提交 5cb29a8f 编写于 作者: G Guo Sheng 提交者: GitHub

Merge pull request #3083 from guoshengCS/add-L2NormLayer

add RowL2NormLayer
......@@ -105,6 +105,11 @@ cross_channel_norm
.. autoclass:: paddle.v2.layer.cross_channel_norm
:noindex:
row_l2_norm
-----------
.. autoclass:: paddle.v2.layer.row_l2_norm
:noindex:
Recurrent Layers
================
......
/* 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"
namespace paddle {
/**
* A layer for L2 normalization in each row,
* \f[
* out[i] = \frac{in[i]}{\sqrt{\sum_{k=1}^N in[k]^{2}}}
* \f]
* where the size of \f$in\f$ is (batchSize x dataDim),
* and the size of \f$out\f$ is (batchSize x dataDim).
*/
class RowL2NormLayer : public Layer {
protected:
MatrixPtr inSquare_;
MatrixPtr l2NormReciprocal_;
MatrixPtr dotSum_;
public:
explicit RowL2NormLayer(const LayerConfig& config) : Layer(config) {}
bool init(const LayerMap& layerMap,
const ParameterMap& parameterMap) override;
void forward(PassType passType) override;
void backward(const UpdateCallback& callback = nullptr) override;
};
REGISTER_LAYER(row_l2_norm, RowL2NormLayer);
bool RowL2NormLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
Layer::init(layerMap, parameterMap);
CHECK_EQ(inputLayers_.size(), 1U);
return true;
}
void RowL2NormLayer::forward(PassType passType) {
Layer::forward(passType);
MatrixPtr inV = getInputValue(0);
/* malloc memory for the output_ if necessary */
size_t batchSize = inV->getHeight();
size_t dataDim = getSize();
CHECK_EQ(dataDim, inV->getWidth());
resetOutput(batchSize, dataDim);
MatrixPtr outV = getOutputValue();
Matrix::resizeOrCreate(inSquare_, batchSize, dataDim, false, useGpu_);
inV->square2(*inSquare_);
Matrix::resizeOrCreate(l2NormReciprocal_, batchSize, 1, false, useGpu_);
inSquare_->rowSum(*l2NormReciprocal_);
l2NormReciprocal_->sqrt2(*l2NormReciprocal_);
l2NormReciprocal_->scalarDiv(*l2NormReciprocal_, 1.0);
outV->rowScale(0, *inV, *l2NormReciprocal_);
}
void RowL2NormLayer::backward(const UpdateCallback& callback) {
MatrixPtr inV = getInputValue(0);
MatrixPtr inG = getInputGrad(0);
MatrixPtr outV = getOutputValue();
MatrixPtr outG = getOutputGrad();
size_t batchSize = inV->getHeight();
// inG[ij] += outG[ij] / l2NormReciprocal
// inG[ij] += -inV[ij] * l2NormReciprocal * l2NormReciprocal * DotMul(outG[i],
// inV[i])
if (inG) {
Matrix::resizeOrCreate(dotSum_, batchSize, 1, false, useGpu_);
dotSum_->zeroMem();
dotSum_->rowDotMul(0, *outG, *outV);
dotSum_->dotMul(*dotSum_, *l2NormReciprocal_);
dotSum_->dotMul(*dotSum_, *l2NormReciprocal_);
inSquare_->rowScale(0, *inV, *dotSum_);
inG->sub(*inSquare_);
inG->addRowScale(0, *outG, *l2NormReciprocal_);
}
}
} // namespace paddle
......@@ -1899,6 +1899,19 @@ TEST(Layer, CropLayer) {
}
}
TEST(Layer, RowL2NormLayer) {
const size_t batchSize = 128;
const size_t size = 512;
TestConfig config;
config.layerConfig.set_type("row_l2_norm");
config.layerConfig.set_size(size);
config.inputDefs.push_back({INPUT_DATA, "input", size, 0});
config.layerConfig.add_inputs();
for (auto useGpu : {false, true}) {
testLayerGrad(config, "row_l2_norm", batchSize, false, useGpu, false);
}
}
int main(int argc, char** argv) {
testing::InitGoogleTest(&argc, argv);
initMain(argc, argv);
......
