Merge pull request #5724 from ranqiu92/dot_product

add dot_prod_layer

doc/api/v2/config/layer.rst

@@ -335,6 +335,16 @@ bilinear_interp
 ..  autoclass:: paddle.v2.layer.bilinear_interp
     :noindex:
 
+dot_prod
+---------
+.. autoclass:: paddle.v2.layer.dot_prod
+    :noindex:
+
+out_prod
+--------
+.. autoclass:: paddle.v2.layer.out_prod
+    :noindex:
+
 power
 -----
 ..  autoclass:: paddle.v2.layer.power

paddle/gserver/layers/DotProdLayer.cpp

@@ -0,0 +1,97 @@
+/* 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 "Layer.h"
+#include "paddle/math/Matrix.h"
+#include "paddle/utils/Logging.h"
+#include "paddle/utils/Stat.h"
+
+namespace paddle {
+
+/**
+ * @brief A layer for computing the dot product of two vectors.
+ * Input1: vector (batchSize * dim)
+ * Input2: vector (batchSize * dim)
+ * Output: a matrix: (batchSize * 1)
+ */
+
+class DotProdLayer : public Layer {
+public:
+  explicit DotProdLayer(const LayerConfig& config) : Layer(config) {}
+
+  ~DotProdLayer() {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+};
+
+REGISTER_LAYER(dot_prod, DotProdLayer);
+
+bool DotProdLayer::init(const LayerMap& layerMap,
+                        const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+
+  CHECK_EQ(inputLayers_.size(), 2U);
+  CHECK_EQ(1UL, getSize())
+      << "The output dimensionality of this layer should be fixed to 1.";
+
+  return true;
+}
+
+void DotProdLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  MatrixPtr inV0 = getInputValue(0);
+  MatrixPtr inV1 = getInputValue(1);
+
+  size_t batchSize = inV0->getHeight();
+  CHECK_EQ(inV1->getHeight(), batchSize);
+  CHECK_EQ(inV0->getWidth(), inV1->getWidth());
+
+  {
+    REGISTER_TIMER_INFO("FwResetTimer", getName().c_str());
+    reserveOutput(batchSize, 1);
+  }
+
+  MatrixPtr outV = getOutputValue();
+  {
+    REGISTER_TIMER_INFO("FwDotProdTimer", getName().c_str());
+    outV->sumOfProducts(*inV0, *inV1, 1, 0);
+  }
+}
+
+void DotProdLayer::backward(const UpdateCallback& callback) {
+  MatrixPtr inV0 = getInputValue(0);
+  MatrixPtr inV1 = getInputValue(1);
+  MatrixPtr outG = getOutputGrad();
+  MatrixPtr inG0 = getInputGrad(0);
+  MatrixPtr inG1 = getInputGrad(1);
+
+  {
+    REGISTER_TIMER_INFO("BwDotProdTimer", getName().c_str());
+
+    if (inG0) {
+      inG0->addRowScale(0, *inV1, *outG);
+    }
+
+    if (inG1) {
+      inG1->addRowScale(0, *inV0, *outG);
+    }
+  }
+}
+
+}  // namespace paddle

paddle/gserver/tests/test_LayerGrad.cpp

@@ -1081,6 +1081,21 @@ TEST(Layer, InterpolationLayer) {
   }
 }
 
+TEST(Layer, DotProdLayer) {
+  TestConfig config;
+  config.layerConfig.set_type("dot_prod");
+  config.layerConfig.set_size(1);
+
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0});
+  config.layerConfig.add_inputs();
+  config.inputDefs.push_back({INPUT_DATA, "layer_1", 10, 0});
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "dot_prod", 10, false, useGpu);
+  }
+}
+
 TEST(Layer, OuterProdLayer) {
   TestConfig config;
   config.layerConfig.set_type("out_prod");

python/paddle/trainer/config_parser.py

@@ -3209,6 +3209,18 @@ class SubNestedSequenceLayer(LayerBase):
         self.set_layer_size(size)
 
