Merge pull request #877 from xiebaiyuan/develop

Merge remote-tracking branch 'refs/remotes/upstream/develop' into develop

src/framework/program/program.h

@@ -31,6 +31,8 @@ class Program {
   std::string para_path;
   bool combined = false;
   bool quantification = false;
+  size_t combined_params_len;
+  const uint8_t *combined_params_buf;
 
  private:
 };

src/io/executor.cpp

@@ -63,6 +63,8 @@ Executor<Dtype, P>::Executor(const framework::Program<Dtype> p, int batch_size,
   }
   Variable *variable_ptr = program_.scope->Var("batch_size");
   variable_ptr[0].SetValue<int>(batch_size);
+  PADDLE_MOBILE_ENFORCE(to_predict_program_ != nullptr,
+                        "to_predict_program_ == NULL!");
   const std::vector<std::shared_ptr<framework::BlockDesc>> blocks =
       to_predict_program_->Blocks();
 #ifdef PADDLE_EXECUTOR_MULTITHREAD
@@ -234,8 +236,15 @@ void Executor<Dtype, P>::InitMemory() {
 
 template <typename Dtype, Precision P>
 void Executor<Dtype, P>::InitCombineMemory() {
-  LOG(kLOG_INFO) << " begin init combine memory";
-  char *origin_data = Get_binary_data(program_.para_path);
+  char *origin_data;
+  if (program_.combined_params_buf && program_.combined_params_len) {
+    LOG(kLOG_INFO) << "use outter memory";
+    origin_data = (char *)program_.combined_params_buf;
+  } else {
+    LOG(kLOG_INFO) << " begin init combine memory";
+    origin_data = Get_binary_data(program_.para_path);
+  }
+  PADDLE_MOBILE_ENFORCE(origin_data != nullptr, "origin_data==nullptr!!!");
   char *data = origin_data;
   for (const auto &block : to_predict_program_->Blocks()) {
     for (const auto &var_desc : block->Vars()) {

src/io/loader.cpp

@@ -20,6 +20,62 @@ limitations under the License. */
 namespace paddle_mobile {
 using framework::Variable;
 
+/**
+ * muteandresize tensor as originProgramDesc and scope in loadParams
+ *
+ * @param originProgramDesc
+ * @param scope
+ */
+void InitMemoryFromProgram(
+    std::shared_ptr<framework::ProgramDesc> &originProgramDesc,
+    std::shared_ptr<framework::Scope> &scope) {
+  for (const auto &block : originProgramDesc.get()->Blocks()) {
+    for (const auto &var_desc : block->Vars()) {
+      auto var = scope.get()->Var(var_desc->Name());
+      if (var_desc->Type() == framework::VARTYPE_TYPE_LOD_TENSOR) {
+        if (var_desc->Persistable() &&
+            var_desc->Type() != framework::VARTYPE_TYPE_FEED_MINIBATCH &&
+            var_desc->Type() != framework::VARTYPE_TYPE_FETCH_LIST) {
+          auto dim = var_desc->Tensor_desc().Dims();
+          auto tensor = var->GetMutable<framework::LoDTensor>();
+          tensor->Resize(framework::make_ddim(dim));
+        } else {
+          auto dim = var_desc->Tensor_desc().Dims();
+          PADDLE_MOBILE_ENFORCE(dim.size() > 0, "dim size is 0");
+          dim[0] = 1;
+          auto tensor = var->GetMutable<framework::LoDTensor>();
+          tensor->Resize(framework::make_ddim(dim));
+        }
+      } else {
+        // TODO(codeWorm): some.
+      }
+    }
+  }
+}
+/**
+ * fusion and print someinfos
+ * @tparam Dtype
+ * @tparam P
+ * @param optimize
+ * @param can_add_split
+ * @param program
+ * @param originProgramDesc
+ */
+template <typename Dtype, Precision P>
+void FusionAndPrintInfos(
+    bool &optimize, bool &can_add_split, framework::Program<Dtype, P> &program,
+    const std::shared_ptr<framework::ProgramDesc> &originProgramDesc) {
+  if (optimize) {
+    framework::ProgramOptimize program_optimize;
+    program.optimizeProgram =
+        program_optimize.FusionOptimize(originProgramDesc, can_add_split);
+  }
+  if (optimize) {
+    program.optimizeProgram->Description("optimize: ");
+  } else {
+    originProgramDesc->Description("program: ");
+  }
+}
 static size_t ReadBuffer(const char *file_name, uint8_t **out) {
   FILE *fp;
   fp = fopen(file_name, "rb");
@@ -87,46 +143,52 @@ const framework::Program<Dtype, P> Loader<Dtype, P>::LoadProgram(
   framework::Program<Dtype, P> program;
   program.originProgram = originProgramDesc;
   program.quantification = quantification;
-
+  program.combined_params_len = 0;
+  program.combined_params_buf = nullptr;
   auto scope = std::make_shared<framework::Scope>();
   program.scope = scope;
 
