[deploy] support dynamic_shape and add tensorrt tutorial (#2367) (#2392)

* TensorRT support dynamic shape

dygraph/deploy/cpp/include/object_detector.h

@@ -59,12 +59,17 @@ class ObjectDetector {
   explicit ObjectDetector(const std::string& model_dir, 
                           bool use_gpu=false,
                           const std::string& run_mode="fluid",
-                          const int gpu_id=0) {
+                          const int gpu_id=0,
+                          bool use_dynamic_shape=false,
+                          const int trt_min_shape=1,
+                          const int trt_max_shape=1280,
+                          const int trt_opt_shape=640) {
     config_.load_config(model_dir);
     threshold_ = config_.draw_threshold_;
     image_shape_ = config_.image_shape_;
     preprocessor_.Init(config_.preprocess_info_, image_shape_);
-    LoadModel(model_dir, use_gpu, config_.min_subgraph_size_, 1, run_mode, gpu_id);
+    LoadModel(model_dir, use_gpu, config_.min_subgraph_size_, 1, run_mode, gpu_id,
+    use_dynamic_shape, trt_min_shape, trt_max_shape, trt_opt_shape);
   }
 
   // Load Paddle inference model
@@ -74,7 +79,11 @@ class ObjectDetector {
     const int min_subgraph_size,
     const int batch_size = 1,
     const std::string& run_mode = "fluid",
-    const int gpu_id=0);
+    const int gpu_id=0,
+    bool use_dynamic_shape=false,
+    const int trt_min_shape=1,
+    const int trt_max_shape=1280,
+    const int trt_opt_shape=640);
 
   // Run predictor
   void Predict(const cv::Mat& im,

dygraph/deploy/cpp/include/preprocess_op.h

@@ -37,8 +37,8 @@ class ImageBlob {
   std::vector<float> im_shape_;
   // Buffer for image data after preprocessing
   std::vector<float> im_data_;
-  // input image width, height
-  std::vector<int> input_shape_;
+  // in net data shape(after pad)
+  std::vector<int> in_net_shape_;
   // Evaluation image width and height
   //std::vector<float>  eval_im_size_f_;
   // Scale factor for image size to origin image size
@@ -90,7 +90,7 @@ class Resize : public PreprocessOp {
     keep_ratio_ = item["keep_ratio"].as<bool>();
     target_size_ = item["target_size"].as<std::vector<int>>();
     if (item["keep_ratio"]) {
-      input_shape_ = image_shape;
+      in_net_shape_ = image_shape;
     }
  }
 
@@ -103,7 +103,7 @@ class Resize : public PreprocessOp {
   int interp_;
   bool keep_ratio_;
   std::vector<int> target_size_;
-  std::vector<int> input_shape_;
+  std::vector<int> in_net_shape_;
 };
 
 // Models with FPN need input shape % stride == 0

dygraph/deploy/cpp/scripts/build.sh

@@ -35,11 +35,11 @@ then
   echo "set OPENCV_DIR for x86_64"
   # linux系统通过以下命令下载预编译的opencv
   mkdir -p $(pwd)/deps && cd $(pwd)/deps
-  wget -c https://bj.bcebos.com/paddleseg/deploy/opencv3.4.6gcc4.8ffmpeg.tar.gz2
-  tar xvfj opencv3.4.6gcc4.8ffmpeg.tar.gz2 && cd ..
+  wget -c https://paddledet.bj.bcebos.com/data/opencv3.4.6gcc8.2ffmpeg.zip
+  unzip opencv3.4.6gcc8.2ffmpeg.zip && cd ..
 
   # set OPENCV_DIR
-  OPENCV_DIR=$(pwd)/deps/opencv3.4.6gcc4.8ffmpeg/
+  OPENCV_DIR=$(pwd)/deps/opencv3.4.6gcc8.2ffmpeg
 
 elif [ "$MACHINE_TYPE" = "aarch64" ]
 then

dygraph/deploy/cpp/src/main.cc

@@ -42,6 +42,10 @@ DEFINE_int32(camera_id, -1, "Device id of camera to predict");
 DEFINE_bool(run_benchmark, false, "Whether to predict a image_file repeatedly for benchmark");
 DEFINE_double(threshold, 0.5, "Threshold of score.");
 DEFINE_string(output_dir, "output", "Directory of output visualization files.");
+DEFINE_bool(use_dynamic_shape, false, "Trt use dynamic shape or not");
+DEFINE_int32(trt_min_shape, 1, "Min shape of TRT DynamicShapeI");
+DEFINE_int32(trt_max_shape, 1280, "Max shape of TRT DynamicShapeI");
+DEFINE_int32(trt_opt_shape, 640, "Opt shape of TRT DynamicShapeI");
 
