Merge pull request #4509 from hedaoyuan/inference

Build mobile inference library for minimum size

CMakeLists.txt

@@ -86,6 +86,14 @@ if(ANDROID OR IOS)
         "Disable MKLDNN when cross-compiling for Android and iOS" FORCE)
     set(WITH_MKLML OFF CACHE STRING
         "Disable MKLML package when cross-compiling for Android and iOS" FORCE)
+
+    # Compile PaddlePaddle mobile inference library
+    if (NOT WITH_C_API)
+        set(WITH_C_API ON CACHE STRING
+            "Always compile the C_API when cross-compiling for Android and iOS" FORCE)
+    endif()
+    set(MOBILE_INFERENCE ON)
+    add_definitions(-DPADDLE_MOBILE_INFERENCE)
 endif()
 
 set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
@@ -160,9 +168,11 @@ endif(USE_NNPACK)
 
 add_subdirectory(proto)
 
-# "add_subdirectory(go)" should be placed after the following loine,
-# because it depends on paddle/optimizer.
-add_subdirectory(paddle/optimizer)
+if(NOT MOBILE_INFERENCE)
+    # "add_subdirectory(go)" should be placed after the following loine,
+    # because it depends on paddle/optimizer.
+    add_subdirectory(paddle/optimizer)
+endif()
 
 # "add_subdirectory(paddle)" and "add_subdirectory(python)" should be
 # placed after this block, because they depends on it.

cmake/util.cmake

@@ -73,6 +73,23 @@ function(link_paddle_exe TARGET_NAME)
         generate_rdma_links()
     endif()
 
+    if(MOBILE_INFERENCE)
+        target_circle_link_libraries(${TARGET_NAME}
+            ARCHIVE_START
+            paddle_gserver
+            paddle_function
+            ARCHIVE_END
+            paddle_math
+            paddle_utils
+            paddle_parameter
+            paddle_proto
+            paddle_cuda
+            ${EXTERNAL_LIBS}
+            ${CMAKE_THREAD_LIBS_INIT}
+            ${CMAKE_DL_LIBS}
+            ${RDMA_LD_FLAGS}
+            ${RDMA_LIBS})
+    else()
         target_circle_link_libraries(${TARGET_NAME}
             ARCHIVE_START
             paddle_gserver
@@ -92,6 +109,7 @@ function(link_paddle_exe TARGET_NAME)
             ${CMAKE_DL_LIBS}
             ${RDMA_LD_FLAGS}
             ${RDMA_LIBS})
+    endif()
 
     if(ANDROID)
         target_link_libraries(${TARGET_NAME} log)

paddle/CMakeLists.txt

 add_subdirectory(cuda)
 add_subdirectory(function)
 add_subdirectory(utils)
-add_subdirectory(testing)
 add_subdirectory(math)
-add_subdirectory(parameter)
 add_subdirectory(gserver)
-add_subdirectory(pserver)
-add_subdirectory(trainer)
-add_subdirectory(scripts)
-add_subdirectory(string)
+add_subdirectory(parameter)
+add_subdirectory(testing)
+
+if(MOBILE_INFERENCE)
+  add_subdirectory(capi)
+else()
+  add_subdirectory(pserver)
+  add_subdirectory(trainer)
+  add_subdirectory(string)
+  add_subdirectory(scripts)
 
-if(Boost_FOUND)
+  if(WITH_C_API)
+    add_subdirectory(capi)
+  endif()
+
+  if(Boost_FOUND)
     add_subdirectory(memory)
     add_subdirectory(platform)
     add_subdirectory(framework)
     add_subdirectory(operators)
     add_subdirectory(pybind)
-endif()
-
-if(WITH_C_API)
-  add_subdirectory(capi)
-endif()
+  endif()
 
