merge develop

Author: NHZlX

cmake/external/anakin.cmake

@@ -8,6 +8,7 @@ set(ANAKIN_INCLUDE "${ANAKIN_INSTALL_DIR}" CACHE STRING "root of Anakin header f
 set(ANAKIN_LIBRARY "${ANAKIN_INSTALL_DIR}" CACHE STRING "path of Anakin library")
 
 set(ANAKIN_COMPILE_EXTRA_FLAGS 
+    -Wno-error=unused-but-set-variable -Wno-unused-but-set-variable
     -Wno-error=unused-variable -Wno-unused-variable 
     -Wno-error=format-extra-args -Wno-format-extra-args
     -Wno-error=comment -Wno-comment 
@@ -19,7 +20,7 @@ set(ANAKIN_COMPILE_EXTRA_FLAGS
     -Wno-reorder 
     -Wno-error=cpp)
 
-set(ANAKIN_LIBRARY_URL "https://github.com/pangge/Anakin/releases/download/3.0/anakin_release_simple.tar.gz")
+set(ANAKIN_LIBRARY_URL "https://github.com/pangge/Anakin/releases/download/Version0.1.0/anakin.tar.gz")
 
 # A helper function used in Anakin, currently, to use it, one need to recursively include
 # nearly all the header files.
@@ -41,9 +42,9 @@ if (NOT EXISTS "${ANAKIN_INSTALL_DIR}")
     message(STATUS "Download Anakin library from ${ANAKIN_LIBRARY_URL}")
     execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_INSTALL_DIR}")
     execute_process(COMMAND bash -c "rm -rf ${ANAKIN_INSTALL_DIR}/*")
-    execute_process(COMMAND bash -c "cd ${ANAKIN_INSTALL_DIR}; wget -q ${ANAKIN_LIBRARY_URL}")
+    execute_process(COMMAND bash -c "cd ${ANAKIN_INSTALL_DIR}; wget --no-check-certificate -q ${ANAKIN_LIBRARY_URL}")
     execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_INSTALL_DIR}")
-    execute_process(COMMAND bash -c "cd ${ANAKIN_INSTALL_DIR}; tar xzf anakin_release_simple.tar.gz")
+    execute_process(COMMAND bash -c "cd ${ANAKIN_INSTALL_DIR}; tar xzf anakin.tar.gz")
 endif()
 
 if (WITH_ANAKIN)

paddle/fluid/inference/analysis/argument.h

@@ -23,6 +23,7 @@
 
 #pragma once
 
+#include <string>
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/inference/analysis/data_flow_graph.h"
 

paddle/fluid/inference/analysis/data_flow_graph.h

@@ -176,7 +176,7 @@ struct GraphTraits<DataFlowGraph> {
 // sub-graph is the inputs nodes and output nodes that doesn't inside the
 // sub-graph.
 std::pair<std::vector<Node *>, std::vector<Node *>>
-ExtractInputAndOutputOfSubGraph(std::vector<Node *> &graph);
+ExtractInputAndOutputOfSubGraph(std::vector<Node *> &graph);  // NOLINT
 
 }  // namespace analysis
 }  // namespace inference

paddle/fluid/inference/analysis/model_store_pass.cc

@@ -12,11 +12,13 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/inference/analysis/model_store_pass.h"
 #include <stdio.h>
 #include <stdlib.h>
+#include <string>
+
 #include "paddle/fluid/inference/analysis/analyzer.h"
 #include "paddle/fluid/inference/analysis/argument.h"
+#include "paddle/fluid/inference/analysis/model_store_pass.h"
 
 namespace paddle {
 namespace inference {

paddle/fluid/inference/analysis/model_store_pass.h

@@ -17,6 +17,8 @@
  * model in the disk, and that model can be reloaded for prediction.
  */
 
