Created by: BryceLuminary

我使用 c++ 和 python 对同一个模型进行预测，c++ api 遇到了2个疑问： 1、请问 PaddleBuf 接口第二个参数的单位是字节吗？ 2、为什么同一个模型两种接口预测的输出个数不同？

代码如下，谢谢！

#include <gflags/gflags.h>
#include <glog/logging.h>
#include <memory>
#include <thread>  //NOLINT
#include "paddle/fluid/inference/paddle_inference_api.h"
#include "paddle/fluid/platform/enforce.h"

DEFINE_string(dirname, "", "Directory of the inference model.");
DEFINE_bool(use_gpu, false, "Whether use gpu.");

#define N 1
#define C 3
#define H 225
#define W 225

#define IN_SIZE (N * C * H * W)

namespace paddle {
namespace demo {

void Main(bool use_gpu) {
  //# 1. Create PaddlePredictor with a config.
  NativeConfig config;
  config.prog_file = "/home/fluid_models/model";
  config.param_file = "/home/fluid_models/params";
  config.use_gpu = use_gpu;
  config.fraction_of_gpu_memory = 0.15;
  config.device = 0;
  auto predictor =
      CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
  //# 2. Prepare input.
  float data[IN_SIZE];

  for (int i = 0; i < IN_SIZE; i++) {
    data[i] = 1.f;
  }

  std::cout << "size of data: " << sizeof(data) << std::endl;

  PaddleTensor tensor;
  tensor.shape = std::vector<int>({N, C, H, W});
  tensor.data = PaddleBuf(data, IN_SIZE);   // Question 1: Should IN_SIZE be count of bytes?
  tensor.dtype = PaddleDType::FLOAT32;

  // For simplicity, we set all the slots with the same data.
  std::vector<PaddleTensor> slots(1, tensor);

  //# 3. Run
  std::vector<PaddleTensor> outputs;


  CHECK(predictor->Run(slots, &outputs));

  //# 4. Get output.
  std::cout << "num of outputs: " << outputs.size() << std::endl;       // Question 2: The output of c++ is different from python
  for (int i = 0; i < outputs.size(); i++) {
    std::cout << "output " << i << " name " << outputs[i].name << std::endl;
    std::cout << "output " << i << " size " << outputs[i].data.length() << std::endl;
    std::cout << "output " << i << " dtype " << outputs[i].dtype << std::endl;
  }
  // The outputs' buffers are in CPU memory.
  for (size_t i = 0; i < outputs[0].data.length() / 4; i++) {
    std::cout << static_cast<float*>(outputs[0].data.data())[i] << std::endl;
  }
}


}  // namespace demo
}  // namespace paddle

int main(int argc, char** argv) {
  google::ParseCommandLineFlags(&argc, &argv, true);
  paddle::demo::Main(true /* use_gpu*/);
  return 0;
}