 static std::string DirName(const std::string &filepath) {
   auto pos = filepath.rfind(OS_PATH_SEP);
@@ -198,14 +202,17 @@ int main(int argc, char** argv) {
     return -1;
   }
   if (!(FLAGS_run_mode == "fluid" || FLAGS_run_mode == "trt_fp32"
-      || FLAGS_run_mode == "trt_fp16")) {
-    std::cout << "run_mode should be 'fluid', 'trt_fp32' or 'trt_fp16'.";
+      || FLAGS_run_mode == "trt_fp16" || FLAGS_run_mode == "trt_int8")) {
+    std::cout << "run_mode should be 'fluid', 'trt_fp32', 'trt_fp16' or 'trt_int8'.";
     return -1;
   }
   // Load model and create a object detector
-  
-  PaddleDetection::ObjectDetector det(FLAGS_model_dir, FLAGS_use_gpu,
-    FLAGS_run_mode, FLAGS_gpu_id);
+  const std::vector<int> trt_min_shape = {1, FLAGS_trt_min_shape, FLAGS_trt_min_shape};
+  const std::vector<int> trt_max_shape = {1, FLAGS_trt_max_shape, FLAGS_trt_max_shape};
+  const std::vector<int> trt_opt_shape = {1, FLAGS_trt_opt_shape, FLAGS_trt_opt_shape};
+  PaddleDetection::ObjectDetector det(FLAGS_model_dir, FLAGS_use_gpu, FLAGS_run_mode,
+                        FLAGS_gpu_id, FLAGS_use_dynamic_shape, FLAGS_trt_min_shape,
+                        FLAGS_trt_max_shape, FLAGS_trt_opt_shape);
   // Do inference on input video or image
   if (!FLAGS_video_path.empty() || FLAGS_use_camera) {
     PredictVideo(FLAGS_video_path, &det);

dygraph/deploy/cpp/src/object_detector.cc

@@ -27,7 +27,11 @@ void ObjectDetector::LoadModel(const std::string& model_dir,
                                const int min_subgraph_size,
                                const int batch_size,
                                const std::string& run_mode,
-                               const int gpu_id) {
+                               const int gpu_id,
+                               bool use_dynamic_shape,
+                               const int trt_min_shape,
+                               const int trt_max_shape,
+                               const int trt_opt_shape) {
   paddle_infer::Config config;
   std::string prog_file = model_dir + OS_PATH_SEP + "model.pdmodel";
   std::string params_file = model_dir + OS_PATH_SEP + "model.pdiparams";
@@ -35,26 +39,48 @@ void ObjectDetector::LoadModel(const std::string& model_dir,
   if (use_gpu) {
     config.EnableUseGpu(200, gpu_id);
     config.SwitchIrOptim(true);
+    // use tensorrt
+    bool use_calib_mode = false;
     if (run_mode != "fluid") {
       auto precision = paddle_infer::Config::Precision::kFloat32;
-      if (run_mode == "trt_fp16") {
+      if (run_mode == "trt_fp32") {
+        precision = paddle_infer::Config::Precision::kFloat32;
+      }
+      else if (run_mode == "trt_fp16") {
         precision = paddle_infer::Config::Precision::kHalf;
-      } else if (run_mode == "trt_int8") {
-        printf("TensorRT int8 mode is not supported now, "
-               "please use 'trt_fp32' or 'trt_fp16' instead");
+      }
+      else if (run_mode == "trt_int8") {
+        precision = paddle_infer::Config::Precision::kInt8;
+        use_calib_mode = true;
       } else {
-        if (run_mode != "trt_fp32") {
-          printf("run_mode should be 'fluid', 'trt_fp32' or 'trt_fp16'");
-        }
+          printf("run_mode should be 'fluid', 'trt_fp32', 'trt_fp16' or 'trt_int8'");
       }
+      // set tensorrt
       config.EnableTensorRtEngine(
-          1 << 10,
+          1 << 30,
           batch_size,
           min_subgraph_size,
           precision,
           false,
-          false);
-   }
+          use_calib_mode);
+
+      // set use dynamic shape
+      if (use_dynamic_shape) {
+        // set DynamicShsape for image tensor
+        const std::vector<int> min_input_shape = {1, trt_min_shape, trt_min_shape};
+        const std::vector<int> max_input_shape = {1, trt_max_shape, trt_max_shape};
+        const std::vector<int> opt_input_shape = {1, trt_opt_shape, trt_opt_shape};
+        const std::map<std::string, std::vector<int>> map_min_input_shape = {{"image", min_input_shape}};
+        const std::map<std::string, std::vector<int>> map_max_input_shape = {{"image", max_input_shape}};
+        const std::map<std::string, std::vector<int>> map_opt_input_shape = {{"image", opt_input_shape}};
+
+        config.SetTRTDynamicShapeInfo(map_min_input_shape,
+                                      map_max_input_shape,
+                                      map_opt_input_shape);
+        std::cout << "TensorRT dynamic shape enabled" << std::endl;
+      }
+    }
+
   } else {
     config.DisableGpu();
   }
@@ -171,8 +197,8 @@ void ObjectDetector::Predict(const cv::Mat& im,
   for (const auto& tensor_name : input_names) {
     auto in_tensor = predictor_->GetInputHandle(tensor_name);
     if (tensor_name == "image") {
-      int rh = inputs_.input_shape_[0];
-      int rw = inputs_.input_shape_[1];
+      int rh = inputs_.in_net_shape_[0];
+      int rw = inputs_.in_net_shape_[1];
       in_tensor->Reshape({1, 3, rh, rw});
       in_tensor->CopyFromCpu(inputs_.im_data_.data());
     } else if (tensor_name == "im_shape") {

