NANO自带的tensorrt运行卡慢，每帧图像处理速度在3s左右

问题描述: 在运行NANO自带的tensorrt包中的 tensorrt/samples/python/yolov3_onnx/onnx_to_tensorrt.py 时，运行卡慢，每帧图像处理速度在3s左右，相比于直接在darknet下使用yolov3进行检测时，无明显加速效果，求指点。

代码(onnx_to_tensorrt.py 见附件。
经测试，主要耗时的程序在boxes, classes, scores = postprocessor.process(trt_outputs, (shape_orig_WH))，每张图片该行程序耗时在2-3s.求助!
本人运行的步骤如下:
1. $ sudo python yolov3_to_onnx.py
2. $ sudo python onnx_to_tensorrt.py

onnx_to_tensorrt.py代码:

1. #!/usr/bin/env python2
   
2. #
   
3. # Copyright 1993-2019 NVIDIA Corporation.  All rights reserved.
   
4. #
   
5. # NOTICE TO LICENSEE:
   
6. #
   
7. # This source code and/or documentation ("Licensed Deliverables") are
   
8. # subject to NVIDIA intellectual property rights under U.S. and
   
9. # international Copyright laws.
   
10. #
    
11. # These Licensed Deliverables contained herein is PROPRIETARY and
    
12. # CONFIDENTIAL to NVIDIA and is being provided under the terms and
    
13. # conditions of a form of NVIDIA software license agreement by and
    
14. # between NVIDIA and Licensee ("License Agreement") or electronically
    
15. # accepted by Licensee.  Notwithstanding any terms or conditions to
    
16. # the contrary in the License Agreement, reproduction or disclosure
    
17. # of the Licensed Deliverables to any third party without the express
    
18. # written consent of NVIDIA is prohibited.
    
19. #
    
20. # NOTWITHSTANDING ANY TERMS OR CONDITIONS TO THE CONTRARY IN THE
    
21. # LICENSE AGREEMENT, NVIDIA MAKES NO REPRESENTATION ABOUT THE
    
22. # SUITABILITY OF THESE LICENSED DELIVERABLES FOR ANY PURPOSE.  IT IS
    
23. # PROVIDED "AS IS" WITHOUT EXPRESS OR IMPLIED WARRANTY OF ANY KIND.
    
24. # NVIDIA DISCLAIMS ALL WARRANTIES WITH REGARD TO THESE LICENSED
    
25. # DELIVERABLES, INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY,
    
26. # NONINFRINGEMENT, AND FITNESS FOR A PARTICULAR PURPOSE.
    
27. # NOTWITHSTANDING ANY TERMS OR CONDITIONS TO THE CONTRARY IN THE
    
28. # LICENSE AGREEMENT, IN NO EVENT SHALL NVIDIA BE LIABLE FOR ANY
    
29. # SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL DAMAGES, OR ANY
    
30. # DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS,
    
31. # WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS
    
32. # ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE
    
33. # OF THESE LICENSED DELIVERABLES.
    
34. #
    
35. # U.S. Government End Users.  These Licensed Deliverables are a
    
36. # "commercial item" as that term is defined at 48 C.F.R. 2.101 (OCT
    
37. # 1995), consisting of "commercial computer software" and "commercial
    
38. # computer software documentation" as such terms are used in 48
    
39. # C.F.R. 12.212 (SEPT 1995) and is provided to the U.S. Government
    
40. # only as a commercial end item.  Consistent with 48 C.F.R.12.212 and
    
41. # 48 C.F.R. 227.7202-1 through 227.7202-4 (JUNE 1995), all
    
42. # U.S. Government End Users acquire the Licensed Deliverables with
    
43. # only those rights set forth herein.
    
44. #
    
45. # Any use of the Licensed Deliverables in individual and commercial
    
46. # software must include, in the user documentation and internal
    
47. # comments to the code, the above Disclaimer and U.S. Government End
    
48. # Users Notice.
    
49. #
    
50. 
    
