Add TRT deploy (#520)

* Add TRT deploy * code clean * doc refine

Add TRT deploy (#520)
* Add TRT deploy * code clean * doc refine
1a54970a · Bai Yifan · GitHub · 9f43bbcc · 1a54970a · 1a54970a
10 changed file
--- a/README.md
+++ b/README.md
@@ -237,6 +237,7 @@ pip install paddleslim==1.2.0 -i https://pypi.tuna.tsinghua.edu.cn/simple
    - [SlimFaceNet](demo/slimfacenet/README.md)
    - [OCR模型压缩(基于PaddleOCR)](demo/ocr/README.md)
    - [检测模型压缩(基于PaddleDetection)](demo/detection/README.md)
+- [TensorRT部署](demo/quant/deploy/TensorRT): 介绍如何使用TensorRT部署PaddleSlim量化得到的模型。
 ## 部分压缩策略效果

--- a/README_en.md
+++ b/README_en.md
@@ -90,6 +90,7 @@ pip install paddleslim==1.2.0 -i https://pypi.tuna.tsinghua.edu.cn/simple
 - [PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection/tree/master/slim): Introduce how to use PaddleSlim in PaddleDetection library.
 - [PaddleSeg](https://github.com/PaddlePaddle/PaddleSeg/tree/develop/slim): Introduce how to use PaddleSlim in PaddleSeg library.
 - [PaddleLite](https://paddlepaddle.github.io/Paddle-Lite/): How to use PaddleLite to deploy models generated by PaddleSlim.
+- [TensorRT Deploy](demo/quant/deploy/TensorRT): How to use TensorRT to deploy models generated by PaddleSlim.
 ## Performance

--- a/demo/quant/deploy/TensorRT/CMakeLists.txt
+++ b/demo/quant/deploy/TensorRT/CMakeLists.txt
+cmake_minimum_required(VERSION 3.0)
+project(inference_test CXX C)
+option(WITH_MKL        "Compile demo with MKL/OpenBlas support, default use MKL."       OFF)
+option(WITH_GPU        "Compile demo with GPU/CPU, default use CPU."                    OFF)
+option(WITH_STATIC_LIB "Compile demo with static/shared library, default use static."   OFF)
+option(USE_TENSORRT "Compile demo with TensorRT."   OFF)
+if(NOT DEFINED PADDLE_LIB)
+  message(FATAL_ERROR "please set PADDLE_LIB with -DPADDLE_LIB=/path/paddle/lib")
+endif()
+# check file system
+file(READ "/etc/issue" ETC_ISSUE)
+string(REGEX MATCH "Debian|Ubuntu|CentOS" DIST ${ETC_ISSUE})
+if(DIST STREQUAL "Debian")
+  message(STATUS ">>>> Found Debian <<<<")
+elseif(DIST STREQUAL "Ubuntu")
+  message(STATUS ">>>> Found Ubuntu <<<<")
+elseif(DIST STREQUAL "CentOS")
+  message(STATUS ">>>> Found CentOS <<<<")
+else()
+  message(STATUS ">>>> Found unknown distribution <<<<")
+endif()
+include_directories("${PADDLE_LIB}/")
+set(PADDLE_LIB_THIRD_PARTY_PATH "${PADDLE_LIB}/third_party/install/")
+include_directories("${PADDLE_LIB_THIRD_PARTY_PATH}protobuf/include")
+include_directories("${PADDLE_LIB_THIRD_PARTY_PATH}glog/include")
+include_directories("${PADDLE_LIB_THIRD_PARTY_PATH}gflags/include")
+include_directories("${PADDLE_LIB_THIRD_PARTY_PATH}xxhash/include")
+include_directories("${PADDLE_LIB_THIRD_PARTY_PATH}zlib/include")
+include_directories("${PADDLE_LIB}/third_party/boost")
+include_directories("${PADDLE_LIB}/third_party/eigen3")
+include_directories("${PADDLE_LIB}/paddle/include")
+include_directories("${PADDLE_LIB}/paddle")
+link_directories("${PADDLE_LIB_THIRD_PARTY_PATH}zlib/lib")
+link_directories("${PADDLE_LIB_THIRD_PARTY_PATH}protobuf/lib")
+link_directories("${PADDLE_LIB_THIRD_PARTY_PATH}glog/lib")
