m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge pull request #2776 from NHZlX/mobilenet_gpu

Browse Source 
Mobilenet gpu implementation
cblas_new
Zhaolong Xing 8 years ago committed by GitHub
parent e2880f16c8 6c528cbc2a
commit 91d2a57a2e
7 changed files with 993 additions and 22 deletions
Show Stats Download Patch File Download Diff File

134
paddle/function/ConvOpTest.cpp
Unescape View File

@ -31,13 +31,22 @@ public:
ConvolutionTest(const std::string& conv1,
const std::string& conv2,
TestType type,
bool useGroups = true,
std::string algo = "auto") {
for (size_t batchSize : {1, 32}) {
for (size_t inputSize : {7, 14, 54}) {
for (size_t filterSize : {1, 3, 5}) {
for (size_t inputChannels : {3, 64}) {
for (size_t outputChannels : {3, 64, 128}) {
if (inputChannels < outputChannels) break;
for (size_t outputChannels : {3, 64}) {
if (inputChannels > outputChannels) break;
size_t groups;
if (!useGroups) {
groups = 1;
} else {
if (outputChannels % inputChannels != 0) continue;
groups = inputChannels;
}
for (size_t stride : {1, 2}) {
for (size_t padding : {0, 1}) {
if (padding >= filterSize) break;
@ -62,13 +71,24 @@ public:
FuncConfig()
.set("paddings", paddings)
.set("strides", strides)
.set("groups", (size_t)1)
.set("groups", groups)
.set("algo", algo));
TensorShape input{
batchSize, inputChannels, inputSize, inputSize};
TensorShape filter{
outputChannels, inputChannels, filterSize, filterSize};
TensorShape filter;
if (groups > 1)
filter = TensorShape({groups,
outputChannels / groups,
inputChannels / groups,
filterSize,
filterSize});
else
filter = TensorShape({outputChannels,
inputChannels,
filterSize,
filterSize});
TensorShape output{
batchSize, outputChannels, outputSize, outputSize};
@ -85,7 +105,8 @@ public:
} else if (type == kBackwardFilterTest) {
test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output));
test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input));
test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, filter));
test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, filter),
ADD_TO);
test.run();
}
}
@ -106,6 +127,7 @@ public:
ConvolutionTest2(const std::string& conv1,
const std::string& conv2,
TestType type,
bool useGroups = true,
std::string algo = "auto") {
for (size_t batchSize : {16}) {
for (size_t inputHeight : {7, 31}) {
@ -113,7 +135,15 @@ public:
for (size_t filterHeight : {1, 5}) {
for (size_t filterWidth : {3, 7}) {
for (size_t inputChannels : {7}) {
for (size_t outputChannels : {32}) {
for (size_t outputChannels : {7}) {
size_t groups;
if (!useGroups) {
groups = 1;
} else {
if (outputChannels % inputChannels != 0) continue;
groups = inputChannels;
}
size_t stride = 1;
size_t padding = 0;
size_t outputHeight =
@ -141,13 +171,24 @@ public:
FuncConfig()
.set("paddings", paddings)
.set("strides", strides)
.set("groups", (size_t)1)
.set("groups", groups)
.set("algo", algo));
TensorShape input{
batchSize, inputChannels, inputHeight, inputWidth};
TensorShape filter{
outputChannels, inputChannels, filterHeight, filterWidth};
TensorShape filter;
if (groups > 1)
filter = TensorShape({groups,
outputChannels / groups,
inputChannels / groups,
filterHeight,
filterWidth});
else
filter = TensorShape({outputChannels,
inputChannels,
filterHeight,
filterWidth});
TensorShape output{