+#pragma once
+#include <string>
 #include "paddle/fluid/inference/analysis/pass.h"
 
 namespace paddle {

paddle/fluid/inference/api/CMakeLists.txt

@@ -19,6 +19,7 @@ endif(APPLE)
 
 
 set(inference_deps paddle_inference_api paddle_fluid_api)
+
 if(WITH_GPU AND TENSORRT_FOUND)
     set(inference_deps ${inference_deps} paddle_inference_tensorrt_subgraph_engine)
 endif()
@@ -63,6 +64,8 @@ endif()
 if (WITH_ANAKIN) # only needed in CI
     # Due to Anakin do not have official library releases and the versions of protobuf and cuda do not match Paddle's,
     # so anakin library will not be merged to our official inference library. To use anakin prediction API, one need to
+    # compile the libinference_anakin_api.a and compile with anakin.so.
+    fetch_include_recursively(${ANAKIN_INCLUDE})
     # compile the libinference_anakin_api.a and anakin.so.
     nv_library(inference_anakin_api SRCS api.cc api_anakin_engine.cc)
     nv_library(inference_anakin_api_shared SHARED SRCS api.cc api_anakin_engine.cc)
@@ -73,7 +76,7 @@ if (WITH_ANAKIN) # only needed in CI
     if (WITH_TESTING)
         cc_test(inference_anakin_test SRCS api_anakin_engine_tester.cc
                                   ARGS --model=${ANAKIN_INSTALL_DIR}/mobilenet_v2.anakin.bin
-                                  DEPS inference_anakin_api)
+                                  DEPS inference_anakin_api_shared)
         target_compile_options(inference_anakin_test BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
      endif(WITH_TESTING)
 endif()

paddle/fluid/inference/api/api_anakin_engine.cc

@@ -18,26 +18,36 @@
 
 namespace paddle {
 
-PaddleInferenceAnakinPredictor::PaddleInferenceAnakinPredictor(
+template <typename Target>
+PaddleInferenceAnakinPredictor<Target>::PaddleInferenceAnakinPredictor(
     const AnakinConfig &config) {
   CHECK(Init(config));
 }
 
-bool PaddleInferenceAnakinPredictor::Init(const AnakinConfig &config) {
+template <typename Target>
+bool PaddleInferenceAnakinPredictor<Target>::Init(const AnakinConfig &config) {
   if (!(graph_.load(config.model_file))) {
+    LOG(FATAL) << "fail to load graph from " << config.model_file;
     return false;
   }
-  graph_.ResetBatchSize("input_0", config.max_batch_size);
+  auto inputs = graph_.get_ins();
+  for (auto &input_str : inputs) {
+    graph_.ResetBatchSize(input_str, config.max_batch_size);
+  }
   // optimization for graph
   if (!(graph_.Optimize())) {
     return false;
   }
   // construct executer
-  executor_.init(graph_);
+  if (executor_p_ == nullptr) {
+    executor_p_ = new anakin::Net<Target, anakin::saber::AK_FLOAT,
+                                  anakin::Precision::FP32>(graph_, true);
+  }
   return true;
 }
 
-bool PaddleInferenceAnakinPredictor::Run(
+template <typename Target>
+bool PaddleInferenceAnakinPredictor<Target>::Run(
     const std::vector<PaddleTensor> &inputs,
     std::vector<PaddleTensor> *output_data, int batch_size) {
   for (const auto &input : inputs) {
@@ -46,7 +56,29 @@ bool PaddleInferenceAnakinPredictor::Run(
                  << "'s type is not float";
       return false;
     }
-    auto d_tensor_in_p = executor_.get_in(input.name);
+    auto d_tensor_in_p = executor_p_->get_in(input.name);
+    auto net_shape = d_tensor_in_p->valid_shape();
+    if (net_shape.size() != input.shape.size()) {
+      LOG(ERROR) << " input  " << input.name
+                 << "'s shape size should be equal to that of net";
+      return false;
+    }
+    int sum = 1;
+    for_each(input.shape.begin(), input.shape.end(), [&](int n) { sum *= n; });
+    if (sum > net_shape.count()) {
+      graph_.Reshape(input.name, input.shape);
+      delete executor_p_;
+      executor_p_ = new anakin::Net<Target, anakin::saber::AK_FLOAT,
+                                    anakin::Precision::FP32>(graph_, true);
+      d_tensor_in_p = executor_p_->get_in(input.name);
+    }
+
+    anakin::saber::Shape tmp_shape;
+    for (auto s : input.shape) {
+      tmp_shape.push_back(s);
+    }
+    d_tensor_in_p->reshape(tmp_shape);
+
     float *d_data_p = d_tensor_in_p->mutable_data();
     if (cudaMemcpy(d_data_p, static_cast<float *>(input.data.data()),
                    d_tensor_in_p->valid_size() * sizeof(float),
@@ -56,16 +88,17 @@ bool PaddleInferenceAnakinPredictor::Run(
     }
     cudaStreamSynchronize(NULL);
   }
-
-  executor_.prediction();
+  cudaDeviceSynchronize();
+  executor_p_->prediction();
+  cudaDeviceSynchronize();
 