+link_directories("${PADDLE_LIB_THIRD_PARTY_PATH}gflags/lib")
+link_directories("${PADDLE_LIB_THIRD_PARTY_PATH}xxhash/lib")
+link_directories("${PADDLE_LIB}/paddle/lib")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -O0 -std=c++11")
+message("flags" ${CMAKE_CXX_FLAGS})
+if (USE_TENSORRT AND WITH_GPU)
+  message("=====> TENSORRT_INCLUDE_DIR is ${TENSORRT_INCLUDE_DIR}")
+  message("=====> TENSORRT_LIB_DIR is ${TENSORRT_LIB_DIR}")
+  include_directories("${TENSORRT_INCLUDE_DIR}")
+  link_directories("${TENSORRT_LIB_DIR}")
+endif()
+if(WITH_MKL)
+  set(MATH_LIB_PATH "${PADDLE_LIB_THIRD_PARTY_PATH}mklml")
+  include_directories("${MATH_LIB_PATH}/include")
+  set(MATH_LIB ${MATH_LIB_PATH}/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX}
+                 ${MATH_LIB_PATH}/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX})
+  set(MKLDNN_PATH "${PADDLE_LIB_THIRD_PARTY_PATH}mkldnn")
+  if(EXISTS ${MKLDNN_PATH})
+    include_directories("${MKLDNN_PATH}/include")
+    set(MKLDNN_LIB ${MKLDNN_PATH}/lib/libmkldnn.so.0)
+  endif()
+else()
+  set(MATH_LIB ${PADDLE_LIB_THIRD_PARTY_PATH}openblas/lib/libopenblas${CMAKE_STATIC_LIBRARY_SUFFIX})
+endif()
+# Note: libpaddle_inference_api.so/a must put before libpaddle_fluid.so/a
+if(WITH_STATIC_LIB)
+  set(DEPS ${PADDLE_LIB}/paddle/lib/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX})
+else()
+  set(DEPS ${PADDLE_LIB}/paddle/lib/libpaddle_fluid${CMAKE_SHARED_LIBRARY_SUFFIX})
+endif()
+set(EXTERNAL_LIB "-lrt -ldl -lpthread -lprotobuf")
+set(DEPS ${DEPS}
+  ${MATH_LIB} ${MKLDNN_LIB}
+  glog gflags protobuf z xxhash
+  ${EXTERNAL_LIB})
+if(WITH_GPU)
+    if (USE_TENSORRT)
+      set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/libnvinfer${CMAKE_SHARED_LIBRARY_SUFFIX})
+      set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/libnvinfer_plugin${CMAKE_SHARED_LIBRARY_SUFFIX})
+    endif()
+    set(DEPS ${DEPS} $ENV{CUDA_LIB})
+endif()
+set(TestFiles "trt_clas.cc";
+              "trt_gen_calib_table_test.cc";
+              "test_acc.cc";)
+foreach(testsourcefile ${TestFiles})
+  message("====> ${testsourcefile} will be compiled")
+  # add executable for all test files
+  string(REPLACE ".cc" "" testname ${testsourcefile})
+  add_executable(${testname} ${testsourcefile})
+  # link libs
+  target_link_libraries(${testname} ${DEPS})
+endforeach()
--- a/demo/quant/deploy/TensorRT/README.md
+++ b/demo/quant/deploy/TensorRT/README.md
+# PaddleSlim量化模型的TensorRT预测
+本教程将介绍使用TensortRT部署PaddleSlim量化得到的模型的详细步骤。
+## 1. 准备环境
+* 有2种方式获取Paddle预测库，下面进行详细介绍。
+### 1.1 直接下载安装
+* [Paddle预测库官网](https://www.paddlepaddle.org.cn/documentation/docs/zh/advanced_guide/inference_deployment/inference/build_and_install_lib_cn.html)上提供了不同cuda版本的Linux预测库，可以在官网查看并选择带有TensorRT的预测库版本。
+* 下载之后使用下面的方法解压。
+```
+tar -xf fluid_inference.tgz
+```
+最终会在当前的文件夹中生成`fluid_inference/`的子文件夹。
+### 1.2 预测库源码编译
+* 如果希望获取最新预测库特性，可以从Paddle github上克隆最新代码，源码编译预测库。
+* 可以参考[Paddle预测库官网](https://www.paddlepaddle.org.cn/documentation/docs/zh/advanced_guide/inference_deployment/inference/build_and_install_lib_cn.html)的说明，从github上获取Paddle代码，然后进行编译，生成最新的预测库。使用git获取代码方法如下。