-if(WITH_SWIG_PY)
+  if(WITH_SWIG_PY)
     add_subdirectory(api)
+  endif()
 endif()

paddle/capi/CMakeLists.txt

@@ -37,9 +37,7 @@ set(PADDLE_CAPI_INFER_LIBS
     paddle_cuda
     paddle_function
     paddle_gserver
-    paddle_proto
-    paddle_pserver
-    paddle_network)
+    paddle_proto)
 
 cc_library(paddle_capi_whole DEPS paddle_capi ${PADDLE_CAPI_INFER_LIBS})
 

paddle/capi/tests/CMakeLists.txt

@@ -4,11 +4,12 @@ add_unittest(capi_test_mats test_Vector.cpp
 target_include_directories(capi_test_mats PUBLIC ${PADDLE_CAPI_INC_PATH})
 target_link_libraries(capi_test_mats paddle_capi)
 
-
-add_unittest_without_exec(capi_test_gradientMachine test_GradientMachine.cpp)
-target_include_directories(capi_test_gradientMachine PUBLIC
+if(NOT MOBILE_INFERENCE)
+    add_unittest_without_exec(capi_test_gradientMachine test_GradientMachine.cpp)
+    target_include_directories(capi_test_gradientMachine PUBLIC
       ${PADDLE_CAPI_INC_PATH})
-target_link_libraries(capi_test_gradientMachine paddle_capi)
-add_test(NAME capi_test_gradientMachine
+    target_link_libraries(capi_test_gradientMachine paddle_capi)
+    add_test(NAME capi_test_gradientMachine
       COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/capi_test_gradientMachine
       WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/capi/tests)
+endif()

paddle/gserver/CMakeLists.txt

@@ -60,6 +60,36 @@ if(NOT WITH_PYTHON)
             dataproviders/PyDataProvider.h)
 endif()
 
+if(MOBILE_INFERENCE)
+    # Remove evaluators
+    list(REMOVE_ITEM GSERVER_SOURCES
+         layers/ValidationLayer.cpp
+         evaluators/Evaluator.cpp
+         evaluators/DetectionMAPEvaluator.cpp
+         evaluators/CTCErrorEvaluator.cpp
+         evaluators/ChunkEvaluator.cpp)
+
+    # Remove dataproviders
+    list(REMOVE_ITEM GSERVER_SOURCES
+         dataproviders/DataProvider.cpp
+         dataproviders/MultiDataProvider.cpp
+         dataproviders/ProtoDataProvider.cpp
+         dataproviders/PyDataProvider2.cpp
+         dataproviders/PyDataProvider.cpp)
+
+    # Remove useless gradientmachines
+    list(REMOVE_ITEM GSERVER_SOURCES
+         gradientmachines/MultiNetwork.cpp
+         gradientmachines/RecurrentGradientMachine.cpp
+         gradientmachines/ParallelNeuralNetwork.cpp
+         gradientmachines/GradientMachineMode.cpp
+         gradientmachines/MultiGradientMachine.cpp)
+
+    # Remove useless layers
+    list(REMOVE_ITEM GSERVER_SOURCES
+    	 layers/RecurrentLayerGroup.cpp)
+endif()
+
 if(WITH_GPU)
     cuda_add_library(paddle_gserver ${GSERVER_SOURCES})
 else()

paddle/gserver/gradientmachines/GradientMachine.cpp

@@ -17,12 +17,15 @@ limitations under the License. */
 #include <fstream>
 #include "paddle/utils/Logging.h"
 
