Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

fix bug in ExpandLayer when seq expands to sub-seq #5044

Closed
wants to merge 5 commits into from
+179 −133
Closed

fix bug in ExpandLayer when seq expands to sub-seq #5044

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
8ea1863
fix bug in ExpandLayer when seq expands to sub-seq
luotao1 Oct 23, 2017
50556b8
Merge branch 'develop' into expand
luotao1 Oct 24, 2017
2f64a0d
simplify the gserver/tests/CMakelists.txt
luotao1 Oct 24, 2017
22881be
add unitest for ExpandLayer
luotao1 Oct 24, 2017
ca385e1
Merge branch 'develop' into expand
luotao1 Oct 25, 2017
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 3 additions & 7 deletions paddle/gserver/layers/ExpandLayer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,18 +52,15 @@ void ExpandLayer::forward(PassType passType) {
const Argument& shapeInput = getInput(1);
const Argument& dataInput = getInput(0);
size_t outputBatchSize = shapeInput.getBatchSize();
auto startPositions = type_ ? shapeInput.subSequenceStartPositions
: shapeInput.sequenceStartPositions;
auto startPositions = shapeInput.sequenceStartPositions;
size_t numSequences = startPositions->getSize() - 1;
const int* starts = startPositions->getData(false);

CHECK_EQ(starts[numSequences], shapeInput.getBatchSize());
CHECK_EQ(dataInput.getBatchSize(), shapeInput.getNumSequences());
if (type_) {
// when trans_type = seq, input[1] must hasSubseq
CHECK_EQ(shapeInput.hasSubseq(), 1UL);
CHECK_EQ(dataInput.getNumSequences(), shapeInput.getNumSequences());
} else {
CHECK_EQ(dataInput.getBatchSize(), shapeInput.getNumSequences());
}

// set output sequence info as shape sequence
Expand Down Expand Up @@ -105,8 +102,7 @@ void ExpandLayer::backward(const UpdateCallback& callback) {
if (!getInputGrad(0)) return;
MatrixPtr inputGrad = getInputGrad(0);
MatrixPtr outputGrad = getOutputGrad();
auto cpuSeqStartPos = type_ ? getInput(1).subSequenceStartPositions
: getInput(1).sequenceStartPositions;
auto cpuSeqStartPos = getInput(1).sequenceStartPositions;
size_t numSequences = cpuSeqStartPos->getSize() - 1;
const int* starts = cpuSeqStartPos->getData(false);

Expand Down
165 changes: 49 additions & 116 deletions paddle/gserver/tests/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,24 +1,29 @@
# gserver pacakge unittests

if(NOT MOBILE_INFERENCE)
################### test_ProtoDataProvider ############
add_unittest_without_exec(test_ProtoDataProvider
test_ProtoDataProvider.cpp)

# test_ProtoDataProvider will mkdir as same name,
# so if WORKING_DIRECTORY is default directory, then
# mkdir will get error.
add_test(NAME test_ProtoDataProvider
COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_ProtoDataProvider
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
endif()
add_simple_unittest(test_LinearChainCRF)
add_simple_unittest(test_MultinomialSampler)
add_simple_unittest(test_RecurrentLayer)

################# test_LayerGrad #######################
add_unittest_without_exec(test_LayerGrad
test_LayerGrad.cpp
LayerGradUtil.cpp)
add_test(NAME test_LayerGrad
COMMAND test_LayerGrad)
function(gserver_test TARGET)
add_unittest_without_exec(${TARGET}
${TARGET}.cpp
LayerGradUtil.cpp)
add_test(NAME ${TARGET}
COMMAND ${TARGET})
endfunction()

gserver_test(test_LayerGrad)
gserver_test(test_CRFLayerGrad)
gserver_test(test_CrossEntropyOverBeamGrad)
gserver_test(test_SeqSliceLayerGrad)
gserver_test(test_ActivationGrad)
gserver_test(test_ConvTrans)
gserver_test(test_PriorBox)
gserver_test(test_DetectionOutput)
gserver_test(test_ConvUnify)
gserver_test(test_BatchNorm)
gserver_test(test_KmaxSeqScore)
gserver_test(test_Expand)

########## test_Mkldnn layers and activations ##########
if(WITH_MKLDNN)
Expand All @@ -32,89 +37,6 @@ if(WITH_MKLDNN)
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
endif()

################ test_CRFLayerGrad ####################
add_unittest_without_exec(test_CRFLayerGrad
test_CRFLayerGrad.cpp
LayerGradUtil.cpp)
add_test(NAME test_CRFLayerGrad
COMMAND test_CRFLayerGrad)

################ test_CrossEntropyOverBeam ####################
add_unittest_without_exec(test_CrossEntropyOverBeam
test_CrossEntropyOverBeamGrad.cpp
LayerGradUtil.cpp)
add_test(NAME test_CrossEntropyOverBeam
COMMAND test_CrossEntropyOverBeam)

