Merge pull request #324 from cocodark/develop

add impl for executor'predict

CMakeLists.txt

 cmake_minimum_required(VERSION 3.0)
 project(paddle-mobile)
-add_definitions(-DPADDLE_MOBILE_DEBUG="true")
+add_definitions(-DPADDLE_MOBILE_DEBUG)
+add_definitions(-DENABLE_EXCEPTION)
 
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
 set(CMAKE_BUILD_TYPE RelWithDebInfo)

scripts/push2android.sh

+#!/usr/bin/env sh
+
+push_fn () {
+MODELS_PATH="../test/models/*"
+EXE_FILE="../test/build/*"
+EXE_DIR="data/local/tmp/bin"
+MODELS_DIR="data/local/tmp/models"
+LIB_PATH="../build/release/arm-v7a/build/*"
+adb push ${EXE_FILE} ${EXE_DIR}
+adb push ${LIB_PATH} ${EXE_DIR}
+adb push ${MODELS_PATH} ${MODELS_DIR}
+echo "test files sync completed"
+}
+push_fn

src/common/enforce.h

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #pragma once
 
-#ifdef PADDLE_MOBILE_DEBUG
+#ifdef ENABLE_EXCEPTION
 #include <stdio.h>
 #include <exception>
 #include <sstream>
@@ -25,7 +25,7 @@ limitations under the License. */
 
 namespace paddle_mobile {
 
-#ifdef PADDLE_MOBILE_DEBUG
+#ifdef ENABLE_EXCEPTION
 struct PaddleMobileException : public std::exception {
   const std::string exception_prefix = "paddle mobile C++ Exception: \n";
   std::string message;
@@ -64,7 +64,7 @@ struct PaddleMobileException : public std::exception {
   }
 #else
 #define PADDLE_MOBILE_THROW_EXCEPTION(...)
-#define PADDLE_MOBILE_ASSERT(stat, ...)
+#define PADDLE_MOBILE_ENFORCE(stat, ...)
 #endif
 
 }  // namespace paddle_mobile

src/framework/operator.cpp

@@ -28,18 +28,6 @@ vector<string> OperatorBase<Dtype>::GetOutKeys() const {
   return it->second.second;
 }
 
-template <typename T>
-static T *GetVarValue(const string &key, const VariableNameMap &var_map,
-                      const Scope &scope) {
-  auto var_vec = var_map.at(key);
-  if (!var_vec.empty()) {
-    auto var = scope.FindVar(var_vec[0]);
-    return var->GetMutable<T>();
-  } else {
-    return nullptr;
-  }
-}
-
 template <typename Dtype>
 OperatorBase<Dtype>::OperatorBase(const std::string &type,
                                   const VariableNameMap &inputs,
@@ -60,7 +48,7 @@ void OperatorBase<Dtype>::CheckAllInputOutputSet() const {}
 template <typename Dtype>
 void OperatorBase<Dtype>::Run() const {
   RunImpl();
-#ifdef PADDLE_MOBILE_DEBUG
+#if (PADDLE_MOBILE_DEBUG)
   vector<string> output_keys = GetOutKeys();
   for (const auto key : output_keys) {
     Tensor *out_ = GetVarValue<framework::LoDTensor>(key, outputs_, *scope_);

src/framework/operator.h

@@ -39,6 +39,18 @@ namespace framework {
 using std::string;
 using std::vector;
 
+template <typename T>
+static T *GetVarValue(const string &key, const VariableNameMap &var_map,
+                      const Scope &scope) {
+  auto var_vec = var_map.at(key);
+  if (!var_vec.empty()) {
+    auto var = scope.FindVar(var_vec[0]);
+    return var->GetMutable<T>();
+  } else {
+    return nullptr;
+  }
+}
+
 template <typename Dtype>
 class OperatorBase : PaddleMobileObject {
  public:

src/io.cpp

@@ -371,31 +371,47 @@ void Executor<Dtype, P>::InitMemory() {
 }
 
 template <typename Dtype, Precision P>
-void Executor<Dtype, P>::Predict(const framework::Tensor &t, int block_id) {
+std::shared_ptr<framework::Tensor> Executor<Dtype, P>::Predict(
+    const framework::Tensor &t) {
   framework::Variable *g_feed_value = program_.scope->Var("feed");
   framework::Tensor *feed_tensor =
       g_feed_value->GetMutable<framework::LoDTensor>();
   feed_tensor->Resize(t.dims());
   feed_tensor->ShareDataWith(t);
   std::shared_ptr<framework::BlockDesc> to_predict_block =
-      to_predict_program_->Block(block_id);
+      to_predict_program_->Block(0);
   for (int j = 0; j < ops_of_block_[*to_predict_block.get()].size(); ++j) {
     auto op = ops_of_block_[*to_predict_block.get()][j];
     op->Run();
   }
+  auto ops = ops_of_block_[*to_predict_program_->Block(0)];
+  auto last_op = ops.rbegin();
+  auto output_map = (*last_op)->Outputs();
+  std::vector<std::string> out_keys = (*last_op)->GetOutKeys();
+  PADDLE_MOBILE_ENFORCE(out_keys.size() > 0, "the last op contains no output");
+  framework::LoDTensor *output_tensor =
+      framework::GetVarValue<framework::LoDTensor>(out_keys[0], output_map,
+                                                   *(program_.scope));
+  return std::shared_ptr<framework::Tensor>(output_tensor);
+}
+template <typename Dtype, Precision P>
+std::shared_ptr<framework::Tensor> Executor<Dtype, P>::Predict(
+    const framework::Tensor &t, int block_id) {
+  return Predict(t);
 }
 