batchSize, outputChannels, outputHeight, outputWidth};
@ -164,7 +205,8 @@ public:
} else if (type == kBackwardFilterTest) {
test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output));
test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input));
test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, filter));
test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, filter),
ADD_TO);
test.run();
}
}
@ -177,34 +219,88 @@ public:
}
};
// ======Start Convolution TEST======
TEST(Forward, GEMM) {
ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU> test(
"NaiveConv-CPU", "GemmConv-CPU", kForwardTest);
"NaiveConv-CPU", "GemmConv-CPU", kForwardTest, false);
ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU> test2(
"NaiveConv-CPU", "GemmConv-CPU", kForwardTest);
"NaiveConv-CPU", "GemmConv-CPU", kForwardTest, false);
}
#ifndef PADDLE_ONLY_CPU
TEST(Forward, GEMM2) {
ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
"GemmConv-CPU", "GemmConv-GPU", kForwardTest);
"GemmConv-CPU", "GemmConv-GPU", kForwardTest, false);
ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
"GemmConv-CPU", "GemmConv-GPU", kForwardTest);
"GemmConv-CPU", "GemmConv-GPU", kForwardTest, false);
}
TEST(BackwardInput, GEMM) {
ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
"GemmConvGradInput-CPU", "GemmConvGradInput-GPU", kBackwardInputTest);
"GemmConvGradInput-CPU",
"GemmConvGradInput-GPU",
kBackwardInputTest,
false);
ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
"GemmConvGradInput-CPU", "GemmConvGradInput-GPU", kBackwardInputTest);
"GemmConvGradInput-CPU",
"GemmConvGradInput-GPU",
kBackwardInputTest,
false);
}
TEST(BackwardFilter, GEMM) {
ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
"GemmConvGradFilter-CPU", "GemmConvGradFilter-GPU", kBackwardFilterTest);
"GemmConvGradFilter-CPU",
"GemmConvGradFilter-GPU",
kBackwardFilterTest,
false);
ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
"GemmConvGradFilter-CPU", "GemmConvGradFilter-GPU", kBackwardFilterTest);
"GemmConvGradFilter-CPU",
"GemmConvGradFilter-GPU",
kBackwardFilterTest,
false);
}
#endif
// ======End Convolution TEST======
// ======Start DepthwiseConvolution TEST======
// TODO(zhaolong) The depthwise convolution cpu test will be added when the cpu
// version of depthwiseConv is implemented.
#ifndef PADDLE_ONLY_CPU
TEST(DepthwiseConvForward, GEMM2) {
ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
"GemmConv-CPU", "DepthwiseConv-GPU", kForwardTest);
ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
"GemmConv-CPU", "DepthwiseConv-GPU", kForwardTest);
}
TEST(DepthwiseConvBackwardInput, GEMM) {
ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
"GemmConvGradInput-CPU",
"DepthwiseConvGradInput-GPU",
kBackwardInputTest);
ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
"GemmConvGradInput-CPU",
"DepthwiseConvGradInput-GPU",
kBackwardInputTest);
}
TEST(DepthwiseConvBackwardFilter, GEMM) {
ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
"GemmConvGradFilter-CPU",
"DepthwiseConvGradFilter-GPU",
kBackwardFilterTest);
ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
"GemmConvGradFilter-CPU",
"DepthwiseConvGradFilter-GPU",
kBackwardFilterTest);
}
#endif
// ======End DepthwiseConvolution TEST======
} // namespace paddle

