d_layer test_BatchNorm3D test_resize_layer
+test_conv3d_layer test_deconv3d_layer test_BatchNorm3D test_resize_layer
+test_scale_sub_region_layer test_dot_prod_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: "y"
+  }
+}
+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)))