51. from __future__ import print_function
    
52. import cv2 as cv
    
53. import numpy as np
    
54. import tensorrt as trt
    
55. import pycuda.driver as cuda
    
56. import pycuda.autoinit
    
57. from PIL import ImageDraw
    
58. import time
    
59. from yolov3_to_onnx import download_file
    
60. from data_processing import PreprocessYOLO, PostprocessYOLO, ALL_CATEGORIES
    
61. 
    
62. import sys, os
    
63. sys.path.insert(1, os.path.join(sys.path[0], ".."))
    
64. import common
    
65. 
    
66. TRT_LOGGER = trt.Logger()
    
67. cap = cv.VideoCapture(0)
    
68. 
    
69. def load_label_categories(label_file_path):
    
70.     categories = [line.rstrip('\n') for line in open(label_file_path)]
    
71.     return categories
    
72. 
    
73. LABEL_FILE_PATH = os.path.join(os.path.dirname(os.path.realpath(__file__)), 'coco_labels.txt')
    
74. ALL_CATEGORIES = load_label_categories(LABEL_FILE_PATH)
    
75. 
    
76. def get_engine1(vv):
    
77.     with trt.Runtime(TRT_LOGGER) as runtime:
    
78.         return runtime.deserialize_cuda_engine(vv)
    
79. 
    
80. def get_engine(onnx_file_path, engine_file_path=""):
    
81.     def build_engine():
    
82.         """Takes an ONNX file and creates a TensorRT engine to run inference with"""
    
83.         with trt.Builder(TRT_LOGGER) as builder,\
    
84.               builder.create_network() as network, \
    
85.               trt.OnnxParser(network, TRT_LOGGER) as parser:
    
86. 
    
87.             builder.max_workspace_size = 1 << 30 # 1GB
    
88.             builder.max_batch_size = 1
    
89. 
    
90.             if not os.path.exists(onnx_file_path):
    
91.                 print('ONNX file {} not found, please run yolov3_to_onnx.py first to generate it.'.format(onnx_file_path))
    
92.                 exit(0)
    
93. 
    
94.             print('Loading ONNX file from path {}...'.format(onnx_file_path))
    
95.             with open(onnx_file_path, 'rb') as model:
    
96.                 print('Beginning ONNX file parsing')
    
97.                 parser.parse(model.read())
    
98.             print('Completed parsing of ONNX file')
    
99. 
    
100.             print('Building an engine from file {}; this may take a while...'.format(onnx_file_path))
     
101.             engine = builder.build_cuda_engine(network)
     
102.             print("Completed creating Engine")
     
103. 
     
104.             with open(engine_file_path, "wb") as f:
     
105.                 f.write(engine.serialize())
     
106.             return engine
     
107. 
     
108.     if os.path.exists(engine_file_path):
     
109.         print("Reading engine from file {}".format(engine_file_path))
     
110.         with open(engine_file_path, "rb") as f, \
     
111.              trt.Runtime(TRT_LOGGER) as runtime:
     
112.             return runtime.deserialize_cuda_engine(f.read())
     
113.     else:
     
114.         return build_engine()
     
115. 
     
116. def main():
     
117.     """Create a TensorRT engine for ONNX-based YOLOv3-608 and run inference."""
     
118. 
     
119.     # Try to load a previously generated YOLOv3-608 network graph in ONNX format:
     
120.     onnx_file_path = './yolov3.onnx'
     
121.     engine_file_path = "./yolov3.trt"
     
122.     file = open(engine_file_path,"rb")
     
123.     f = file.read()
     
124.     file.close()
     
125.     engine= get_engine1(f)
     
126.     context = engine.create_execution_context()
     
127.     # with get_engine(onnx_file_path, engine_file_path) as engine:
     
128.     #     print("finished")
     
129.     # Download a dog image and save it to the following file path:
     
130.     while True:
     
131.         ret,frame = cap.read()
     
132.         if ret:
     
133.             x, y = frame.shape[0:2]
     
134.                 # Two-dimensional tuple with the target network's (spatial) input resolution in HW ordered
     
135.             input_resolution_yolov3_HW = (608, 608)
     
136.             # Create a pre-processor object by specifying the required input resolution for YOLOv3
     
137.             preprocessor = PreprocessYOLO(input_resolution_yolov3_HW)
     
138.             # Load an image from the specified input path, and return it together with  a pre-processed version
     
139.             image_raw, image = preprocessor.process(frame)
     
140.             # Store the shape of the original input image in WH format, we will need it for later
     
141.             shape_orig_WH = image_raw.size
     
142. 
     