306
paddle/function/DepthwiseConvOp.cpp
View File

File diff suppressed because it is too large Load Diff

159
paddle/function/DepthwiseConvOp.h
Unescape View File

@ -0,0 +1,159 @@
/* 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 "TensorType.h"
namespace paddle {
/**
*\brief Depthwise convolution forward. The outputData
* of depthwise convolution is same with ExpandConvLayer
* when groups equals inputChannels in ExpandConvLayer.
*
* \param[in] inputData input data.
* \param[in] filterData the Paramters of the depthwise conv layer..
* \param[in] batchSize batch size of input data.
* \param[in] outputChannels channels of outputData.
* \param[in] outputHeight height of outputData.
* \param[in] outputWidth width of outputData.
* \param[in] inputChannels channels of inputData.
* \param[in] inputHeight height of inputData.
* \param[in] inputWidth width of inputData..
* \param[in] filterMultiplier equals to outputChannels/groups_.
* \param[in] filterHeight height of filter.
* \param[in] filterWidth widht of filter.
* \param[in] strideH stride size in height direction.
* \param[in] strideW stride size in width direction.
* \param[in] paddingH padding size in height direction.
* \param[in] paddingW padding size in width direction.
* \param[out] outputData outputData.
*
*/
template <DeviceType Device, class T>
class DepthwiseConvFunctor {
public:
void operator()(const T* inputData,
const T* filterData,
int batchSize,
int outputChannels,
int outputHeight,
int outputWidth,
int inputChannels,
int inputHeight,
int inputWidth,
int filterMultiplier,
int filterHeight,
int filterWidth,
int strideH,
int strideW,
int paddingH,
int paddingW,
T* outputData);
};
/**
*\brief Functor tot compute the depthwise convolution backprop w.r.t input.
*
*
* \param[in] outputGradData the grad data of output.
* \param[in] filterData the Paramters of the depthwise conv layer..
* \param[in] batchSize batch size of input data.
* \param[in] outputChannels channels of outputData.
* \param[in] outputHeight height of outputData.
* \param[in] outputWidth width of outputData.
* \param[in] inputChannels channels of input data.
* \param[in] inputHeight height of inputData.
* \param[in] inputWidth width of inputData.
* \param[in] filterMultiplier equals to outputChannels/groups_.
* \param[in] filterHeight height of filter.
* \param[in] filterWidth widht of filter.
* \param[in] strideH stride size in height direction.
* \param[in] strideW stride size in width direction.
* \param[in] paddingH padding size in height direction.
* \param[in] paddingW padding size in width direction.
* \param[out] inputGrad the grad data of input.
*
*/
template <DeviceType Device, class T>
class DepthwiseConvGradInputFunctor {
public:
void operator()(const T* outputGrad,
const T* filterData,
int batchSize,
int outputChannels,
int outputHeight,
int outputWidth,
int inputChannels,
int inputHeight,
int inputWidth,
int filterMultiplier,
int filterHeight,
int filterWidth,
int strideH,
int strideW,
int paddingH,
int paddingW,
T* inputGrad);
};
/**
*\brief Functor tot compute the depthwise convolution backprop w.r.t filter.
*
* \param[in] outputGradData the grad data of output.
* \param[in] inputData inputData.
* \param[in] batchSize batch size of input data.
* \param[in] outputChannels channels of outputData.
* \param[in] outputHeight height of outputData.
* \param[in] outputWidth width of outputData.
* \param[in] inputChannels channels of input data.
* \param[in] inputHeight height of inputData.
* \param[in] inputWidth width of inputData.
* \param[in] filterMultiplier equals to outputChannels/groups_.
* \param[in] filterHeight height of filter.
* \param[in] filterWidth widht of filter.
* \param[in] strideH stride size in height direction.
* \param[in] strideW stride size in width direction.
* \param[in] paddingH padding size in height direction.
* \param[in] paddingW padding size in width direction.
* \param[in] colData Auxiliary data when calculating filterGrad.
* \param[in] multiplierData Auxiliary data when calculating filterGrad.
* \param[out] filterGrad the grad data of filter.
*
*/
template <DeviceType Device, class T>
class DepthwiseConvGradFilterFunctor {
public:
void operator()(const T* outputGrad,
const T* inputData,
int batchSize,
int outputChannels,
int outputHeight,
int outputWidth,
int inputChannels,
int inputHeight,
int inputWidth,
int filterMultiplier,
int filterHeight,
int filterWidth,
int strideH,
int strideW,
int paddingH,
int paddingW,
T* colData,
T* filterGrad);
};
} // namespace paddle