dygraph/deploy/cpp/src/preprocess_op.cc

@@ -25,7 +25,7 @@ void InitInfo::Run(cv::Mat* im, ImageBlob* data) {
       static_cast<float>(im->cols)
   };
   data->scale_factor_ = {1., 1.};
-  data->input_shape_ = {
+  data->in_net_shape_ = {
       static_cast<int>(im->rows),
       static_cast<int>(im->cols)
   };
@@ -62,7 +62,11 @@ void Permute::Run(cv::Mat* im, ImageBlob* data) {
 
 void Resize::Run(cv::Mat* im, ImageBlob* data) {
   auto resize_scale = GenerateScale(*im);
-  data->input_shape_ = {
+  data->im_shape_ = {
+      static_cast<int>(im->cols * resize_scale.first),
+      static_cast<int>(im->rows * resize_scale.second)
+  };
+  data->in_net_shape_ = {
       static_cast<int>(im->cols * resize_scale.first),
       static_cast<int>(im->rows * resize_scale.second)
   };
@@ -121,7 +125,7 @@ void PadStride::Run(cv::Mat* im, ImageBlob* data) {
     nw - rw,
     cv::BORDER_CONSTANT,
     cv::Scalar(0));
-  data->input_shape_ = {
+  data->in_net_shape_ = {
     static_cast<int>(im->rows),
     static_cast<int>(im->cols),
   };

dygraph/deploy/python/infer.py

@@ -46,6 +46,11 @@ class Detector(object):
         model_dir (str): root path of model.pdiparams, model.pdmodel and infer_cfg.yml
         use_gpu (bool): whether use gpu
         run_mode (str): mode of running(fluid/trt_fp32/trt_fp16)
+        use_dynamic_shape (bool): use dynamic shape or not
+        trt_min_shape (int): min shape for dynamic shape in trt
+        trt_max_shape (int): max shape for dynamic shape in trt
+        trt_opt_shape (int): opt shape for dynamic shape in trt
+        run_mode (str): mode of running(fluid/trt_fp32/trt_fp16)
         threshold (float): threshold to reserve the result for output.
     """
 
@@ -54,13 +59,21 @@ class Detector(object):
                  model_dir,
                  use_gpu=False,
                  run_mode='fluid',
+                 use_dynamic_shape=False,
+                 trt_min_shape=1,
+                 trt_max_shape=1280,
+                 trt_opt_shape=640,
                  threshold=0.5):
         self.pred_config = pred_config
         self.predictor = load_predictor(
             model_dir,
             run_mode=run_mode,
             min_subgraph_size=self.pred_config.min_subgraph_size,
-            use_gpu=use_gpu)
+            use_gpu=use_gpu,
+            use_dynamic_shape=use_dynamic_shape,
+            trt_min_shape=trt_min_shape,
+            trt_max_shape=trt_max_shape,
+            trt_opt_shape=trt_opt_shape)
 