+```shell
+git clone https://github.com/PaddlePaddle/Paddle.git
+```
+* 在[Nvidia官网](https://developer.nvidia.com/TensorRT)下载TensorRT并解压, 本示例以TensorRT 6.0为例。
+* 进入Paddle目录后，编译方法如下。
+```shell
+rm -rf build
+mkdir build
+cd build
+cmake  .. \
+    -DWITH_MKL=ON \
+    -DWITH_MKLDNN=ON  \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DWITH_INFERENCE_API_TEST=OFF \
+    -DTENSORRT_ROOT=TensorRT-6.0.1.5 \
+    -DFLUID_INFERENCE_INSTALL_DIR=LIB_ROOT \
+    -DON_INFER=ON \
+    -DWITH_PYTHON=ON
+make -j
+make inference_lib_dist
+```
+其中`DFLUID_INFERENCE_INSTALL_DIR`代表编译完成后预测库生成的地址，`DTENSORRT_ROOT`代表下载解压后的TensorRT路径。
+更多编译参数选项可以参考Paddle C++预测库官网：[https://www.paddlepaddle.org.cn/documentation/docs/zh/advanced_guide/inference_deployment/inference/build_and_install_lib_cn.html](https://www.paddlepaddle.org.cn/documentation/docs/zh/advanced_guide/inference_deployment/inference/build_and_install_lib_cn.html)。
+* 编译完成之后，可以在`LIB_ROOT`路径下看到生成了以下文件及文件夹。
+```
+LIB_ROOT/
+|-- CMakeCache.txt
+|-- paddle
+|-- third_party
+|-- version.txt
+```
+其中`paddle`就是之后进行TensorRT预测时所需的Paddle库，`version.txt`中包含当前预测库的版本信息。
+## 2 开始运行
+### 2.1 将模型导出为inference model
+* 可以参考[量化训练教程](https://paddleslim.readthedocs.io/zh_CN/latest/quick_start/quant_aware_tutorial.html#id9)，在训练完成后导出inference model。
+```
+inference/
+|-- model
+|-- params
+```
+### 2.2 编译TensorRT预测demo
+* 编译命令如下，其中Paddle, TensorRT地址需要换成自己机器上的实际地址。
+```shell
+sh tools/build.sh
+```
+具体地，`tools/build.sh`中内容如下。
+```shell
+PADDLE_LIB_PATH=trt_inference # change to your path
+USE_GPU=ON
+USE_MKL=ON
+USE_TRT=ON
+TENSORRT_INCLUDE_DIR=TensorRT-6.0.1.5/include # change to your path
+TENSORRT_LIB_DIR=TensorRT-6.0.1.5/lib # change to your path
+if [ $USE_GPU -eq ON ]; then
+  export CUDA_LIB=`find /usr/local -name libcudart.so`
+fi
+BUILD=build
+mkdir -p $BUILD
+cd $BUILD
+cmake .. \
+      -DPADDLE_LIB=${PADDLE_LIB_PATH} \
+      -DWITH_GPU=${USE_GPU} \
+      -DWITH_MKL=${USE_MKL} \
+      -DCUDA_LIB=${CUDA_LIB} \
+      -DUSE_TENSORRT=${USE_TRT} \
+      -DTENSORRT_INCLUDE_DIR=${TENSORRT_INCLUDE_DIR} \
+      -DTENSORRT_LIB_DIR=${TENSORRT_LIB_DIR}
+make -j4
+```
+`PADDLE_LIB_PATH`为下载(`fluid_inference`文件夹)或者编译生成的Paddle预测库地址(`build/fluid_inference_install_dir`文件夹)；`TENSORRT_INCLUDE_DIR`和`TENSORRT_LIB_DIR`分别代表TensorRT的include和lib目录路径。
+* 编译完成之后，会在`build`文件夹下生成可执行文件。
+### 2.3 数据预处理转化
+在精度和性能预测中，需要先对数据进行二进制转化。运行脚本如下可转化完整ILSVRC2012 val数据集。使用`--local`可以转化用户自己的数据。在Paddle所在目录运行下面的脚本。脚本在官网位置为[full_ILSVRC2012_val_preprocess.py](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/api/full_ILSVRC2012_val_preprocess.py)
+```
+python Paddle/paddle/fluid/inference/tests/api/full_ILSVRC2012_val_preprocess.py --local --data_dir=/PATH/TO/USER/DATASET/  --output_file=/PATH/TO/SAVE/BINARY/FILE
+```
+可选参数：
+- 不设置任何参数。脚本将下载 ILSVRC2012_img_val数据集，并转化为二进制文件。
+- **local:** 设置便为true，表示用户将提供自己的数据
+- **data_dir:** 用户自己的数据目录
+- **label_list:** 图片路径-图片类别列表文件，类似于`val_list.txt`
+- **output_file:** 生成的binary文件路径。
+- **data_dim:** 预处理图片的长和宽。默认值 224。
+用户自己的数据集目录结构应该如下
+```
+imagenet_user
+├── val
+│   ├── ILSVRC2012_val_00000001.jpg
+│   ├── ILSVRC2012_val_00000002.jpg
+|   |── ...