   if (output_data->empty()) {
     LOG(ERROR) << "At least one output should be set with tensors' names.";
     return false;
   }
   for (auto &output : *output_data) {
-    auto *tensor = executor_.get_out(output.name);
-    output.shape = tensor->shape();
+    auto *tensor = executor_p_->get_out(output.name);
+    output.shape = tensor->valid_shape();
     if (output.data.length() < tensor->valid_size() * sizeof(float)) {
       output.data.Resize(tensor->valid_size() * sizeof(float));
     }
@@ -81,19 +114,23 @@ bool PaddleInferenceAnakinPredictor::Run(
   return true;
 }
 
-anakin::Net<anakin::NV, anakin::saber::AK_FLOAT, anakin::Precision::FP32>
-    &PaddleInferenceAnakinPredictor::get_executer() {
-  return executor_;
+template <typename Target>
+anakin::Net<Target, anakin::saber::AK_FLOAT, anakin::Precision::FP32>
+    &PaddleInferenceAnakinPredictor<Target>::get_executer() {
+  return *executor_p_;
 }
 
 // the cloned new Predictor of anakin share the same net weights from original
 // Predictor
-std::unique_ptr<PaddlePredictor> PaddleInferenceAnakinPredictor::Clone() {
+template <typename Target>
+std::unique_ptr<PaddlePredictor>
+PaddleInferenceAnakinPredictor<Target>::Clone() {
   VLOG(3) << "Anakin Predictor::clone";
-  std::unique_ptr<PaddlePredictor> cls(new PaddleInferenceAnakinPredictor());
+  std::unique_ptr<PaddlePredictor> cls(
+      new PaddleInferenceAnakinPredictor<Target>());
   // construct executer from other graph
   auto anakin_predictor_p =
-      dynamic_cast<PaddleInferenceAnakinPredictor *>(cls.get());
+      dynamic_cast<PaddleInferenceAnakinPredictor<Target> *>(cls.get());
   if (!anakin_predictor_p) {
     LOG(ERROR) << "fail to call Init";
     return nullptr;
@@ -103,14 +140,28 @@ std::unique_ptr<PaddlePredictor> PaddleInferenceAnakinPredictor::Clone() {
   return std::move(cls);
 }
 
+template class PaddleInferenceAnakinPredictor<anakin::NV>;
+template class PaddleInferenceAnakinPredictor<anakin::X86>;
+
 // A factory to help create difference predictor.
 template <>
 std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
     AnakinConfig, PaddleEngineKind::kAnakin>(const AnakinConfig &config) {
   VLOG(3) << "Anakin Predictor create.";
-  std::unique_ptr<PaddlePredictor> x(
-      new PaddleInferenceAnakinPredictor(config));
-  return x;
-}
+  if (config.target_type == AnakinConfig::NVGPU) {
+    VLOG(3) << "Anakin Predictor create on [ NVIDIA GPU ].";
+    std::unique_ptr<PaddlePredictor> x(
+        new PaddleInferenceAnakinPredictor<anakin::NV>(config));
+    return x;
+  } else if (config.target_type == AnakinConfig::X86) {
+    VLOG(3) << "Anakin Predictor create on [ Intel X86 ].";
+    std::unique_ptr<PaddlePredictor> x(
+        new PaddleInferenceAnakinPredictor<anakin::X86>(config));
+    return x;
+  } else {
+    VLOG(3) << "Anakin Predictor create on unknown platform.";
+    return nullptr;
+  }
+};
 
 }  // namespace paddle

paddle/fluid/inference/api/api_anakin_engine.h

@@ -20,14 +20,16 @@ limitations under the License. */
 #pragma once
 
 #include <vector>
-#include "paddle/fluid/inference/api/paddle_inference_api.h"
 
-// from anakin
 #include "framework/core/net/net.h"
+#include "framework/graph/graph.h"
+#include "paddle/fluid/inference/api/paddle_inference_api.h"
+#include "saber/core/shape.h"
 #include "saber/saber_types.h"
 
 namespace paddle {
 
+template <typename Target>
 class PaddleInferenceAnakinPredictor : public PaddlePredictor {
  public:
   PaddleInferenceAnakinPredictor() {}
@@ -42,19 +44,21 @@ class PaddleInferenceAnakinPredictor : public PaddlePredictor {
 
   std::unique_ptr<PaddlePredictor> Clone() override;
 