     def preprocess(self, im):
         preprocess_ops = []
@@ -154,6 +167,10 @@ class DetectorSOLOv2(Detector):
         model_dir (str): root path of model.pdiparams, model.pdmodel and infer_cfg.yml
         use_gpu (bool): whether use gpu
         run_mode (str): mode of running(fluid/trt_fp32/trt_fp16)
+        use_dynamic_shape (bool): use dynamic shape or not
+        trt_min_shape (int): min shape for dynamic shape in trt
+        trt_max_shape (int): max shape for dynamic shape in trt
+        trt_opt_shape (int): opt shape for dynamic shape in trt
         threshold (float): threshold to reserve the result for output.
     """
 
@@ -162,13 +179,21 @@ class DetectorSOLOv2(Detector):
                  model_dir,
                  use_gpu=False,
                  run_mode='fluid',
+                 use_dynamic_shape=False,
+                 trt_min_shape=1,
+                 trt_max_shape=1280,
+                 trt_opt_shape=640,
                  threshold=0.5):
         self.pred_config = pred_config
         self.predictor = load_predictor(
             model_dir,
             run_mode=run_mode,
             min_subgraph_size=self.pred_config.min_subgraph_size,
-            use_gpu=use_gpu)
+            use_gpu=use_gpu,
+            use_dynamic_shape=use_dynamic_shape,
+            trt_min_shape=trt_min_shape,
+            trt_max_shape=trt_max_shape,
+            trt_opt_shape=trt_opt_shape)
 
     def predict(self,
                 image,
@@ -287,11 +312,20 @@ def load_predictor(model_dir,
                    run_mode='fluid',
                    batch_size=1,
                    use_gpu=False,
-                   min_subgraph_size=3):
+                   min_subgraph_size=3,
+                   use_dynamic_shape=False,
+                   trt_min_shape=1,
+                   trt_max_shape=1280,
+                   trt_opt_shape=640):
     """set AnalysisConfig, generate AnalysisPredictor
     Args:
         model_dir (str): root path of __model__ and __params__
         use_gpu (bool): whether use gpu
+        run_mode (str): mode of running(fluid/trt_fp32/trt_fp16)
+        use_dynamic_shape (bool): use dynamic shape or not
+        trt_min_shape (int): min shape for dynamic shape in trt
+        trt_max_shape (int): max shape for dynamic shape in trt
+        trt_opt_shape (int): opt shape for dynamic shape in trt
     Returns:
         predictor (PaddlePredictor): AnalysisPredictor
     Raises:
@@ -301,9 +335,12 @@ def load_predictor(model_dir,
         raise ValueError(
             "Predict by TensorRT mode: {}, expect use_gpu==True, but use_gpu == {}"
             .format(run_mode, use_gpu))
-    if run_mode == 'trt_int8':
-        raise ValueError("TensorRT int8 mode is not supported now, "
-                         "please use trt_fp32 or trt_fp16 instead.")
+    if run_mode == 'trt_int8' and not os.path.exists(
+            os.path.join(model_dir, '_opt_cache')):
+        raise ValueError(
+            "TensorRT int8 must calibration first, and model_dir must has _opt_cache dir"
+        )
+    use_calib_mode = True if run_mode == 'trt_int8' else False
     config = Config(
         os.path.join(model_dir, 'model.pdmodel'),
         os.path.join(model_dir, 'model.pdiparams'))
@@ -316,11 +353,7 @@ def load_predictor(model_dir,
         # initial GPU memory(M), device ID
         config.enable_use_gpu(200, 0)
         # optimize graph and fuse op
-        # FIXME(dkp): ir optimize may prune variable inside graph
-        #             and incur error in Paddle 2.0, e.g. in SSDLite
-        #             FCOS model, set as False currently and should
-        #             be set as True after switch_ir_optim fixed
-        config.switch_ir_optim(False)
+        config.switch_ir_optim(True)
     else:
         config.disable_gpu()
 
@@ -331,7 +364,16 @@ def load_predictor(model_dir,
             min_subgraph_size=min_subgraph_size,
             precision_mode=precision_map[run_mode],
             use_static=False,
-            use_calib_mode=False)
+            use_calib_mode=use_calib_mode)
+
+        if use_dynamic_shape:
+            print('use_dynamic_shape')
+            min_input_shape = {'image': [1, 3, trt_min_shape, trt_min_shape]}
+            max_input_shape = {'image': [1, 3, trt_max_shape, trt_max_shape]}
+            opt_input_shape = {'image': [1, 3, trt_opt_shape, trt_opt_shape]}
+            config.set_trt_dynamic_shape_info(min_input_shape, max_input_shape,
+                                              opt_input_shape)
+            print('trt set dynamic shape done!')
 
