Merge pull request #4509 from hedaoyuan/inference

Build mobile inference library for minimum size

Author: NHZlX

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

@@ -73,25 +73,43 @@ function(link_paddle_exe TARGET_NAME)
         generate_rdma_links()
     endif()
 
-    target_circle_link_libraries(${TARGET_NAME}
-        ARCHIVE_START
-        paddle_gserver
-        paddle_function
-        ARCHIVE_END
-        paddle_pserver
-        paddle_trainer_lib
-        paddle_network
-        paddle_math
-        paddle_utils
-        paddle_parameter
-        paddle_proto
-        paddle_cuda
-        paddle_optimizer
-        ${EXTERNAL_LIBS}
-        ${CMAKE_THREAD_LIBS_INIT}
-        ${CMAKE_DL_LIBS}
-        ${RDMA_LD_FLAGS}
-        ${RDMA_LIBS})
+    if(MOBILE_INFERENCE)
+        target_circle_link_libraries(${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
+            paddle_function
+            ARCHIVE_END
+            paddle_pserver
+            paddle_trainer_lib
+            paddle_network
+            paddle_math
+            paddle_utils
+            paddle_parameter
+            paddle_proto
+            paddle_cuda
+            paddle_optimizer
+            ${EXTERNAL_LIBS}
+            ${CMAKE_THREAD_LIBS_INIT}
+            ${CMAKE_DL_LIBS}
+            ${RDMA_LD_FLAGS}
+            ${RDMA_LIBS})
+    endif()
 
     if(ANDROID)
         target_link_libraries(${TARGET_NAME} log)

paddle/CMakeLists.txt

@@ -1,27 +1,32 @@
 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)
-
-if(Boost_FOUND)
-  add_subdirectory(memory)
-  add_subdirectory(platform)
-  add_subdirectory(framework)
-  add_subdirectory(operators)
-  add_subdirectory(pybind)
-endif()
+add_subdirectory(parameter)
+add_subdirectory(testing)
 
-if(WITH_C_API)
+if(MOBILE_INFERENCE)
   add_subdirectory(capi)
-endif()
+else()
+  add_subdirectory(pserver)
+  add_subdirectory(trainer)
+  add_subdirectory(string)
+  add_subdirectory(scripts)
+
+  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_SWIG_PY)
-  add_subdirectory(api)
+  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
-  ${PADDLE_CAPI_INC_PATH})
-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)
+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
+      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

@@ -17,26 +17,32 @@ 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 {
 
 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 @@ class GradientMachine {
 
   virtual void onPassEnd() = 0;
 
+#ifndef PADDLE_MOBILE_INFERENCE
   /**
    * Create an evaluator which can be used for eval()
    */
@@ -156,6 +160,7 @@ class GradientMachine {
    * 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,
@@ -54,13 +56,17 @@ 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") {
     return new MultiNetwork("root");
   } 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 @@ class NeuralNetwork : public GradientMachine {
 
   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);