+#include "NeuralNetwork.h"
+#include "hl_gpu.h"
+
+#ifndef PADDLE_MOBILE_INFERENCE
 #include "GradientMachineMode.h"
 #include "MultiGradientMachine.h"
 #include "MultiNetwork.h"
-#include "NeuralNetwork.h"
 #include "ParallelNeuralNetwork.h"
-#include "hl_gpu.h"
+#endif
 
 namespace paddle {
 
@@ -30,13 +33,16 @@ GradientMachine* GradientMachine::create(
     const ModelConfig& config,
     int mode,
     const std::vector<ParameterType>& parameterTypes) {
+#ifndef PADDLE_MOBILE_INFERENCE
   if (auto gm = IGradientMachineMode::tryCreateGradientMachine(mode, config)) {
     return gm;
   }
   if (FLAGS_trainer_count > 1) {
     return new MultiGradientMachine(config, FLAGS_use_gpu);
   }
+#endif
   if (FLAGS_trainer_count == 1) {  // single
+#ifndef PADDLE_MOBILE_INFERENCE
     NeuralNetwork* nn;
     if (config.type() == "multi_nn") {
       /* multi submodel calculate, thread(s) will be initialized inside */
@@ -48,6 +54,9 @@ GradientMachine* GradientMachine::create(
       /* single thread calculate */
       nn = NeuralNetwork::create(config);
     }
+#else
+    NeuralNetwork* nn = NeuralNetwork::create(config);
+#endif
     ParamInitCallback testParamInitCb = [](int paramId, Parameter* para) {
       para->enableType(PARAMETER_VALUE);
     };

paddle/gserver/gradientmachines/GradientMachine.h

@@ -20,13 +20,16 @@ limitations under the License. */
 #include "ModelConfig.pb.h"
 #include "TrainerConfig.pb.h"
 #include "paddle/gserver/dataproviders/DataProvider.h"
-#include "paddle/gserver/evaluators/Evaluator.h"
 #include "paddle/gserver/layers/Layer.h"
 #include "paddle/math/Matrix.h"
 #include "paddle/parameter/Parameter.h"
 #include "paddle/parameter/ParameterUpdaterBase.h"
 #include "paddle/utils/Thread.h"
 
+#ifndef PADDLE_MOBILE_INFERENCE
+#include "paddle/gserver/evaluators/Evaluator.h"
+#endif
+
 namespace paddle {
 /**
  * @brief A gradient machine is capable of calculating some outputs given
@@ -147,6 +150,7 @@ public:
 
   virtual void onPassEnd() = 0;
 
+#ifndef PADDLE_MOBILE_INFERENCE
   /**
    * Create an evaluator which can be used for eval()
    */
@@ -156,6 +160,7 @@ public:
    * evaluate using the given evaluator
    */
   virtual void eval(Evaluator* evaluator) const = 0;
+#endif
 
   std::vector<ParameterPtr>& getParameters() { return parameters_; }
 

paddle/gserver/gradientmachines/NeuralNetwork.cpp

@@ -14,15 +14,17 @@ limitations under the License. */
 
 #include "paddle/utils/Util.h"
 
+#include "NeuralNetwork.h"
+#include "hl_gpu.h"
+#include "paddle/gserver/layers/AgentLayer.h"
 #include "paddle/utils/CustomStackTrace.h"
 #include "paddle/utils/Logging.h"
+#include "paddle/utils/Stat.h"
 
+#ifndef PADDLE_MOBILE_INFERENCE
 #include "MultiNetwork.h"
-#include "NeuralNetwork.h"
 #include "RecurrentGradientMachine.h"
-#include "hl_gpu.h"
-#include "paddle/gserver/layers/AgentLayer.h"
-#include "paddle/utils/Stat.h"
+#endif
 
 namespace paddle {
 void parameterInitNN(int paramId,
@@ -54,6 +56,7 @@ void parameterInitNN(int paramId,
 }
 
 NeuralNetwork* NeuralNetwork::create(const ModelConfig& config) {
+#ifndef PADDLE_MOBILE_INFERENCE
   if (config.type() == "recurrent_nn") {
     return newNeuralNetwork("root");
   } else if (config.type() == "multi_nn") {
@@ -61,6 +64,9 @@ NeuralNetwork* NeuralNetwork::create(const ModelConfig& config) {
   } else {
     return newNeuralNetwork();
   }
+#else
+  return new NeuralNetwork();
+#endif
 }
 
 std::map<std::string, bool> NeuralNetwork::dllInitMap;
@@ -304,6 +310,8 @@ void NeuralNetwork::onPassEnd() {
   }
 }
 