     # disable print log when predict
     config.disable_glog_info()
@@ -424,13 +466,21 @@ def main():
         pred_config,
         FLAGS.model_dir,
         use_gpu=FLAGS.use_gpu,
-        run_mode=FLAGS.run_mode)
+        run_mode=FLAGS.run_mode,
+        use_dynamic_shape=FLAGS.use_dynamic_shape,
+        trt_min_shape=FLAGS.trt_min_shape,
+        trt_max_shape=FLAGS.trt_max_shape,
+        trt_opt_shape=FLAGS.trt_opt_shape)
     if pred_config.arch == 'SOLOv2':
         detector = DetectorSOLOv2(
             pred_config,
             FLAGS.model_dir,
             use_gpu=FLAGS.use_gpu,
-            run_mode=FLAGS.run_mode)
+            run_mode=FLAGS.run_mode,
+            use_dynamic_shape=FLAGS.use_dynamic_shape,
+            trt_min_shape=FLAGS.trt_min_shape,
+            trt_max_shape=FLAGS.trt_max_shape,
+            trt_opt_shape=FLAGS.trt_opt_shape)
     # predict from image
     if FLAGS.image_file != '':
         predict_image(detector)
@@ -480,6 +530,23 @@ if __name__ == '__main__':
         type=str,
         default="output",
         help="Directory of output visualization files.")
+    parser.add_argument(
+        "--use_dynamic_shape",
+        type=ast.literal_eval,
+        default=False,
+        help="Dynamic_shape for TensorRT.")
+    parser.add_argument(
+        "--trt_min_shape", type=int, default=1, help="min_shape for TensorRT.")
+    parser.add_argument(
+        "--trt_max_shape",
+        type=int,
+        default=1280,
+        help="max_shape for TensorRT.")
+    parser.add_argument(
+        "--trt_opt_shape",
+        type=int,
+        default=640,
+        help="opt_shape for TensorRT.")
 
     FLAGS = parser.parse_args()
     print_arguments(FLAGS)