-  for (const auto &block : originProgramDesc->Blocks()) {
-    for (auto var_desc : block->Vars()) {
-      auto var = scope->Var(var_desc->Name());
+  // use  originProgramDesc and scope to init tensors
+  InitMemoryFromProgram(originProgramDesc, scope);
+  // perform fusion and print infos
+  FusionAndPrintInfos(optimize, can_add_split, program, originProgramDesc);
 
-      if (var_desc->Type() == framework::VARTYPE_TYPE_LOD_TENSOR) {
-        if (var_desc->Persistable() &&
-            var_desc->Type() != framework::VARTYPE_TYPE_FEED_MINIBATCH &&
-            var_desc->Type() != framework::VARTYPE_TYPE_FETCH_LIST) {
-          auto dim = var_desc->Tensor_desc().Dims();
-          auto tensor = var->GetMutable<framework::LoDTensor>();
-          tensor->Resize(framework::make_ddim(dim));
-        } else {
-          auto dim = var_desc->Tensor_desc().Dims();
-          PADDLE_MOBILE_ENFORCE(dim.size() > 0, "dim size is 0");
-          dim[0] = 1;
-          auto tensor = var->GetMutable<framework::LoDTensor>();
-          tensor->Resize(framework::make_ddim(dim));
-        }
-      } else {
-        // TODO(codeWorm): some.
-      }
-    }
-  }
+  paddle_mobile__framework__proto__program_desc__free_unpacked(c_program, NULL);
+  return program;
+}
 
-  if (optimize) {
-    framework::ProgramOptimize program_optimize;
-    program.optimizeProgram =
-        program_optimize.FusionOptimize(originProgramDesc, can_add_split);
-  }
-  if (optimize) {
-    program.optimizeProgram->Description("optimize: ");
-  } else {
-    originProgramDesc->Description("program: ");
-  }
+template <typename Dtype, Precision P>
+const framework::Program<Dtype, P> Loader<Dtype, P>::LoadCombinedMemory(
+    size_t read_size, const uint8_t *buf, size_t combined_params_len,
+    const uint8_t *combined_params_buf, bool optimize, bool quantification) {
+  bool can_add_split = false;
 
-  paddle_mobile__framework__proto__program_desc__free_unpacked(c_program, NULL);
+  PaddleMobile__Framework__Proto__ProgramDesc *c_program;
+  PADDLE_MOBILE_ENFORCE(buf != nullptr, "read from __model__ is null");
+
+  c_program = paddle_mobile__framework__proto__program_desc__unpack(
+      nullptr, read_size, buf);
+  //
+  PADDLE_MOBILE_ENFORCE(c_program != nullptr, "program is null");
+  //
+  DLOG << "n_ops: " << (*c_program->blocks)->n_ops;
+  //
+
+  auto originProgramDesc = std::make_shared<framework::ProgramDesc>(c_program);
+
+  framework::Program<Dtype, P> program;
+  program.combined = true;
+  program.originProgram = originProgramDesc;
+  program.quantification = quantification;
+  program.combined_params_len = combined_params_len;
+  program.combined_params_buf = combined_params_buf;
+
+  auto scope = std::make_shared<framework::Scope>();
+  program.scope = scope;
+  InitMemoryFromProgram(originProgramDesc, scope);
+  FusionAndPrintInfos(optimize, can_add_split, program, originProgramDesc);
+  paddle_mobile__framework__proto__program_desc__free_unpacked(c_program,
+                                                               nullptr);
   return program;
 }
 

src/io/loader.h

@@ -42,6 +42,11 @@ class Loader {
                                           bool optimize = false,
                                           bool quantification = false);
 
+  const framework::Program<Dtype, P> LoadCombinedMemory(
+      size_t model_len, const uint8_t *model_buf, size_t combined_params_len,
+      const uint8_t *combined_params_buf, bool optimize = false,
+      bool quantification = false);
+
  private:
   const framework::Program<Dtype, P> LoadProgram(const std::string &model_path,
                                                  bool optimize = false,

src/io/paddle_mobile.cpp

@@ -64,6 +64,32 @@ bool PaddleMobile<Dtype, P>::Load(const std::string &model_path,
   return true;
 }
 