+#ifndef PADDLE_MOBILE_INFERENCE
+
 class CombinedEvaluator : public Evaluator {
 public:
   void addEvaluator(std::unique_ptr<Evaluator>&& evaluator) {
@@ -466,6 +474,8 @@ Evaluator* NeuralNetwork::makeEvaluator() const {
 
 void NeuralNetwork::eval(Evaluator* evaluator) const { evaluator->eval(*this); }
 
+#endif
+
 void NeuralNetwork::setOutputGrad(const std::vector<Argument>& args) {
   CHECK_GE(outputLayers_.size(), args.size());
   for (size_t i = 0; i < args.size(); ++i) {

paddle/gserver/gradientmachines/NeuralNetwork.h

@@ -97,9 +97,12 @@ public:
 
   virtual void onPassEnd();
 
+#ifndef PADDLE_MOBILE_INFERENCE
   virtual Evaluator* makeEvaluator() const;
 
   virtual void eval(Evaluator* evaluator) const;
+#endif
+
   virtual void resetState();
   virtual void setOutputGrad(const std::vector<Argument>& args);
 

paddle/gserver/layers/Layer.cpp

@@ -15,11 +15,14 @@ limitations under the License. */
 #include "paddle/utils/Util.h"
 
 #include "CostLayer.h"
-#include "ValidationLayer.h"
 #include "paddle/math/SparseMatrix.h"
 #include "paddle/utils/Error.h"
 #include "paddle/utils/Logging.h"
 
+#ifndef PADDLE_MOBILE_INFERENCE
+#include "ValidationLayer.h"
+#endif
+
 DEFINE_bool(log_error_clipping, false, "enable log error clipping or not");
 
 namespace paddle {
@@ -103,10 +106,12 @@ LayerPtr Layer::create(const LayerConfig& config) {
     return LayerPtr(new MultiClassCrossEntropy(config));
   else if (type == "rank-cost")
     return LayerPtr(new RankingCost(config));
+#ifndef PADDLE_MOBILE_INFERENCE
   else if (type == "auc-validation")
     return LayerPtr(new AucValidation(config));
   else if (type == "pnpair-validation")
     return LayerPtr(new PnpairValidation(config));
+#endif
 
   return LayerPtr(registrar_.createByType(config.type(), config));
 }

paddle/gserver/tests/CMakeLists.txt

 # gserver pacakge unittests
 
+if(NOT MOBILE_INFERENCE)
 ################### test_ProtoDataProvider ############
-add_unittest_without_exec(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
+    # 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()
 
 ################# test_LayerGrad #######################
 add_unittest_without_exec(test_LayerGrad
@@ -98,9 +100,11 @@ add_unittest_without_exec(test_KmaxSeqScore
 add_test(NAME test_KmaxSeqScore
     COMMAND test_KmaxSeqScore)
 
+if(NOT MOBILE_INFERENCE)
 ################## test_Evaluator #######################
-add_unittest(test_Evaluator
+    add_unittest(test_Evaluator
         test_Evaluator.cpp)
+endif()
 
 ################ test_LinearChainCRF ####################
 add_simple_unittest(test_LinearChainCRF)
@@ -131,27 +135,31 @@ if(NOT WITH_DOUBLE)
         WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
 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
+  # 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
+  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()
 
-add_unittest_without_exec(test_NetworkCompare
+if(NOT MOBILE_INFERENCE)
+    add_unittest_without_exec(test_NetworkCompare
         test_NetworkCompare.cpp)
-if(WITH_GPU)
+    if(WITH_GPU)
         add_test(NAME test_NetworkCompare
             COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=true
             WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
-else()
+    else()
         add_test(NAME test_NetworkCompare
             COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=false
             WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
+    endif()
 endif()
 