dygraph/deploy/python/trt_int8_calib.py

+# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+
+import os
+import argparse
+import time
+import yaml
+import ast
+from functools import reduce
+
+from PIL import Image
+import cv2
+import numpy as np
+import glob
+import paddle
+from preprocess import preprocess, Resize, NormalizeImage, Permute, PadStride
+from visualize import visualize_box_mask
+from paddle.inference import Config
+from paddle.inference import create_predictor
+
+# Global dictionary
+SUPPORT_MODELS = {
+    'YOLO',
+    'RCNN',
+    'SSD',
+    'FCOS',
+    'SOLOv2',
+    'TTFNet',
+}
+
+
+class Detector(object):
+    """
+    Args:
+        config (object): config of model, defined by `Config(model_dir)`
+        model_dir (str): root path of model.pdiparams, model.pdmodel and infer_cfg.yml
+        use_gpu (bool): whether use gpu
+    """
+
+    def __init__(self, pred_config, model_dir, use_gpu=False):
+        self.pred_config = pred_config
+        self.predictor = load_predictor(
+            model_dir,
+            min_subgraph_size=self.pred_config.min_subgraph_size,
+            use_gpu=use_gpu)
+
+    def preprocess(self, im):
+        preprocess_ops = []
+        for op_info in self.pred_config.preprocess_infos:
+            new_op_info = op_info.copy()
+            op_type = new_op_info.pop('type')
+            preprocess_ops.append(eval(op_type)(**new_op_info))
+        im, im_info = preprocess(im, preprocess_ops,
+                                 self.pred_config.input_shape)
+        inputs = create_inputs(im, im_info)
+        return inputs
+
+    def postprocess(self, np_boxes, np_masks, inputs, threshold=0.5):
+        # postprocess output of predictor
+        results = {}
+        if self.pred_config.arch in ['Face']:
+            h, w = inputs['im_shape']
+            scale_y, scale_x = inputs['scale_factor']
+            w, h = float(h) / scale_y, float(w) / scale_x
+            np_boxes[:, 2] *= h
+            np_boxes[:, 3] *= w
+            np_boxes[:, 4] *= h
+            np_boxes[:, 5] *= w
+        results['boxes'] = np_boxes
+        if np_masks is not None:
+            results['masks'] = np_masks
+        return results
+
+    def predict(self,
+                image,
+                threshold=0.5,
+                warmup=0,
+                repeats=1,
+                run_benchmark=False):
+        '''
+        Args:
+            image (str/np.ndarray): path of image/ np.ndarray read by cv2
+            threshold (float): threshold of predicted box' score
+        Returns:
+            results (dict): include 'boxes': np.ndarray: shape:[N,6], N: number of box,
+                            matix element:[class, score, x_min, y_min, x_max, y_max]
+                            MaskRCNN's results include 'masks': np.ndarray:
+                            shape: [N, im_h, im_w]
+        '''
+        inputs = self.preprocess(image)
+        np_boxes, np_masks = None, None
+        input_names = self.predictor.get_input_names()
+        for i in range(len(input_names)):
+            input_tensor = self.predictor.get_input_handle(input_names[i])
+            input_tensor.copy_from_cpu(inputs[input_names[i]])
+
+        for i in range(warmup):
+            self.predictor.run()
+            output_names = self.predictor.get_output_names()
+            boxes_tensor = self.predictor.get_output_handle(output_names[0])
+            np_boxes = boxes_tensor.copy_to_cpu()
+            if self.pred_config.mask:
+                masks_tensor = self.predictor.get_output_handle(output_names[2])
+                np_masks = masks_tensor.copy_to_cpu()
+
+        t1 = time.time()
+        for i in range(repeats):
+            self.predictor.run()
+            output_names = self.predictor.get_output_names()
+            boxes_tensor = self.predictor.get_output_handle(output_names[0])
+            np_boxes = boxes_tensor.copy_to_cpu()
+            if self.pred_config.mask:
+                masks_tensor = self.predictor.get_output_handle(output_names[2])
+                np_masks = masks_tensor.copy_to_cpu()
+        t2 = time.time()
+        ms = (t2 - t1) * 1000.0 / repeats
+        print("Inference: {} ms per batch image".format(ms))
+
+        # do not perform postprocess in benchmark mode
+        results = []
+        if not run_benchmark:
+            if reduce(lambda x, y: x * y, np_boxes.shape) < 6:
+                print('[WARNNING] No object detected.')
+                results = {'boxes': np.array([])}
+            else:
+                results = self.postprocess(
+                    np_boxes, np_masks, inputs, threshold=threshold)
+
+        return results
+
+
+def create_inputs(im, im_info):
+    """generate input for different model type
+    Args:
+        im (np.ndarray): image (np.ndarray)
+        im_info (dict): info of image
+        model_arch (str): model type
+    Returns:
+        inputs (dict): input of model
+    """
+    inputs = {}
+    inputs['image'] = np.array((im, )).astype('float32')