......@@ -2754,6 +2754,16 @@ class SumToOneNormLayer(LayerBase):
self.set_layer_size(input_layer0.size)
@config_layer('row_l2_norm')
class RowL2NormLayer(LayerBase):
def __init__(self, name, inputs, **xargs):
super(RowL2NormLayer, self).__init__(
name, 'row_l2_norm', 0, inputs=inputs, **xargs)
config_assert(len(self.inputs) == 1, 'RowL2NormLayer must have 1 input')
input_layer = self.get_input_layer(0)
self.set_layer_size(input_layer.size)
@config_layer('cos_vm')
class CosSimVecMatLayer(LayerBase):
def __init__(self, name, size, inputs, cos_scale=1.0, device=None):
......
......@@ -76,6 +76,7 @@ __all__ = [
'trans_layer',
'rotate_layer',
'sum_to_one_norm_layer',
'row_l2_norm_layer',
'get_output_layer',
'LayerType',
'context_projection',
......@@ -160,6 +161,7 @@ class LayerType(object):
BATCH_NORM_LAYER = 'batch_norm'
NORM_LAYER = 'norm'
SUM_TO_ONE_NORM_LAYER = 'sum_to_one_norm'
ROW_L2_NORM_LAYER = 'row_l2_norm'
ADDTO_LAYER = 'addto'
CONCAT_LAYER = 'concat'
......@@ -2889,6 +2891,42 @@ def sum_to_one_norm_layer(input, name=None, layer_attr=None):
name, LayerType.SUM_TO_ONE_NORM_LAYER, parents=[input], size=input.size)
@wrap_name_default()
@layer_support()
def row_l2_norm_layer(input, name=None, layer_attr=None):
"""
A layer for L2-normalization in each row.
.. math::
out[i] = \frac{in[i]}{\sqrt{\sum_{k=1}^N in[k]^{2}}}
where the size of :math:`in` is (batchSize x dataDim) ,
and the size of :math:`out` is a (batchSize x dataDim) .
The example usage is:
.. code-block:: python
row_l2_norm_layer = row_l2_norm_layer(input=layer)
:param input: Input layer.
:type input: LayerOutput
:param name: Layer name.
:type name: basestring
:param layer_attr: extra layer attributes.
:type layer_attr: ExtraLayerAttribute.
:return: LayerOutput object.
:rtype: LayerOutput
"""
Layer(
name=name,
type=LayerType.ROW_L2_NORM_LAYER,
inputs=[input.name],
**ExtraAttr.to_kwargs(layer_attr))
return LayerOutput(
name, LayerType.ROW_L2_NORM_LAYER, parents=[input], size=input.size)
@wrap_name_default("addto")
@wrap_act_default(act=LinearActivation())
@wrap_bias_attr_default(has_bias=False)
......
......@@ -7,6 +7,6 @@ test_rnn_group shared_fc shared_lstm shared_gru test_cost_layers_with_weight
test_spp_layer test_bilinear_interp test_maxout test_bi_grumemory math_ops
test_seq_concat_reshape test_pad test_smooth_l1 test_multiplex_layer
test_prelu_layer test_row_conv test_detection_output_layer test_multibox_loss_layer
test_recursive_topology test_gated_unit_layer)
test_recursive_topology test_gated_unit_layer test_row_l2_norm_layer)
export whole_configs=(test_split_datasource)
type: "nn"
layers {
name: "input"
type: "data"
size: 300
active_type: ""
}
layers {
name: "__row_l2_norm_layer_0__"
type: "row_l2_norm"
size: 300
active_type: ""
inputs {
input_layer_name: "input"
}
}
input_layer_names: "input"
output_layer_names: "__row_l2_norm_layer_0__"
sub_models {
name: "root"
layer_names: "input"
layer_names: "__row_l2_norm_layer_0__"
input_layer_names: "input"
output_layer_names: "__row_l2_norm_layer_0__"
is_recurrent_layer_group: false
}
from paddle.trainer_config_helpers import *
data = data_layer(name='input', size=300)
row_l2_norm = row_l2_norm_layer(input=data)
outputs(row_l2_norm)
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