342
paddle/function/DepthwiseConvOpGpu.cu
View File

File diff suppressed because it is too large Load Diff

21
paddle/gserver/layers/ExpandConvLayer.cpp
Unescape View File

@ -38,10 +38,25 @@ bool ExpandConvLayer::init(const LayerMap &layerMap,
inputShape_.resize(numInputs);
filterShape_.resize(numInputs);
outputShape_.resize(numInputs);
std::string convType;
std::string convGradInputType;
std::string convGradFilterType;
for (int i = 0; i < config_.inputs_size(); i++) {
std::vector<size_t> paddings = {(size_t)paddingY_[i], (size_t)padding_[i]};
std::vector<size_t> strides = {(size_t)strideY_[i], (size_t)stride_[i]};
if (useGpu_ && (size_t)groups_[i] == (size_t)channels_[i] && !isDeconv_) {
convType = "DepthwiseConv";
convGradInputType = "DepthwiseConvGradInput";
convGradFilterType = "DepthwiseConvGradFilter";
} else {
convType = "GemmConv";
convGradInputType = "GemmConvGradInput";
convGradFilterType = "GemmConvGradFilter";
}
if (FLAGS_use_nnpack) {
CHECK_EQ(isDeconv_, false);
createFunction(forward_,
@ -53,21 +68,21 @@ bool ExpandConvLayer::init(const LayerMap &layerMap,
.set("algo", std::string("auto")));
} else {
createFunction(forward_,
!isDeconv_ ? "GemmConv" : "GemmConvGradInput",
!isDeconv_ ? convType : convGradInputType,
FuncConfig()
.set("paddings", paddings)
.set("strides", strides)
.set("groups", (size_t)groups_[i]));
createFunction(backward_,
!isDeconv_ ? "GemmConvGradInput" : "GemmConv",
!isDeconv_ ? convGradInputType : convType,
FuncConfig()
.set("paddings", paddings)
.set("strides", strides)
.set("groups", (size_t)groups_[i]));
createFunction(backward_,
"GemmConvGradFilter",
convGradFilterType,
FuncConfig()
.set("paddings", paddings)
.set("strides", strides)

49
paddle/gserver/tests/test_LayerGrad.cpp
Unescape View File

@ -347,6 +347,55 @@ TEST(Layer, CosSimVecMatLayer) {
}
}
void testDepthwiseConvLayer(const string& type, bool useGpu) {
TestConfig config;
config.biasSize = 32;
config.layerConfig.set_type(type);
config.layerConfig.set_num_filters(32);
config.layerConfig.set_partial_sum(1);
config.layerConfig.set_shared_biases(true);
config.inputDefs.push_back({INPUT_DATA, "layer_0", 2048, 192});
LayerInputConfig* input = config.layerConfig.add_inputs();
ConvConfig* conv = input->mutable_conv_conf();
conv->set_filter_size(2);
conv->set_filter_size_y(3);
conv->set_channels(16);
conv->set_padding(0);
conv->set_padding_y(1);
conv->set_stride(2);
conv->set_stride_y(2);
conv->set_groups(16);
conv->set_filter_channels(conv->channels() / conv->groups());
conv->set_img_size(16);
conv->set_img_size_y(8);
conv->set_output_x(outputSize(conv->img_size(),
conv->filter_size(),
conv->padding(),
conv->stride(),
/* caffeMode */ true));
conv->set_output_y(outputSize(conv->img_size_y(),
conv->filter_size_y(),
conv->padding_y(),
conv->stride_y(),
/* caffeMode */ true));
config.layerConfig.set_size(conv->output_x() * conv->output_y() *
config.layerConfig.num_filters());
testLayerGrad(config, "depthwise_conv", 100, false, useGpu);
// Use small batch_size and useWeight=true to test biasGrad
testLayerGrad(config, "depthwise_conv", 2, false, useGpu, true, 0.02);
}
TEST(Layer, depthwiseConvLayer) {
// 'depthwise_conv' is a sepecial case of 'exconv' whose
// groups size equals to the input channels size.
testDepthwiseConvLayer("exconv", /* useGpu= */ false);
#ifndef PADDLE_ONLY_CPU
testDepthwiseConvLayer("exconv", /* useGpu= */ true);
#endif
}
void testConvLayer(const string& type, bool trans, bool useGpu) {
TestConfig config;
config.biasSize = 16;

4
python/paddle/trainer/config_parser.py
Unescape View File

@ -3219,6 +3219,10 @@ def ParameterHook(type, **kwargs):
if sparsity_ratio is not None:
hook.sparsity_ratio = sparsity_ratio
return hook
elif type == 'dpruning':
hook = ParameterUpdaterHookConfig()
hook.type = type
return hook
else:
return None

Write Preview
Loading…
Cancel
Save