+    inputs['im_shape'] = np.array((im_info['im_shape'], )).astype('float32')
+    inputs['scale_factor'] = np.array(
+        (im_info['scale_factor'], )).astype('float32')
+
+    return inputs
+
+
+class PredictConfig():
+    """set config of preprocess, postprocess and visualize
+    Args:
+        model_dir (str): root path of model.yml
+    """
+
+    def __init__(self, model_dir):
+        # parsing Yaml config for Preprocess
+        deploy_file = os.path.join(model_dir, 'infer_cfg.yml')
+        with open(deploy_file) as f:
+            yml_conf = yaml.safe_load(f)
+        self.check_model(yml_conf)
+        self.arch = yml_conf['arch']
+        self.preprocess_infos = yml_conf['Preprocess']
+        self.min_subgraph_size = yml_conf['min_subgraph_size']
+        self.labels = yml_conf['label_list']
+        self.mask = False
+        if 'mask' in yml_conf:
+            self.mask = yml_conf['mask']
+        self.input_shape = yml_conf['image_shape']
+        self.print_config()
+
+    def check_model(self, yml_conf):
+        """
+        Raises:
+            ValueError: loaded model not in supported model type 
+        """
+        for support_model in SUPPORT_MODELS:
+            if support_model in yml_conf['arch']:
+                return True
+        raise ValueError("Unsupported arch: {}, expect {}".format(yml_conf[
+            'arch'], SUPPORT_MODELS))
+
+    def print_config(self):
+        print('-----------  Model Configuration -----------')
+        print('%s: %s' % ('Model Arch', self.arch))
+        print('%s: ' % ('Transform Order'))
+        for op_info in self.preprocess_infos:
+            print('--%s: %s' % ('transform op', op_info['type']))
+        print('--------------------------------------------')
+
+
+def load_predictor(model_dir, batch_size=1, use_gpu=False, min_subgraph_size=3):
+    """set AnalysisConfig, generate AnalysisPredictor
+    Args:
+        model_dir (str): root path of __model__ and __params__
+        use_gpu (bool): whether use gpu
+    Returns:
+        predictor (PaddlePredictor): AnalysisPredictor
+    Raises:
+        ValueError: predict by TensorRT need use_gpu == True.
+    """
+    run_mode = 'trt_int8'
+    if not use_gpu and not run_mode == 'fluid':
+        raise ValueError(
+            "Predict by TensorRT mode: {}, expect use_gpu==True, but use_gpu == {}"
+            .format(run_mode, use_gpu))
+    config = Config(
+        os.path.join(model_dir, 'model.pdmodel'),
+        os.path.join(model_dir, 'model.pdiparams'))
+    precision_map = {
+        'trt_int8': Config.Precision.Int8,
+        'trt_fp32': Config.Precision.Float32,
+        'trt_fp16': Config.Precision.Half
+    }
+    if use_gpu:
+        # initial GPU memory(M), device ID
+        config.enable_use_gpu(200, 0)
+        # optimize graph and fuse op
+        config.switch_ir_optim(True)
+    else:
+        config.disable_gpu()
+
+    if run_mode in precision_map.keys():
+        config.enable_tensorrt_engine(
+            workspace_size=1 << 10,
+            max_batch_size=batch_size,
+            min_subgraph_size=min_subgraph_size,
+            precision_mode=precision_map[run_mode],
+            use_static=False,
+            use_calib_mode=True)
+
+    # disable print log when predict
+    config.disable_glog_info()
+    # enable shared memory
+    config.enable_memory_optim()
+    # disable feed, fetch OP, needed by zero_copy_run
+    config.switch_use_feed_fetch_ops(False)
+    predictor = create_predictor(config)
+    return predictor
+
+
+def print_arguments(args):
+    print('-----------  Running Arguments -----------')
+    for arg, value in sorted(vars(args).items()):
+        print('%s: %s' % (arg, value))
+    print('------------------------------------------')
+
+
+def predict_image_dir(detector):
+    for image_file in glob.glob(FLAGS.image_dir + '/*.jpg'):
+        print('image_file is', image_file)
+        results = detector.predict(image_file, threshold=0.5)
+
+
+def main():
+    pred_config = PredictConfig(FLAGS.model_dir)
+    detector = Detector(pred_config, FLAGS.model_dir, use_gpu=FLAGS.use_gpu)
+    # predict from image
+    if FLAGS.image_dir != '':
+        predict_image_dir(detector)
+
+
+if __name__ == '__main__':
+    paddle.enable_static()
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--model_dir",
+        type=str,
+        default=None,
+        help=("Directory include:'model.pdiparams', 'model.pdmodel', "
+              "'infer_cfg.yml', created by tools/export_model.py."),
+        required=True)
+    parser.add_argument(
+        "--image_dir", type=str, default='', help="Directory of image file.")
+    parser.add_argument(
+        "--use_gpu",
+        type=ast.literal_eval,
+        default=False,
+        help="Whether to predict with GPU.")
+    print('err?')
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="output",
+        help="Directory of output visualization files.")
+    FLAGS = parser.parse_args()
+    print_arguments(FLAGS)
+
+    main()