+template <typename Dtype, Precision P>
+bool PaddleMobile<Dtype, P>::LoadCombinedMemory(
+    size_t model_len, const uint8_t *model_buf, size_t combined_params_len,
+    const uint8_t *combined_params_buf) {
+  int batch_size = 1;
+  bool optimise = true;
+  bool quantification = false;
+
+  if (loader_.get() == nullptr) {
+    loader_ = std::make_shared<Loader<Dtype, P>>();
+  } else {
+    LOG(kLOG_INFO) << "loader inited";
+  }
+
+  if (executor_.get() == nullptr) {
+    executor_ = std::make_shared<Executor<Dtype, P>>(
+        loader_->LoadCombinedMemory(model_len, model_buf, combined_params_len,
+                                    combined_params_buf, optimise,
+                                    quantification),
+        batch_size, optimise);
+  } else {
+    LOG(kLOG_INFO) << "executor inited";
+  }
+
+  return true;
+}
 template <typename Dtype, Precision P>
 std::shared_ptr<framework::Tensor> PaddleMobile<Dtype, P>::Predict(
     const framework::Tensor &t) {

src/io/paddle_mobile.h

@@ -66,6 +66,19 @@ class PaddleMobile {
   std::vector<Ptype> Predict(const std::vector<Ptype> &input,
                              const std::vector<int64_t> &dims);
 
+  /**
+   * 从内存加载model 以及 combinedparams的接口
+   *
+   * @param model_len model 文件的内存大小
+   * @param model_buf model文件的内存
+   * @param combined_params_len  params文件的内存大小
+   * @param combined_params_buf  params文件的内存
+   * @return
+   */
+  bool LoadCombinedMemory(size_t model_len, const uint8_t *model_buf,
+                          size_t combined_params_len,
+                          const uint8_t *combined_params_buf);
+
   void Clear();
 
   ~PaddleMobile();

test/CMakeLists.txt

@@ -117,6 +117,10 @@ else ()
     ADD_EXECUTABLE(test-load framework/test_load.cpp)
     target_link_libraries(test-load paddle-mobile)
 
+    # gen test log
+    ADD_EXECUTABLE(test-loadmemory framework/test_load_memory.cpp)
+    target_link_libraries(test-loadmemory paddle-mobile)
+
     ADD_EXECUTABLE(test-inference-api framework/test_inference_api.cpp)
     target_link_libraries(test-inference-api paddle-mobile)
 

test/framework/test_load_memory.cpp

+/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
+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 <string>
+
+#include "../test_helper.h"
+#include "../test_include.h"
+static size_t ReadBuffer(const char *file_name, uint8_t **out) {
+  FILE *fp;
+  fp = fopen(file_name, "rb");
+  PADDLE_MOBILE_ENFORCE(fp != nullptr, " %s open failed !", file_name);
+  fseek(fp, 0, SEEK_END);
+  auto size = static_cast<size_t>(ftell(fp));
+  rewind(fp);
+  DLOG << "model size: " << size;
+  *out = reinterpret_cast<uint8_t *>(malloc(size));
+  size_t cur_len = 0;
+  size_t nread;
+  while ((nread = fread(*out + cur_len, 1, size - cur_len, fp)) != 0) {
+    cur_len += nread;
+  }
+  fclose(fp);
+  return cur_len;
+}
+
+static char *Get_binary_data(std::string filename) {
+  FILE *file = fopen(filename.c_str(), "rb");
+  PADDLE_MOBILE_ENFORCE(file != nullptr, "can't open file: %s ",
+                        filename.c_str());
+  fseek(file, 0, SEEK_END);
+  int64_t size = ftell(file);
+  PADDLE_MOBILE_ENFORCE(size > 0, "size is too small");
+  rewind(file);
+  auto *data = new char[size];
+  size_t bytes_read = fread(data, 1, size, file);
+  PADDLE_MOBILE_ENFORCE(bytes_read == size,
+                        "read binary file bytes do not match with fseek");
+  fclose(file);
+  return data;
+}
+
+int main() {
+  paddle_mobile::PaddleMobile<paddle_mobile::CPU> paddle_mobile;
+  auto model_path = std::string(g_genet_combine) + "/model";
+  auto params_path = std::string(g_genet_combine) + "/params";
+  uint8_t *bufModel = nullptr;
+  size_t sizeBuf = ReadBuffer(model_path.c_str(), &bufModel);
+  uint8_t *bufParams = nullptr;
+
+  DLOG << "sizeBuf: " << sizeBuf;
+  size_t sizeParams = ReadBuffer(params_path.c_str(), &bufParams);
+  DLOG << "sizeParams: " << sizeParams;
+
+  paddle_mobile.LoadCombinedMemory(sizeBuf, bufModel, sizeParams, bufParams);
+  return 0;
+}