simplify inference api (#11104)

paddle/contrib/inference/paddle_inference_api.h

@@ -40,14 +40,23 @@ struct PaddleBuf {
 struct PaddleTensor {
   std::string name;  // variable name.
   std::vector<int> shape;
+  // TODO(Superjomn) for LoD support, add a vector<vector<int>> field if needed.
   PaddleBuf data;  // blob of data.
   PaddleDType dtype;
 };
 
+enum class PaddleEngineKind {
+  kNative = 0,  // Use the native Fluid facility.
+  // TODO(Superjomn) support following engines latter.
+  // kAnakin,             // Use Anakin for inference.
+  // kTensorRT,           // Use TensorRT for inference.
+  // kAutoMixedAnakin,    // Automatically mix Fluid with Anakin.
+  // kAutoMixedTensorRT,  // Automatically mix Fluid with TensorRT.
+};
+
 /*
  * A simple Inference API for Paddle. Currently this API can be used by
  * non-sequence scenerios.
- * TODO(Superjomn) Support another API for NLP-related usages.
  */
 class PaddlePredictor {
  public:
@@ -69,15 +78,6 @@ class PaddlePredictor {
   // Destroy the Predictor.
   virtual ~PaddlePredictor() {}
 
-  enum class EngineKind {
-    kNative = -1,  // Use the native Fluid facility.
-    // TODO(Superjomn) support latter.
-    // kAnakin,             // Use Anakin for inference.
-    // kTensorRT,           // Use TensorRT for inference.
-    // kAutoMixedAnakin,    // Automatically mix Fluid with Anakin.
-    // kAutoMixedTensorRT,  // Automatically mix Fluid with TensorRT.
-  };
-
   // The common configs for all the predictors.
   struct Config {
     std::string model_dir;      // path to the model directory.
@@ -86,18 +86,24 @@ class PaddlePredictor {
 };
 
 struct NativeConfig : public PaddlePredictor::Config {
+  // GPU related fields.
   bool use_gpu{false};
-  int device;
-  float fraction_of_gpu_memory;
+  int device{0};
+  float fraction_of_gpu_memory{-1.f};  // Negative to notify initialization.
+
   std::string prog_file;
   std::string param_file;
-  bool share_variables;
 };
 
-// A factory to help create difference predictor.
-template <
-    typename ConfigT,
-    PaddlePredictor::EngineKind engine = PaddlePredictor::EngineKind::kNative>
+// A factory to help create different predictors.
+//
+// FOR EXTENSION DEVELOPER:
+// Different predictors are designated by config type and engine kind. Similar
+// configs can be merged, but there shouldn't be a huge config containing
+// different fields for more than one kind of predictors.
+//
+// Similarly, each engine kind should map to a unique predictor implementation.
+template <typename ConfigT, PaddleEngineKind engine = PaddleEngineKind::kNative>
 std::unique_ptr<PaddlePredictor> CreatePaddlePredictor(const ConfigT& config);
 
 }  // namespace paddle

paddle/contrib/inference/paddle_inference_api_impl.cc

@@ -57,8 +57,7 @@ std::string num2str(T a) {
 bool NativePaddlePredictor::Init() {
   VLOG(3) << "Predictor::init()";
 
-  // TODO(panyx0718): Should CPU vs GPU device be decided by id?
-  if (config_.device >= 0) {
+  if (config_.use_gpu) {
     place_ = paddle::platform::CUDAPlace(config_.device);
   } else {
     place_ = paddle::platform::CPUPlace();
@@ -85,11 +84,13 @@ bool NativePaddlePredictor::Init() {
   }
   ctx_ = executor_->Prepare(*inference_program_, 0);
 
-  // Create variables
-  // TODO(panyx0718): Why need to test share_variables here?
-  if (config_.share_variables) {
+  // Create temporary variables first, so that the first batch do not need to
+  // create variables in the runtime. This is the logics of the old inference
+  // API.
+  // TODO(Superjomn) this should be modified when `Clone` is valid for
+  // multi-thread application.
   executor_->CreateVariables(*inference_program_, scope_.get(), 0);
-  }
+
   // Get the feed_target_names and fetch_target_names
   feed_target_names_ = inference_program_->GetFeedTargetNames();
   fetch_target_names_ = inference_program_->GetFetchTargetNames();
@@ -124,7 +125,7 @@ bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
                                 scope_.get(),
                                 &feed_targets,
                                 &fetch_targets,
-                                !config_.share_variables);
+                                false /* don't create variable eatch time */);
   if (!GetFetch(fetchs, output_data)) {
     LOG(ERROR) << "fail to get fetchs";
     return false;
@@ -242,11 +243,14 @@ bool NativePaddlePredictor::GetFetch(
 
 template <>
 std::unique_ptr<PaddlePredictor>
-CreatePaddlePredictor<NativeConfig, PaddlePredictor::EngineKind::kNative>(
+CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(
     const NativeConfig &config) {
   VLOG(3) << "create NativePaddlePredictor";
   if (config.use_gpu) {
     // 1. GPU memeroy
+    PADDLE_ENFORCE(
+        config.fraction_of_gpu_memory > 0.f,
+        "fraction_of_gpu_memory in the config should be set to range (0., 1.]");
     std::vector<std::string> flags;
     if (config.fraction_of_gpu_memory >= 0.0f ||
         config.fraction_of_gpu_memory <= 0.95f) {

paddle/contrib/inference/test_paddle_inference_api_impl.cc

@@ -47,7 +47,6 @@ NativeConfig GetConfig() {
   config.fraction_of_gpu_memory = 0.15;
   config.use_gpu = true;
   config.device = 0;
-  config.share_variables = true;
   return config;
 }