提交 aa250718 编写于 作者: R ranqiu

add dot_prod_layer

上级 08bc08d6
/* 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(1, getSize()) << "Dimension mismatch";
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);
{
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
...@@ -1081,6 +1081,21 @@ TEST(Layer, InterpolationLayer) { ...@@ -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", 100, false, useGpu);
}
}
TEST(Layer, OuterProdLayer) { TEST(Layer, OuterProdLayer) {
TestConfig config; TestConfig config;
config.layerConfig.set_type("out_prod"); config.layerConfig.set_type("out_prod");
......
...@@ -3209,6 +3209,15 @@ class SubNestedSequenceLayer(LayerBase): ...@@ -3209,6 +3209,15 @@ class SubNestedSequenceLayer(LayerBase):
self.set_layer_size(size) 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')
self.set_layer_size(1)
@config_layer('out_prod') @config_layer('out_prod')
class OuterProdLayer(LayerBase): class OuterProdLayer(LayerBase):
def __init__(self, name, inputs, device=None): def __init__(self, name, inputs, device=None):
......
...@@ -115,6 +115,7 @@ __all__ = [ ...@@ -115,6 +115,7 @@ __all__ = [
'huber_classification_cost', 'huber_classification_cost',
'block_expand_layer', 'block_expand_layer',
'maxout_layer', 'maxout_layer',
'dot_prod_layer',
'out_prod_layer', 'out_prod_layer',
'printer_layer', 'printer_layer',
'print_layer', 'print_layer',
...@@ -197,6 +198,7 @@ class LayerType(object): ...@@ -197,6 +198,7 @@ class LayerType(object):
SCALING_LAYER = 'scaling' SCALING_LAYER = 'scaling'
TRANS_LAYER = 'trans' TRANS_LAYER = 'trans'
ROTATE_LAYER = 'rotate' ROTATE_LAYER = 'rotate'
DOT_PROD_LAYER = 'dot_prod'
OUT_PROD_LAYER = 'out_prod' OUT_PROD_LAYER = 'out_prod'
FEATURE_MAP_EXPAND_LAYER = 'featmap_expand' FEATURE_MAP_EXPAND_LAYER = 'featmap_expand'
...@@ -4140,6 +4142,45 @@ def maxid_layer(input, name=None, layer_attr=None): ...@@ -4140,6 +4142,45 @@ def maxid_layer(input, name=None, layer_attr=None):
size=l.config.size) 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() @wrap_name_default()
def out_prod_layer(input1, input2, name=None, layer_attr=None): def out_prod_layer(input1, input2, name=None, layer_attr=None):
""" """
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册