143.             # Output shapes expected by the post-processor
     
144.             output_shapes = [(1, 255, 19, 19), (1, 255, 38, 38), (1, 255, 76, 76)]
     
145.             # Do inference with TensorRT
     
146.             trt_outputs = [] #get_engine(onnx_file_path, engine_file_path) as engine,
     
147.             
     
148.             
     
149.             inputs, outputs, bindings, stream = common.allocate_buffers(engine)
     
150.                 # Do inference
     
151.                 #print('Running inference on image {}...'.format(input_image_path))
     
152.                 # Set host input to the image. The common.do_inference function will copy the input to the GPU before executing.
     
153.             inputs[0].host = image
     
154.             trt_outputs = common.do_inference(context, bindings=bindings, inputs=inputs, outputs=outputs, stream=stream)
     
155.             #b = time.clock()
     
156.             # Before doing post-processing, we need to reshape the outputs as the common.do_inference will give us flat arrays.
     
157.             trt_outputs = [output.reshape(shape) for output, shape in zip(trt_outputs, output_shapes)]
     
158. 
     
159.             postprocessor_args = {"yolo_masks": [(6, 7, 8), (3, 4, 5), (0, 1, 2)],                    # A list of 3 three-dimensional tuples for the YOLO masks
     
160.                             "yolo_anchors": [(10, 13), (16, 30), (33, 23), (30, 61), (62, 45),  # A list of 9 two-dimensional tuples for the YOLO anchors
     
161.                                             (59, 119), (116, 90), (156, 198), (373, 326)],
     
162.                             "obj_threshold": 0.6,                                               # Threshold for object coverage, float value between 0 and 1
     
163.                             "nms_threshold": 0.5,                                               # Threshold for non-max suppression algorithm, float value between 0 and 1
     
164.                             "yolo_input_resolution": input_resolution_yolov3_HW}
     
165. 
     
166.             postprocessor = PostprocessYOLO(**postprocessor_args)
     
167. 
     
168.             # Run the post-processing algorithms on the TensorRT outputs and get the bounding box details of detected objects
     
169.             boxes, classes, scores = postprocessor.process(trt_outputs, (shape_orig_WH))
     
170.             if boxes is None:
     
171.                 pass
     
172.             else:
     
173.                 for box, score, category in zip(boxes, scores, classes):
     
174.                     x_coord, y_coord, width, height = box
     
175.                     left = max(0, np.floor(x_coord + 0.5).astype(int))
     
176.                     top = max(0, np.floor(y_coord + 0.5).astype(int))
     
177.                     right = min(image_raw.width, np.floor(x_coord + width + 0.5).astype(int))
     
178.                     bottom = min(image_raw.height, np.floor(y_coord + height + 0.5).astype(int))
     
179.                     cv.rectangle(frame,(left,top),(right,bottom),(255,0,0),thickness=2)
     
180.                     cv.putText(frame,"%s:%.2f"%(ALL_CATEGORIES[category],score),(left, top - 12),cv.FONT_HERSHEY_SIMPLEX,0.7,(0,0,255),2,0)
     
181.                     cv.imshow("obj_detected_img",frame)
     
182. 
     
183.             c = cv.waitKey(20)
     
184.             if c==27:
     
185.                 break
     
186.         else:
     
187.             pass
     
188.         
     
189. if __name__ == '__main__':
     
190.     main()
     

复制代码

你好楼主，请问你这个问题解决了吗？我和你相同的问题。