+└── val_list.txt
+```
+其中，val_list.txt 内容应该如下：
+```
+val/ILSVRC2012_val_00000001.jpg 0
+val/ILSVRC2012_val_00000002.jpg 0
+```
+注意：
+- 为什么将数据集转化为二进制文件？因为paddle中的数据预处理（resize, crop等）都使用pythong.Image模块进行，训练出的模型也是基于Python预处理的图片，但是我们发现Python测试性能开销很大，导致预测性能下降。为了获得良好性能，在量化模型预测阶段，我们需要使用C++测试，而C++只支持Open-CV等库，Paddle不建议使用外部库，因此我们使用Python将图片预处理然后放入二进制文件，再在C++测试中读出。用户根据自己的需要，可以更改C++测试以直接读数据并预处理，精度不会有太大下降。
+### 2.4 部署预测
+### 运行demo
+* 执行以下命令，完成一个分类模型的TensorRT预测。
+```shell
+sh tools/run.sh
+```
+其中`MODEL_DIR`和`DATA_FILE`分别代表模型文件和数据文件, 需要在预测时替换为自己实际要用的地址。
+可以看到类似下面的预测结果：
+```shell
+I1123 11:30:49.160024 10999 trt_clas.cc:103] finish prediction
+I1123 11:30:49.160050 10999 trt_clas.cc:136] pred image class is : 65, ground truth label is : 65
+```
+* 修改`tools/run.sh`中的repeat_times大于1，通过多次预测取平均完成对一个模型的TensorRT速度评测。
+```shell
+sh tools/run.sh
+```
+可以看到类似下面的评测结果：
+```shell
+I1123 11:40:30.936796 11681 trt_clas.cc:83] finish warm up 10 times
+I1123 11:40:30.947906 11681 trt_clas.cc:101] total predict cost is : 11.042 ms, repeat 10 times
+I1123 11:40:30.947947 11681 trt_clas.cc:102] average predict cost is : 1.1042 ms
+```
+* 执行以下命令，完成对一个模型的TensorRT精度评测。
+```shell
+sh tools/test_acc.sh
+```
+同上，在预测时需要将其中路径替换为自己实际要用的地址。
+可以看到类似下面的评测结果：
+```shell
+I1123 11:23:11.856046 10913 test_acc.cc:64] 5000
+I1123 11:23:50.318663 10913 test_acc.cc:64] 10000
+I1123 11:24:28.793603 10913 test_acc.cc:64] 15000
+I1123 11:25:07.277580 10913 test_acc.cc:64] 20000
+I1123 11:25:45.698241 10913 test_acc.cc:64] 25000
+I1123 11:26:24.195798 10913 test_acc.cc:64] 30000
+I1123 11:27:02.625052 10913 test_acc.cc:64] 35000
+I1123 11:27:41.178545 10913 test_acc.cc:64] 40000
+I1123 11:28:19.798691 10913 test_acc.cc:64] 45000
+I1123 11:28:58.457620 10913 test_acc.cc:107] final result:
+I1123 11:28:58.457688 10913 test_acc.cc:108] top1 acc:0.70664
+I1123 11:28:58.457712 10913 test_acc.cc:109] top5 acc:0.89494
+```
+## 3 Benchmark
+GPU: NVIDIA® Tesla® P4
+数据集： ImageNet-2012
+预测引擎： Paddle-TensorRT
+|    模型     | FP32精度(Top1/Top5) | INT8精度(Top1/Top5) | FP32预测时延(ms) | INT8预测时延(ms) | 量化加速比 |
+| :---------: | :-----------------: | :-----------------: | :----------: | :----------: | :--------: |
+| MobileNetV1 |    71.00%/89.69%    |    70.66%/89.27%    |    1.083     |    0.568     |   47.55%   |
+| MobileNetV2 |    72.16%/90.65%    |    71.09%/90.16%    |    1.821     |    0.980     |   46.19%   |
+|  ResNet50   |    76.50%/93.00%    |    76.27%/92.95%    |    4.960     |    2.014     |   59.39%   |
--- a/demo/quant/deploy/TensorRT/test_acc.cc
+++ b/demo/quant/deploy/TensorRT/test_acc.cc
+#include <gflags/gflags.h>
+#include <glog/logging.h>
+#include <algorithm>  // std::sort
+#include <numeric>  // std::iota
+#include <vector>
+#include <numeric>
+#include <iostream>
+#include <fstream>
+#include <sstream>
+#include <memory>
+#include <chrono>
+#include "paddle/include/paddle_inference_api.h"
+namespace paddle {
+using paddle::AnalysisConfig;
+DEFINE_string(model_dir, "resnet50_quant", "Directory of the inference model.");
+DEFINE_string(data_dir, "imagenet-eval-binary", "Directory of the data.");
+DEFINE_bool(int8, true, "use int8 or not");
+using Time = decltype(std::chrono::high_resolution_clock::now());