 
+@config_layer('dot_prod')
+class DotProdLayer(LayerBase):
+    def __init__(self, name, inputs, device=None):
+        super(DotProdLayer, self).__init__(
+            name, 'dot_prod', 0, inputs, device=device)
+        config_assert(len(inputs) == 2, 'DotProdLayer must have 2 inputs.')
+        config_assert(
+            self.get_input_layer(0).size == self.get_input_layer(1).size,
+            "Two inputs should have the same size.")
+        self.set_layer_size(1)
+
+
 @config_layer('out_prod')
 class OuterProdLayer(LayerBase):
     def __init__(self, name, inputs, device=None):

python/paddle/trainer_config_helpers/layers.py

@@ -115,6 +115,7 @@ __all__ = [
     'huber_classification_cost',
     'block_expand_layer',
     'maxout_layer',
+    'dot_prod_layer',
     'out_prod_layer',
     'printer_layer',
     'print_layer',
@@ -197,6 +198,7 @@ class LayerType(object):
     SCALING_LAYER = 'scaling'
     TRANS_LAYER = 'trans'
     ROTATE_LAYER = 'rotate'
+    DOT_PROD_LAYER = 'dot_prod'
     OUT_PROD_LAYER = 'out_prod'
     FEATURE_MAP_EXPAND_LAYER = 'featmap_expand'
 
@@ -4140,6 +4142,45 @@ def maxid_layer(input, name=None, layer_attr=None):
         size=l.config.size)
 
 
+@wrap_name_default()
+def dot_prod_layer(input1, input2, name=None, layer_attr=None):
+    """
+    A layer for computing the dot product of two vectors.
+
+    The example usage is:
+
+    .. code-block:: python
+
+        dot_prod = dot_prod_layer(input1=vec1, input2=vec2)
+
+    :param name: The name of this layer. It is optional.
+    :type name: basestring
+    :param input1: The first input layer.
+    :type input: LayerOutput
+    :param input2: The second input layer.
+    :type input2: LayerOutput
+    :param layer_attr: The extra layer attribute. See ExtraLayerAttribute for
+                       details.
+    :type layer_attr: ExtraLayerAttribute.
+    :return: LayerOutput object.
+    :rtype: LayerOutput
+    """
+    assert isinstance(input1, LayerOutput)
+    assert isinstance(input2, LayerOutput)
+    assert input1.size == input2.size, ("Two inputs should have the same size.")
+
+    l = Layer(
+        name=name,
+        type=LayerType.DOT_PROD_LAYER,
+        inputs=[input1.name, input2.name],
+        **ExtraLayerAttribute.to_kwargs(layer_attr))
+    return LayerOutput(
+        name=name,
+        layer_type=LayerType.DOT_PROD_LAYER,
+        parents=[input1, input2],
+        size=l.config.size)
+
+
 @wrap_name_default()
 def out_prod_layer(input1, input2, name=None, layer_attr=None):
     """

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_dot_prod_layer)
 
 export whole_configs=(test_split_datasource)

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

@@ -0,0 +1,38 @@
+type: "nn"
+layers {
+  name: "vector1"
+  type: "data"
+  size: 10
+  active_type: ""
+}
+layers {
+  name: "vector2"
+  type: "data"
+  size: 10
+  active_type: ""
+}
+layers {
+  name: "__dot_prod_layer_0__"
+  type: "dot_prod"
+  size: 1
+  active_type: ""
+  inputs {
+    input_layer_name: "vector1"
+  }
+  inputs {
+    input_layer_name: "vector2"
+  }
+}
+input_layer_names: "vector1"
+input_layer_names: "vector2"
+output_layer_names: "__dot_prod_layer_0__"
+sub_models {
+  name: "root"
+  layer_names: "vector1"
+  layer_names: "vector2"
+  layer_names: "__dot_prod_layer_0__"
+  input_layer_names: "vector1"
+  input_layer_names: "vector2"
+  output_layer_names: "__dot_prod_layer_0__"
+  is_recurrent_layer_group: false
+}

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

@@ -0,0 +1,7 @@
+from paddle.trainer_config_helpers import *
+
+vec1 = data_layer(name='vector1', size=10)
+vec2 = data_layer(name='vector2', size=10)
+dot_product = dot_prod_layer(input1=vec1, input2=vec2)
+
+outputs(dot_product)