################ test_SeqSliceLayerGrad ####################
add_unittest_without_exec(test_SeqSliceLayerGrad
test_SeqSliceLayerGrad.cpp
LayerGradUtil.cpp)
add_test(NAME test_SeqSliceLayerGrad
COMMAND test_SeqSliceLayerGrad)

add_unittest_without_exec(test_ActivationGrad
test_ActivationGrad.cpp
LayerGradUtil.cpp)
add_test(NAME test_ActivationGrad
COMMAND test_ActivationGrad)
################# test_ConvTrans #######################
add_unittest_without_exec(test_ConvTrans
test_ConvTrans.cpp
LayerGradUtil.cpp)

add_test(NAME test_ConvTrans
COMMAND test_ConvTrans)
################# test_PriorBox #######################
add_unittest_without_exec(test_PriorBox
test_PriorBox.cpp
LayerGradUtil.cpp)

add_test(NAME test_PriorBox
COMMAND test_PriorBox)
################# test_DetectionOutput #######################
add_unittest_without_exec(test_DetectionOutput
test_DetectionOutput.cpp
LayerGradUtil.cpp)

add_test(NAME test_DetectionOutput
COMMAND test_DetectionOutput)
################# test_ConvUnify #######################
add_unittest_without_exec(test_ConvUnify
test_ConvUnify.cpp
LayerGradUtil.cpp)

add_test(NAME test_ConvUnify
COMMAND test_ConvUnify)
################# test_BatchNorm #######################
add_unittest_without_exec(test_BatchNorm
test_BatchNorm.cpp
LayerGradUtil.cpp)

add_test(NAME test_BatchNorm
COMMAND test_BatchNorm)


################# test_KmaxSeqScore #######################
add_unittest_without_exec(test_KmaxSeqScore
test_KmaxSeqScore.cpp
LayerGradUtil.cpp)

add_test(NAME test_KmaxSeqScore
COMMAND test_KmaxSeqScore)

if(NOT MOBILE_INFERENCE)
################## test_Evaluator #######################
add_unittest(test_Evaluator
test_Evaluator.cpp)
endif()

################ test_LinearChainCRF ####################
add_simple_unittest(test_LinearChainCRF)

############## test_MultinomialSampler ###################
add_simple_unittest(test_MultinomialSampler)

############## test_PyDataProvider ########################
if(WITH_PYTHON)
add_unittest_without_exec(test_PyDataProvider
Expand All @@ -125,9 +47,6 @@ if(WITH_PYTHON)
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
endif()

############### test_RecurrentLayer #######################
add_simple_unittest(test_RecurrentLayer)

############### test_WarpCTCLayer #######################
if(NOT WITH_DOUBLE)
add_unittest_without_exec(test_WarpCTCLayer
Expand All @@ -139,19 +58,33 @@ if(NOT WITH_DOUBLE)
endif()

if(NOT MOBILE_INFERENCE)
############### test_RecurrentGradientMachine ###############
# TODO(yuyang18): There is some bug in test_RecurrentGradientMachine
# I will fix it.
add_unittest_without_exec(test_RecurrentGradientMachine
test_RecurrentGradientMachine.cpp)
add_test(NAME test_RecurrentGradientMachine
COMMAND .set_python_path.sh -d
${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests
${CMAKE_CURRENT_BINARY_DIR}/test_RecurrentGradientMachine
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
endif()
################### test_ProtoDataProvider ############
add_unittest_without_exec(test_ProtoDataProvider
test_ProtoDataProvider.cpp)

if(NOT MOBILE_INFERENCE)
# test_ProtoDataProvider will mkdir as same name,
# so if WORKING_DIRECTORY is default directory, then
# mkdir will get error.
add_test(NAME test_ProtoDataProvider
COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_ProtoDataProvider
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)

################## test_Evaluator #######################
add_unittest(test_Evaluator
test_Evaluator.cpp)

############### test_RecurrentGradientMachine ###############
# TODO(yuyang18): There is some bug in test_RecurrentGradientMachine
# I will fix it.
add_unittest_without_exec(test_RecurrentGradientMachine
test_RecurrentGradientMachine.cpp)
add_test(NAME test_RecurrentGradientMachine
COMMAND .set_python_path.sh -d
${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests
${CMAKE_CURRENT_BINARY_DIR}/test_RecurrentGradientMachine
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)

############### test_NetworkCompare ###############
add_unittest_without_exec(test_NetworkCompare
test_NetworkCompare.cpp)
if(WITH_GPU)
Expand Down
116 changes: 116 additions & 0 deletions paddle/gserver/tests/test_Expand.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,116 @@
/* 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 <gtest/gtest.h>
#include <string>
#include <vector>

#include "LayerGradUtil.h"
#include "paddle/testing/TestUtil.h"

using namespace paddle; // NOLINT
using namespace std; // NOLINT

// Do one forward pass of expand layer and check to see if its output
// matches the given result.(Test onlyCPU currently.)
void doOneExpandTest(string trans_type,
bool hasSubseq,
bool useGpu,
Argument& input1,
Argument& input2,
Argument& result) {
FLAGS_use_gpu = false;
// Setting up the expand layer
TestConfig config;
config.layerConfig.set_type("expand");

auto inputType1 =
trans_type == "non-seq" ? INPUT_DENSE_DIM_DATA : INPUT_SEQUENCE_DATA;
config.inputDefs.push_back({inputType1, "layer0", 1, 0});
auto inputType2 =
hasSubseq ? INPUT_HASSUB_SEQUENCE_DATA : INPUT_SEQUENCE_DATA;

config.inputDefs.push_back({inputType2, "layer1", 1, 0});
config.layerConfig.add_inputs();
config.layerConfig.add_inputs();
config.layerConfig.set_trans_type(trans_type);