+Time time() { return std::chrono::high_resolution_clock::now(); };
+double time_diff(Time t1, Time t2) {
+  typedef std::chrono::microseconds ms;
+  auto diff = t2 - t1;
+  ms counter = std::chrono::duration_cast<ms>(diff);
+  return counter.count() / 1000.0;
+}
+void PrepareTRTConfig(AnalysisConfig *config, int batch_size, int id = 0) {
+  int min_subgraph_size = 3;
+  config->SetModel(FLAGS_model_dir + "/model", FLAGS_model_dir + "/params");
+  config->EnableUseGpu(1500, id);
+  // We use ZeroCopyTensor here, so we set config->SwitchUseFeedFetchOps(false)
+  config->SwitchUseFeedFetchOps(false);
+  config->SwitchIrDebug(true);
+  if (FLAGS_int8)
+      config->EnableTensorRtEngine(1 << 30, batch_size, min_subgraph_size, AnalysisConfig::Precision::kInt8, false, false);
+  else
+      config->EnableTensorRtEngine(1 << 30, batch_size, min_subgraph_size, AnalysisConfig::Precision::kFloat32, false, false);
+}
+bool test_map_cnn(int batch_size, int repeat) {
+  AnalysisConfig config;
+  PrepareTRTConfig(&config, batch_size);
+  auto predictor = CreatePaddlePredictor(config);
+  int channels = 3;
+  int height = 224;
+  int width = 224;
+  int input_num = channels * height * width * batch_size;
+  // prepare inputs
+  float *input = new float[input_num];
+  memset(input, 0, input_num * sizeof(float));
+  float test_num = 0;
+  float top1_num = 0;
+  float top5_num = 0;
+  std::vector<int> index(1000);
+  for (size_t ind = 0; ind < 50000; ind++) {
+    if(ind % 5000 == 0)
+    LOG(INFO) << ind;
+    std::ifstream fs(FLAGS_data_dir + "/" +  std::to_string(ind)  + ".data", std::ifstream::binary);
+    if (!fs.is_open()) {
+      LOG(FATAL) << "open input file fail.";
+    }    
+    auto input_data_tmp = input;
+    for (int i = 0; i < input_num; ++i) {
+      fs.read(reinterpret_cast<char*>(input_data_tmp), sizeof(*input_data_tmp));
+      input_data_tmp++;
+    }
+    int label = 0;
+    fs.read(reinterpret_cast<char*>(&label), sizeof(label));
+    fs.close();
+    auto input_names = predictor->GetInputNames();
+    auto input_t = predictor->GetInputTensor(input_names[0]);
+    input_t->Reshape({batch_size, channels, height, width});
+    input_t->copy_from_cpu(input);
+    CHECK(predictor->ZeroCopyRun());
+    std::vector<float> out_data;
+    auto output_names = predictor->GetOutputNames();
+    auto output_t = predictor->GetOutputTensor(output_names[0]);
+    std::vector<int> output_shape = output_t->shape();
+    int out_num = std::accumulate(output_shape.begin(), output_shape.end(), 1, std::multiplies<int>());
+    out_data.resize(out_num);
+    output_t->copy_to_cpu(out_data.data());
+    std::iota(index.begin(), index.end(), 0);
+    std::sort(index.begin(), index.end(), [out_data](size_t i1, size_t i2) {
+      return out_data[i1] > out_data[i2];
+    });
+    test_num++;
+    if (label == index[0]) {
+      top1_num++;
+    }
+    for (int i = 0; i < 5; i++) {
+      if (label == index[i]) {
+        top5_num++;
+      }
+    }
+  }
+  LOG(INFO) << "final result:";
+  LOG(INFO) << "top1 acc:" << top1_num / test_num;
+  LOG(INFO) << "top5 acc:" << top5_num / test_num;
+  return true;
+}
+}  // namespace paddle