-  anakin::Net<anakin::NV, anakin::saber::AK_FLOAT, anakin::Precision::FP32>&
+  anakin::Net<Target, anakin::saber::AK_FLOAT, anakin::Precision::FP32>&
   get_executer();
 
-  ~PaddleInferenceAnakinPredictor() override{};
+  ~PaddleInferenceAnakinPredictor() override {
+    delete executor_p_;
+    executor_p_ = nullptr;
+  };
 
  private:
   bool Init(const AnakinConfig& config);
 
-  anakin::graph::Graph<anakin::NV, anakin::saber::AK_FLOAT,
-                       anakin::Precision::FP32>
+  anakin::graph::Graph<Target, anakin::saber::AK_FLOAT, anakin::Precision::FP32>
       graph_;
-  anakin::Net<anakin::NV, anakin::saber::AK_FLOAT, anakin::Precision::FP32>
-      executor_;
+  anakin::Net<Target, anakin::saber::AK_FLOAT, anakin::Precision::FP32>*
+      executor_p_{nullptr};
   AnakinConfig config_;
 };
 

paddle/fluid/inference/api/api_anakin_engine_tester.cc

@@ -12,18 +12,20 @@ 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 <gflags/gflags.h>
 #include <glog/logging.h>
 #include <gtest/gtest.h>
 
+#include "gflags/gflags.h"
 #include "paddle/fluid/inference/api/paddle_inference_api.h"
 
-DEFINE_string(model, "", "Directory of the inference model.");
+DEFINE_string(model, "", "Directory of the inference model(mobile_v2).");
 
 namespace paddle {
 
 AnakinConfig GetConfig() {
   AnakinConfig config;
+  // using AnakinConfig::X86 if you need to use cpu to do inference
+  config.target_type = AnakinConfig::NVGPU;
   config.model_file = FLAGS_model;
   config.device = 0;
   config.max_batch_size = 1;
@@ -36,30 +38,27 @@ TEST(inference, anakin) {
       CreatePaddlePredictor<AnakinConfig, PaddleEngineKind::kAnakin>(config);
 
   float data[1 * 3 * 224 * 224] = {1.0f};
-
   PaddleTensor tensor;
   tensor.name = "input_0";
   tensor.shape = std::vector<int>({1, 3, 224, 224});
   tensor.data = PaddleBuf(data, sizeof(data));
   tensor.dtype = PaddleDType::FLOAT32;
 
   // For simplicity, we set all the slots with the same data.
-  std::vector<PaddleTensor> paddle_tensor_feeds;
-  paddle_tensor_feeds.emplace_back(std::move(tensor));
+  std::vector<PaddleTensor> paddle_tensor_feeds(1, tensor);
 
   PaddleTensor tensor_out;
   tensor_out.name = "prob_out";
-  tensor_out.shape = std::vector<int>({1000, 1});
+  tensor_out.shape = std::vector<int>({});
   tensor_out.data = PaddleBuf();
   tensor_out.dtype = PaddleDType::FLOAT32;
 
-  std::vector<PaddleTensor> outputs;
-  outputs.emplace_back(std::move(tensor_out));
+  std::vector<PaddleTensor> outputs(1, tensor_out);
 
   ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs));
 
   float* data_o = static_cast<float*>(outputs[0].data.data());
-  for (size_t j = 0; j < 1000; ++j) {
+  for (size_t j = 0; j < outputs[0].data.length(); ++j) {
     LOG(INFO) << "output[" << j << "]: " << data_o[j];
   }
 }

paddle/fluid/inference/api/demo_ci/vis_demo.cc

@@ -20,8 +20,8 @@ limitations under the License. */
 #include <glog/logging.h>  // use glog instead of PADDLE_ENFORCE to avoid importing other paddle header files.
 #include <fstream>
 #include <iostream>
+#include "paddle/fluid/inference/demo_ci/utils.h"
 #include "paddle/fluid/platform/enforce.h"
-#include "utils.h"
 
 #ifdef PADDLE_WITH_CUDA
 DECLARE_double(fraction_of_gpu_memory_to_use);