dygraph/deploy/tensorrt/README.md

+# TensorRT预测部署教程
+TensorRT是NVIDIA提出的用于统一模型部署的加速库，可以应用于V100、JETSON Xavier等硬件，它可以极大提高预测速度。Paddle TensorRT教程请参考文档[使用Paddle-TensorRT库预测](https://paddle-inference.readthedocs.io/en/latest/optimize/paddle_trt.html#)
+
+## 1. 安装PaddleInference预测库
+- Python安装包，请从[这里](https://www.paddlepaddle.org.cn/documentation/docs/zh/install/Tables.html#whl-release) 下载带有tensorrt的安装包进行安装
+
+- CPP预测库，请从[这里](https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/05_inference_deployment/inference/build_and_install_lib_cn.html) 下载带有TensorRT编译的预测库
+
+- 如果Python和CPP官网没有提供已编译好的安装包或预测库，请参考[源码安装](https://www.paddlepaddle.org.cn/documentation/docs/zh/install/compile/linux-compile.html) 自行编译
+
+注意，您的机器上TensorRT的版本需要跟您使用的预测库中TensorRT版本保持一致。
+
+## 2. 导出模型
+模型导出具体请参考文档[PaddleDetection模型导出教程](../EXPORT_MODEL.md)。
+
+## 3. 开启TensorRT加速
+### 3.1 配置TensorRT
+在使用Paddle预测库构建预测器配置config时，打开TensorRT引擎就可以了：
+
+```
+config->EnableUseGpu(100, 0); // 初始化100M显存，使用GPU ID为0
+config->GpuDeviceId();        // 返回正在使用的GPU ID
+// 开启TensorRT预测，可提升GPU预测性能，需要使用带TensorRT的预测库
+config->EnableTensorRtEngine(1 << 20             /*workspace_size*/,
+                             batch_size        /*max_batch_size*/,
+                             3                 /*min_subgraph_size*/,
+                             AnalysisConfig::Precision::kFloat32 /*precision*/,
+                             false             /*use_static*/,
+                             false             /*use_calib_mode*/);
+
+```
+
+### 3.2 TensorRT固定尺寸预测
+TensorRT版本<=5时，使用TensorRT预测时，只支持固定尺寸输入。
+
+在导出模型时指定模型输入尺寸，设置`TestReader.inputs_def.image_shape=[3,640,640]`，具体请参考[PaddleDetection模型导出教程](../EXPORT_MODEL.md) 。
+
+`TestReader.inputs_def.image_shape`设置的是输入TensorRT引擎的数据尺寸（在像FasterRCNN中，`TestReader.inputs_def.image_shape`指定的是在`Pad`操作之前的图像数据尺寸）。
+
+可以通过[visualdl](https://www.paddlepaddle.org.cn/paddle/visualdl/demo/graph) 打开`model.pdmodel`文件，查看输入的第一个Tensor尺寸是否是固定的，如果不指定，尺寸会用`？`表示，如下图所示：
+![img](imgs/input_shape.png)
+
+### 3.3 TensorRT动态尺寸预测
+
+TensorRT版本>=6时，使用TensorRT预测时，可以支持动态尺寸输入。
+Paddle预测库关于动态尺寸输入请查看[Paddle CPP预测](https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/05_inference_deployment/inference/native_infer.html) 的`SetTRTDynamicShapeInfo`函数说明。
+
+`python/infer.py`设置动态尺寸输入参数说明：
+
+- use_dynamic_shape 用于设定TensorRT的输入尺寸是否是动态尺寸，默认值：False
+
+- trt_min_shape 用于设定TensorRT的输入图像height、width中的最小尺寸，默认值：1
+
+- trt_max_shape 用于设定TensorRT的输入图像height、width中的最大尺寸，默认值：1280
+
+- trt_opt_shape 用于设定TensorRT的输入图像height、width中的最优尺寸，默认值：640
+
+**注意：`TensorRT`中动态尺寸设置是4维的，这里只设置输入图像的尺寸。**
+
+## 4、常见问题QA
+**Q:** 提示没有`tensorrt_op`</br>
+**A:** 请检查是否使用带有TensorRT的Paddle Python包或预测库。
+
+**Q:** 提示`op out of memory`</br>
+**A:** 检查GPU是否是别人也在使用，请尝试使用空闲GPU
+
+**Q:** 提示`some trt inputs dynamic shape info not set`</br>
+**A:** 这是由于`TensorRT`会把网络结果划分成多个子图，我们只设置了输入数据的动态尺寸，划分的其他子图的输入并未设置动态尺寸。有两个解决方法：
+
+- 方法一：通过增大`min_subgraph_size`，跳过对这些子图的优化。根据提示，设置min_subgraph_size大于并未设置动态尺寸输入的子图中OP个数即可。
+`min_subgraph_size`的意思是，在加载TensorRT引擎的时候，大于`min_subgraph_size`的OP才会被优化，并且这些OP是连续的且是TensorRT可以优化的。
+
+- 方法二：找到子图的这些输入，按照上面方式也设置子图的输入动态尺寸。
+
+**Q:** 如何打开日志</br>
+**A:** 预测库默认是打开日志的，只要注释掉`config.disable_glog_info()`就可以打开日志