+int main(int argc,char* argv[]) {
+     google::ParseCommandLineFlags(&argc, &argv, false);
+  for (int i = 0; i < 1; i++) {
+    paddle::test_map_cnn(1 << i, 1);
+  }
+  return 0;
+}
--- a/demo/quant/deploy/TensorRT/tools/build.sh
+++ b/demo/quant/deploy/TensorRT/tools/build.sh
+#!/bin/bash
+PADDLE_LIB_PATH=trt_inference # change to your path
+USE_GPU=ON
+USE_MKL=ON
+USE_TRT=ON
+TENSORRT_INCLUDE_DIR=TensorRT-6.0.1.5/include # change to your path
+TENSORRT_LIB_DIR=TensorRT-6.0.1.5/lib # change to your path
+if [ $USE_GPU -eq ON ]; then
+  export CUDA_LIB=`find /usr/local -name libcudart.so`
+fi
+rm -rf build
+BUILD=build
+mkdir -p $BUILD
+cd $BUILD
+cmake .. \
+      -DPADDLE_LIB=${PADDLE_LIB_PATH} \
+      -DWITH_GPU=${USE_GPU} \
+      -DWITH_MKL=${USE_MKL} \
+      -DCUDA_LIB=${CUDA_LIB} \
+      -DUSE_TENSORRT=${USE_TRT} \
+      -DTENSORRT_INCLUDE_DIR=${TENSORRT_INCLUDE_DIR} \
+      -DTENSORRT_LIB_DIR=${TENSORRT_LIB_DIR}
+make -j4
--- a/demo/quant/deploy/TensorRT/tools/run.sh
+++ b/demo/quant/deploy/TensorRT/tools/run.sh
+#!/bin/bash
+project_path=$(cd "$(dirname "$0")";pwd)
+echo "${project_path}"
+unset GREP_OPTIONS;
+if [ ! -d "./build" ];then 
+    echo -e "\033[33m run failed! \033[0m";
+    echo -e "\033[33m you should build first \033[0m";
+    exit;
+fi
+MODEL_DIR=MobileNetV1-quant # change to your model
+BATCH_SIZE=1
+USE_CALIB=false
+USE_INT8=true
+DATA_FILE=imagenet-eval-binary/0.data # change to your data file
+build/trt_clas --model_dir=${MODEL_DIR} \
+               --batch_size=${BATCH_SIZE} \
+               --use_calib=${USE_CALIB} \
+               --use_int8=${USE_INT8} \
+               --data_file=${DATA_FILE} \
+               --repeat_times=1
--- a/demo/quant/deploy/TensorRT/tools/test_acc.sh
+++ b/demo/quant/deploy/TensorRT/tools/test_acc.sh
+#!/bin/bash
+unset GREP_OPTIONS;
+if [ ! -d "./build" ];then 
+    echo -e "\033[33m run failed! \033[0m";
+    echo -e "\033[33m you should build first \033[0m";
+    exit;
+fi
+MODEL_DIR=MobileNetV1-quant # change to your model_dir
+DATA_DIR=imagenet-eval-binary # chage to your data_dir
+USE_INT8=True
+./build/test_acc --model_dir=$MODEL_DIR --data_dir=$DATA_DIR --int8=$USE_INT8
--- a/demo/quant/deploy/TensorRT/trt_clas.cc
+++ b/demo/quant/deploy/TensorRT/trt_clas.cc
+#include <numeric>
+#include <iostream>
+#include <fstream>
+#include <sstream>
+#include <memory>
+#include <chrono>
+#include <gflags/gflags.h>
+#include <glog/logging.h>
+#include "paddle/include/paddle_inference_api.h"
+DEFINE_string(model_dir, "", "Directory of the inference model.");
+DEFINE_int32(batch_size, 1, "Batch size.");
+DEFINE_bool(use_calib, true, "Whether to use calib. Set to true if you are using TRT calibration; \
+                             Set to false if you are using PaddleSlim quant models.");
+DEFINE_string(data_file, "", "Path of the inference data file.");
+DEFINE_bool(use_int8, true, "use trt int8 or not");
+DEFINE_int32(repeat_times, 1000, "benchmark repeat time");
+using Time = decltype(std::chrono::high_resolution_clock::now());
+Time time() { return std::chrono::high_resolution_clock::now(); };
+double time_diff(Time t1, Time t2) {
+  typedef std::chrono::microseconds ms;
