update predictor.register_output_hook: support paddle::Tensor

[yuanlehome]

PR types

New features

PR changes

Others

Description

upgrade predictor.register_output_hook api: support paddle::Tensor(C++), paddle.Tensor(Python)

之前register_output_hook打印中间tensor的功能，获取的是paddle_infer::Tensor，使用起来不方便，升级了下支持paddle::Tensor，c++端和python端接口均不变，使用register_output_hook接口。第一版参考 #47050 。

Python 示例:

def hook(op_name: str, output_name: str, tensor: paddle.Tensor):
        print(op_name, output_name, tensor)

predictor.register_output_hook(hook)

C++ 示例：

// 使用 paddle::Tensor 需要包含这个头文件
#include "paddle/extension.h"

void get_output_tensor(const std::string &op_type,
                       const std::string &tensor_name,
                       const paddle::Tensor &tensor) {
  if(tensor.dtype() != paddle::DataType::FLOAT32) return;
  auto cpu_tensor = tensor.copy_to(paddle::CPUPlace{}, true);
  // using TYPE = phi::dtype::float16;
  using TYPE = float;

  std::stringstream ss;

  // op type and tensor name
  ss << std::left << std::setw(20) << op_type << std::setw(40) << tensor_name;

  // tensor shape
  std::string shape_str;
  shape_str += "[" + std::to_string(cpu_tensor.shape()[0]);
  for (size_t i = 1; i < cpu_tensor.shape().size(); i++) {
    shape_str += "," + std::to_string(cpu_tensor.shape()[i]);
  }
  shape_str += "]";
  ss << std::setw(20) << shape_str;

  // tensor data mean and variance
  TYPE sum{0};
  for (size_t i = 0; i < cpu_tensor.numel(); i++) {
    sum += cpu_tensor.data<TYPE>()[i];
  }
  TYPE mean = sum / TYPE(cpu_tensor.numel());
  TYPE accum{0};
  for (size_t i = 0; i < cpu_tensor.numel(); i++) {
    accum += (cpu_tensor.data<TYPE>()[i] - mean) *
             (cpu_tensor.data<TYPE>()[i] - mean);
  }
  TYPE variance = accum / TYPE(cpu_tensor.numel());
  ss << std::setw(20) << mean << std::setw(20) << variance;

  std::cout << ss.str() << std::endl;
}

C++ 输出示例：

update: hook函数原本支持的paddle_infer::Tensor不再支持，以后统一使用paddle::Tensor。

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[jiweibo]

LGTM

[SCP-KAKA]

使用pip安装的paddlepaddle中没有这个接口吗？

[yuanlehome]

使用pip安装的paddlepaddle中没有这个接口吗？

有的。参考这个 https://github.com/PaddlePaddle/Paddle-Inference-Demo/blob/master/python/gpu/resnet50/infer_resnet.py#L65 调predictor.register_output_hook就可以了。