 template <typename Dtype, Precision P>
 std::vector<typename Executor<Dtype, P>::Ptype> Executor<Dtype, P>::Predict(
     const std::vector<Ptype> &input, const std::vector<int64_t> &dims) {
   framework::Tensor tensor(input, framework::make_ddim(dims));
-
-  Predict(tensor, 0);
-
-  framework::Variable *g_feed_value = program_.scope->Var("col");
-  auto feed_tensor = g_feed_value->GetMutable<framework::Tensor>();
-
-  return {};
+  std::shared_ptr<framework::Tensor> output_tensor = Predict(tensor, 0);
+  Executor<Dtype, P>::Ptype *output_ptr =
+      output_tensor->data<typename Executor<Dtype, P>::Ptype>();
+  std::vector<typename Executor<Dtype, P>::Ptype> result_vector;
+  for (int j = 0; j < output_tensor->numel(); ++j) {
+    result_vector.push_back(output_ptr[j]);
+  }
+  return result_vector;
 }
 
 template class Executor<CPU, Precision::FP32>;

src/io.h

@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <memory.h>
+#include <map>
 #include <string>
 #include <vector>
 
@@ -44,24 +45,25 @@ class Executor {
  public:
   typedef typename PrecisionTrait<P>::ptype Ptype;
 
-  Executor() = default;
-
   Executor(const framework::Program<Dtype> p, int batch_size = 1,
            bool use_optimize = true);
 
-  //  std::shared_ptr<framework::Tensor> Predict(framework::Tensor &t);
+  std::shared_ptr<framework::Tensor> Predict(const framework::Tensor &t);
 
   std::vector<Ptype> Predict(const std::vector<Ptype> &input,
                              const std::vector<int64_t> &dims);
 
  protected:
+  Executor() = default;
+
   void InitMemory();
   void LoadMemory(const framework::VarDesc var_desc,
                   framework::LoDTensor *tensor, const std::string &file_path);
   framework::Program<Dtype> program_;
   int batch_size_ = 1;
   std::shared_ptr<framework::ProgramDesc> to_predict_program_;
-  void Predict(const framework::Tensor &t, int block_id);
+  std::shared_ptr<framework::Tensor> Predict(const framework::Tensor &t,
+                                             int block_id);
   std::map<framework::BlockDesc,
            std::vector<std::shared_ptr<framework::OperatorBase<Dtype>>>>
       ops_of_block_;

src/operators/math/softmax.cpp

@@ -136,9 +136,15 @@ class SoftmaxFuntor<CPU, T> {
 
  public:
   void operator()(const framework::Tensor *X, framework::Tensor *Y) {
+    const DDim dDim = X->dims();
+    for (int i = 0; i < dDim[0]; ++i) {
+      framework::Tensor sub_X = X->Slice(i, i + 1);
+      framework::Tensor sub_Y = Y->Slice(i, i + 1);
+
 #if __ARM_NEON
-    SoftmaxCacl(X, Y);
+      SoftmaxCacl(&sub_X, &sub_Y);
 #endif
+    }
   }
 };
 

test/net/test_googlenet.cpp

@@ -18,20 +18,17 @@ limitations under the License. */
 
 int main() {
   paddle_mobile::Loader<paddle_mobile::CPU> loader;
-  //  ../../../test/models/googlenet
-  //  ../../../test/models/mobilenet
   auto time1 = time();
   auto program = loader.Load(g_googlenet, false);
   auto time2 = time();
-  DLOG << "load cost :" << time_diff(time1, time1) << "ms";
+  DLOG << "load cost :" << time_diff(time1, time2) << "ms\n";
   paddle_mobile::Executor<paddle_mobile::CPU> executor(program, 1, false);
-
   std::vector<float> input;
   std::vector<int64_t> dims{1, 3, 224, 224};
   GetInput<float>(g_test_image_1x3x224x224, &input, dims);
   auto time3 = time();
   executor.Predict(input, dims);
   auto time4 = time();
-  DLOG << "predict cost :" << time_diff(time3, time4) << "ms";
+  DLOG << "predict cost :" << time_diff(time3, time4) << "ms\n";
   return 0;
 }

test/net/test_mobilenet.cpp

@@ -22,17 +22,22 @@ int main() {
   auto program = loader.Load(g_mobilenet, false);
   auto time2 = time();
   DLOG << "load cost :" << time_diff(time1, time1) << "ms";
-  paddle_mobile::Executor<paddle_mobile::CPU> executor(program, 1, false);
+  paddle_mobile::Executor<paddle_mobile::CPU> executor(program, 2, false);
 
-  std::vector<int64_t> dims{1, 3, 224, 224};
+  std::vector<int64_t> dims{2, 3, 224, 224};
   Tensor input_tensor;
-  SetupTensor<float>(&input_tensor, {1, 3, 224, 224}, static_cast<float>(0),
+  SetupTensor<float>(&input_tensor, {2, 3, 224, 224}, static_cast<float>(0),
                      static_cast<float>(1));
 
   std::vector<float> input(input_tensor.data<float>(),
                            input_tensor.data<float>() + input_tensor.numel());
   auto time3 = time();
-  executor.Predict(input, dims);
+  auto vec_result = executor.Predict(input, dims);
+  float sum = 0;
+  for (const auto item : vec_result) {
+    sum += item;
+  }
+  DLOG << "mobilenet output sum =" << sum;
   auto time4 = time();
   DLOG << "predict cost :" << time_diff(time3, time4) << "ms";
   return 0;