+  auto diff = t2 - t1;
+  ms counter = std::chrono::duration_cast<ms>(diff);
+  return counter.count() / 1000.0;
+}
+std::pair<int, float> findMax(std::vector<float>& vec) {
+	float max = -1000;
+	int idx = -1;
+	for (int i=0; i<vec.size(); ++i) {
+		if (vec[i] > max) {
+			max = vec[i];
+			idx=i;
+		}
+	}
+	return {idx, max};
+}
+std::unique_ptr<paddle::PaddlePredictor> CreatePredictor() {
+  paddle::AnalysisConfig config;
+  config.SetModel(FLAGS_model_dir + "/model", FLAGS_model_dir + "/params");
+  config.EnableUseGpu(500, 0);
+  // We use ZeroCopy, so we set config.SwitchUseFeedFetchOps(false) here.
+  config.SwitchUseFeedFetchOps(false);
+  if (FLAGS_use_int8){
+    config.EnableTensorRtEngine(1 << 30, \
+                                FLAGS_batch_size, \
+                                5, \
+                                paddle::AnalysisConfig::Precision::kInt8, \
+                                false, \
+                                FLAGS_use_calib);
+  }
+  else{
+    config.EnableTensorRtEngine(1 << 30, \
+                                FLAGS_batch_size, \
+                                5, \
+                                paddle::AnalysisConfig::Precision::kFloat32, \
+                                false, \
+                                false);
+  }
+  return CreatePaddlePredictor(config);
+}
+void run(paddle::PaddlePredictor *predictor,
+         const std::vector<float>& input,
+         const std::vector<int>& input_shape, 
+         std::vector<float> *out_data) {
+  int input_num = std::accumulate(input_shape.begin(), input_shape.end(), 1, std::multiplies<int>());
+  auto input_names = predictor->GetInputNames();
+  auto input_t = predictor->GetInputTensor(input_names[0]);
+  input_t->Reshape(input_shape);
+  input_t->copy_from_cpu(input.data());
+  // warm up 10 times
+  if (FLAGS_repeat_times != 1) {
+      for(int i=0; i<10; i++){
+        CHECK(predictor->ZeroCopyRun());
+      }
+      LOG(INFO) << "finish warm up 10 times";
+  }
+  auto time1 = time(); 
+  for(int i=0; i<FLAGS_repeat_times; i++){
+    CHECK(predictor->ZeroCopyRun());
+    auto output_names = predictor->GetOutputNames();
+    // there is only one output of Resnet50
+    auto output_t = predictor->GetOutputTensor(output_names[0]);
+    std::vector<int> output_shape = output_t->shape();
+    int out_num = std::accumulate(output_shape.begin(), output_shape.end(), 1, std::multiplies<int>());
+    out_data->resize(out_num);
+    output_t->copy_to_cpu(out_data->data());
+  }
+  auto time2 = time(); 
+  auto total_time = time_diff(time1, time2);
+  auto average_time = total_time / FLAGS_repeat_times;
+  LOG(INFO) << "total predict cost is : " << total_time << " ms, repeat " << FLAGS_repeat_times << " times";
+  LOG(INFO) << "average predict cost is : " << average_time << " ms";
+  LOG(INFO) << "finish prediction";
+}
+int main(int argc, char* argv[]) {
+  google::ParseCommandLineFlags(&argc, &argv, true);
+  auto predictor = CreatePredictor();
+  std::vector<int> input_shape = {FLAGS_batch_size, 3, 224, 224}; 
+  // Init input as 1.0 here for example. You can also load preprocessed real pictures to vectors as input.