 

paddle/gserver/tests/LayerGradUtil.h

@@ -15,7 +15,6 @@ limitations under the License. */
 #pragma once
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
-#include "paddle/trainer/Trainer.h"
 
 #include "paddle/testing/TestUtil.h"
 using namespace std;  // NOLINT

paddle/gserver/tests/test_ActivationGrad.cpp

@@ -17,7 +17,6 @@ limitations under the License. */
 #include <vector>
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
-#include "paddle/trainer/Trainer.h"
 
 #include "LayerGradUtil.h"
 #include "paddle/testing/TestUtil.h"

paddle/gserver/tests/test_BatchNorm.cpp

@@ -17,7 +17,6 @@ limitations under the License. */
 #include <vector>
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
-#include "paddle/trainer/Trainer.h"
 #include "paddle/utils/GlobalConstants.h"
 
 #include "LayerGradUtil.h"

paddle/gserver/tests/test_CRFLayerGrad.cpp

@@ -16,7 +16,6 @@ limitations under the License. */
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
 #include "paddle/gserver/layers/LinearChainCRF.h"
-#include "paddle/trainer/Trainer.h"
 
 #include "LayerGradUtil.h"
 #include "paddle/testing/TestUtil.h"

paddle/gserver/tests/test_ConvTrans.cpp

@@ -18,7 +18,6 @@ limitations under the License. */
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
 #include "paddle/math/MathUtils.h"
-#include "paddle/trainer/Trainer.h"
 #include "paddle/utils/GlobalConstants.h"
 
 #include "LayerGradUtil.h"

paddle/gserver/tests/test_ConvUnify.cpp

@@ -18,7 +18,6 @@ limitations under the License. */
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
 #include "paddle/math/MathUtils.h"
-#include "paddle/trainer/Trainer.h"
 #include "paddle/utils/GlobalConstants.h"
 
 #include "LayerGradUtil.h"

paddle/gserver/tests/test_CrossEntropyOverBeamGrad.cpp

@@ -18,7 +18,6 @@ limitations under the License. */
 #include <gtest/gtest.h>
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
-#include "paddle/trainer/Trainer.h"
 
 #include "LayerGradUtil.h"
 #include "paddle/testing/TestUtil.h"

paddle/gserver/tests/test_KmaxSeqScore.cpp

@@ -18,7 +18,6 @@ limitations under the License. */
 #include <vector>
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
-#include "paddle/trainer/Trainer.h"
 #include "paddle/utils/GlobalConstants.h"
 
 #include "LayerGradUtil.h"

paddle/gserver/tests/test_LayerGrad.cpp

@@ -21,7 +21,6 @@ limitations under the License. */
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
 #include "paddle/math/MathUtils.h"
-#include "paddle/trainer/Trainer.h"
 
 #include "LayerGradUtil.h"
 #include "paddle/testing/TestUtil.h"

paddle/gserver/tests/test_SelectiveFCLayer.cpp

@@ -24,7 +24,6 @@ limitations under the License. */
 #include "paddle/gserver/layers/Layer.h"
 #include "paddle/gserver/layers/SelectiveFullyConnectedLayer.h"
 #include "paddle/math/CpuSparseMatrix.h"
-#include "paddle/trainer/Trainer.h"
 
 using namespace paddle;  // NOLINT
 using namespace std;     // NOLINT

paddle/gserver/tests/test_SeqSliceLayerGrad.cpp

@@ -15,7 +15,6 @@ limitations under the License. */
 #include <gtest/gtest.h>
 #include "ModelConfig.pb.h"
 #include "paddle/gserver/layers/DataLayer.h"
-#include "paddle/trainer/Trainer.h"
 
 #include "LayerGradUtil.h"
 #include "paddle/testing/TestUtil.h"