// data layer initialize
std::vector<DataLayerPtr> dataLayers;
LayerMap layerMap;
vector<Argument> datas;
initDataLayer(
config, &dataLayers, &datas, &layerMap, "expand", 1, false, useGpu);
dataLayers[0]->getOutput() = input1;
dataLayers[1]->getOutput() = input2;

// test layer initialize
std::vector<ParameterPtr> parameters;
LayerPtr expandLayer;
initTestLayer(config, &layerMap, &parameters, &expandLayer);
expandLayer->forward(PASS_GC);
checkMatrixEqual(expandLayer->getOutputValue(), result.value);
}

TEST(Layer, ExpandLayerFwd) {
bool useGpu = false;

// CPU case 1. non-seq expand to seq
// input1 = 1,2,3
// input2 = [4,5],[6],[7,8,9]
// result = [1,1],[2],[3,3,3]
Argument input1, input2, result;
input1.value = Matrix::create(3, 1, false, useGpu);
real input1Data[] = {1, 2, 3};
input1.value->setData(input1Data);

input2.value = Matrix::create(6, 1, false, useGpu);
real input2Data[] = {4, 5, 6, 7, 8, 9};
input2.value->setData(input2Data);
input2.sequenceStartPositions = ICpuGpuVector::create(4, useGpu);
int input2Seq[] = {0, 2, 3, 6};
input2.sequenceStartPositions->copyFrom(input2Seq, 4, useGpu);

result.value = Matrix::create(6, 1, false, useGpu);
real resultData[] = {1, 1, 2, 3, 3, 3};
result.value->setData(resultData);

doOneExpandTest("non-seq", false, useGpu, input1, input2, result);

// CPU case 2. non-seq expand to sub-seq
// input1 = 1,2,3
// input2 = [[[4,5],[6]],[[7,8,9]]
// result = [[1,1],[2]],[[3,3,3]]
input2.subSequenceStartPositions = ICpuGpuVector::create(3, useGpu);
int input2SubSeq[] = {0, 3, 6};
input2.subSequenceStartPositions->copyFrom(input2SubSeq, 3, useGpu);

doOneExpandTest("non-seq", true, useGpu, input1, input2, result);

// CPU case 3. seq expand to sub-seq
// input1 = [1,2],[3]
// input2 = [[[4,5],[6]],[[7,8,9]]
// result = [[1,1],[2]],[[3,3,3]]
input1.sequenceStartPositions = ICpuGpuVector::create(3, useGpu);
int input1Seq[] = {0, 2, 3};
input1.sequenceStartPositions->copyFrom(input1Seq, 3, useGpu);

doOneExpandTest("seq", true, useGpu, input1, input2, result);
}

int main(int argc, char** argv) {
testing::InitGoogleTest(&argc, argv);
initMain(argc, argv);
return RUN_ALL_TESTS();
}
21 changes: 11 additions & 10 deletions paddle/gserver/tests/test_LayerGrad.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,6 +33,7 @@ DECLARE_int32(gpu_id);
DECLARE_double(checkgrad_eps);
DECLARE_bool(thread_local_rand_use_global_seed);
DECLARE_bool(prev_batch_state);
DECLARE_int32(fixed_seq_length);

TEST(Operator, dot_mul) {
TestConfig config;
Expand Down Expand Up @@ -883,19 +884,19 @@ TEST(Layer, huber_two_class) {
}

void testExpandLayer(string trans_type, bool hasSubseq) {
if (hasSubseq && trans_type == "seq") {
FLAGS_fixed_seq_length = 1;
}
TestConfig config;
config.layerConfig.set_type("expand");

config.inputDefs.push_back(
{trans_type == "non-seq" ? INPUT_DENSE_DIM_DATA : INPUT_SEQUENCE_DATA,
"layer_0",
10,
0});
config.inputDefs.push_back(
{hasSubseq ? INPUT_HASSUB_SEQUENCE_DATA : INPUT_SEQUENCE_DATA,
"layer_1",
10,
0});
auto inputType1 =
trans_type == "non-seq" ? INPUT_DENSE_DIM_DATA : INPUT_SEQUENCE_DATA;
config.inputDefs.push_back({inputType1, "layer_0", 10, 0});
auto inputType2 =
hasSubseq ? INPUT_HASSUB_SEQUENCE_DATA : INPUT_SEQUENCE_DATA;
config.inputDefs.push_back({inputType2, "layer_1", 10, 0});

config.layerConfig.add_inputs();
config.layerConfig.add_inputs();
config.layerConfig.set_trans_type(trans_type);
Expand Down