+  // std::vector<float> input_vec(FLAGS_batch_size * 3 * 224 * 224, 1.0);
+  int input_num = FLAGS_batch_size * 3 * 224 * 224;
+  float *input_data = new float[input_num];
+  memset(input_data, 0, input_num * sizeof(float));
+  std::ifstream fs(FLAGS_data_file, std::ifstream::binary);
+  if (!fs.is_open()) {
+    LOG(FATAL) << "open input file fail.";
+  }
+  auto input_data_tmp = input_data;
+  for (int i = 0; i < input_num; ++i) {
+      fs.read(reinterpret_cast<char*>(input_data_tmp), sizeof(*input_data_tmp));
+      input_data_tmp++;
+  }
+  int label = 0;
+  fs.read(reinterpret_cast<char*>(&label), sizeof(label));
+  fs.close();
+  std::vector<float> input_vec {input_data, input_data + input_num};
+  std::vector<float> out_data;
+  run(predictor.get(), input_vec, input_shape, &out_data);
+  if (FLAGS_batch_size == 1) {
+      std::pair<int, float> result = findMax(out_data);
+      LOG(INFO) << "pred image class is : " << result.first << ", ground truth label is : " << label;
+  }
+  return 0;
+}
--- a/demo/quant/deploy/TensorRT/trt_gen_calib_table_test.cc
+++ b/demo/quant/deploy/TensorRT/trt_gen_calib_table_test.cc
+#include <numeric>
+#include <iostream>
+#include <memory>
+#include <chrono>
+#include <random>
+#include <gflags/gflags.h>
+#include <glog/logging.h>
+#include "paddle/include/paddle_inference_api.h"
+using paddle::AnalysisConfig;
+DEFINE_string(model_file, "", "Path of the inference model file.");
+DEFINE_string(params_file, "", "Path of the inference params file.");
+DEFINE_string(model_dir, "", "Directory of the inference model.");
+DEFINE_int32(batch_size, 1, "Batch size.");
+float Random(float low, float high) {
+  static std::random_device rd;
+  static std::mt19937 mt(rd());
+  std::uniform_real_distribution<double> dist(low, high);
+  return dist(mt);
+}
+using Time = decltype(std::chrono::high_resolution_clock::now());
+Time time() { return std::chrono::high_resolution_clock::now(); };
+double time_diff(Time t1, Time t2) {
+  typedef std::chrono::microseconds ms;
+  auto diff = t2 - t1;
+  ms counter = std::chrono::duration_cast<ms>(diff);
+  return counter.count() / 1000.0;
+}
+std::unique_ptr<paddle::PaddlePredictor> CreatePredictor() {
+  AnalysisConfig config;
+  if (FLAGS_model_dir != "") {
+    config.SetModel(FLAGS_model_dir);
+  } else {
+    config.SetModel(FLAGS_model_file,
+                    FLAGS_params_file);
+  }
+  config.EnableUseGpu(500, 0);
+  // We use ZeroCopy, so we set config.SwitchUseFeedFetchOps(false) here.
+  config.SwitchUseFeedFetchOps(false);
+  config.EnableTensorRtEngine(1 << 30, FLAGS_batch_size, 5, AnalysisConfig::Precision::kInt8, false, true /*use_calib*/);
+  return CreatePaddlePredictor(config);
+}
+void run(paddle::PaddlePredictor *predictor,
+         std::vector<float>& input,
+         const std::vector<int>& input_shape, 
+         std::vector<float> *out_data) {
+  int input_num = std::accumulate(input_shape.begin(), input_shape.end(), 1, std::multiplies<int>());
+  for (size_t i = 0; i < 500; i++) {
+    // We use random data here for example. Change this to real data in your application. 
+    for (int j = 0; j < input_num; j++) {
+      input[j] = Random(0, 1.0);
+    }
+    auto input_names = predictor->GetInputNames();
+    auto input_t = predictor->GetInputTensor(input_names[0]);
+    input_t->Reshape(input_shape);
+    input_t->copy_from_cpu(input.data());
+    // Run predictor to generate calibration table. Can be very time-consuming.     
+    CHECK(predictor->ZeroCopyRun());
+  }
+}
+int main(int argc, char* argv[]) {
+  google::ParseCommandLineFlags(&argc, &argv, true);
+  auto predictor = CreatePredictor();
+  std::vector<int> input_shape = {FLAGS_batch_size, 3, 224, 224}; 
+  // Init input as 1.0 here for example. You can also load preprocessed real pictures to vectors as input.
+  std::vector<float> input_data(FLAGS_batch_size * 3 * 224 * 224, 1.0); 
+  std::vector<float> out_data;
+  run(predictor.get(), input_data, input_shape, &out_data);
+  return 0;
+}