jetson nx跑yolov3_NX 跑 int8 yolov341630ms

　　最近研究tensorrt 跑 yolov3-416简单记录。下载yolov3的cfg和weights

修改cfg

vim yolov3.cfgbatch = 1width = 416height = 416

并且每一段只有一个空格。

修改yolov3_to_onnx.py  因为我们是４１６

output_tensor_dims['082_convolutional'] = [255, 13, 13]    [255, 19, 19]output_tensor_dims['094_convolutional'] = [255, 26, 26]    [255, 38, 38]output_tensor_dims['106_convolutional'] = [255, 52, 52]    [255, 76, 76]

其中 255 =(80 + 5)* 3  80为类别，5为输出 x, y, h, w, score

修改 weights_file_path = './yolov3.weights'

修改 output_file_path = './yolov3-416.onnx'

修改 cfg_file_path = './yolov3-416.cfg'

在ＮＸ盒子上把onnx转换成 in8 trt格式模型。

cd /usr/src/tensorrt/samles/python/yolov3_onnx/tensorrt_demos-master/yolo

sudo vim onnx_to_tensorrt.py修改 engine_path = './%s.trt'%args.model修改 onnx_path = './%s.onnx'%model_name修改 network.get_input(0).shape = [1, 3, 416, 416]  # [1, 3, 608, 608]

修改 EXPLICIT_BATCH = [] if trt.__version__[0] < '7' else \        [1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)] 注释掉改为：EXPLICIT_BATCH = 1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)修改 builder.create_network(*EXPLICIT_BATCH) as network改为 builder.create_network(EXPLICIT_BATCH) as network

修改 builder.int8_calibrator = YOLOEntropyCalibrator(                    'calib_images', (net_h, net_w), 'calib_%s.bin' % model_name)改为 calib_images 为量化图片地址。

创建一个量化图片文件夹

mkdir calib_imagessudo ln -s yolov3-416.cfg yolov3-int8-416.cfgsudo ln -s yolov3-416.onnx yolov3-int8-416.onnx

NX盒子转换 onnx 转换 int8

sudo python3 onnx_to_tensorrt.py -v --int8 -m yolov3-int8-416  # -c 为检测类别数 默认80 --dla_core 设置dla卡

cd /usr/src/tensorrt/samles/python/yolov3_onnx/tensorrt_demos-mastersudo vim trt_yolo_video_on_line.py

修改 def loop_and_detect 全部内容　测试本地视屏fps

def loop_and_detect(cam, trt_yolo, conf_th, vis):          output_path = './tensorrt_yolov3_int8_car.mp4'          full_scrn = False          writeVideo_flag = True          if writeVideo_flag:                vid = cv2.VideoCapture(cam)                if not vid.isOpened():                     raise IOError("Couldn't open webcam or video")                video_FourCC = cv2.VideoWriter_fourcc('m', 'p', '4', 'v')                video_fps       = vid.get(cv2.CAP_PROP_FPS)                video_size      = (int(vid.get(cv2.CAP_PROP_FRAME_WIDTH)),int(vid.get(cv2.CAP_PROP_FRAME_HEIGHT)))                isOutput = True if output_path != "" else False                if isOutput:                    #print("!!! TYPE:", type(output_path), type(video_FourCC), type(video_fps), type(video_size))                    out = cv2.VideoWriter(output_path, video_FourCC, video_fps, video_size)                    #list_file = open('detection.txt', 'w')                    frame_index = -1          while True:              return_value, frame = vid.read()              if return_value != True:                  break              if return_value:                  # frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)                  image = Image.fromarray(frame)              else:                   raise ValueError("No image!")              tic = time.time()              boxes, confs, clss = trt_yolo.detect(img, conf_th)              img = vis.draw_bboxes(img, boxes, confs, clss)              toc = time.time()              curr_fps = 1.0 / (toc - tic)              info = "time:" + str(round( (toc - tic), 2)) + " ms, FPS: " + str(curr_fps)              cv2.putText(result, text=info, org=(50, 70), fontFace=cv2.FONT_HERSHEY_SIMPLEX,fontScale=1, color=(255, 0, 0), thickness=2)              #cv2.namedWindow("result", cv2.WINDOW_AUTOSIZE)              if writeVideo_flag:                  # save a frame                  out.write(result)              #cv2.imshow("result", result)              if cv2.waitKey(1) & 0xFF == ord('q'): break

修改  cam = Camera(args)      if not cam.isOpened():          raise SystemExit('ERROR: failed to open camera!') 注释掉

修改 open_window(           WINDOW_NAME, 'Camera TensorRT YOLO Demo',           cam.img_width, cam.img_height) 注释掉

修改 cam.release()        cv2.destroyAllWindows()注释掉

添加 arser.add_argument(        '-video', '--video_path', type=str, default='',        help='video_path') 添加在线视屏路径

测试yolov3-int8-416 在tensorrt上运行时间

 sudo python3 trt_yolo_video_on_line.py  -m /usr/src/tensorrt/samples/python/yolov3_onnx/tensorrt_demos-master/yolo/yolov3-int8nx-416 -video ./car.mp4 # -c 检测类别数目，默认为80

NX量化C++引擎　ｅｎｇｉｎｅ　直接把onnx转化成engine格式

 sudo ./trtexec --onnx=./yolov3-416-1.onnx  --int8 --batch=1 --device=0 --saveEngine=./yolov3-416-gpu-int8.engine --verbose=True --allowGPUFallback --calib=/usr/src/tensorrt/samples/python/yolov3_onnx/tensorrt_demos-master/yolo/calib_images/*

可能报错：

/bin/sh: 1: nvcc: not foundMakefile:48: recipe for target 'yolo_layer.o' failedmake: *** [yolo_layer.o] Error 127

解决　找到nvcc 可执行文件路径

 cd /tensorrt_demo-master/plugins vim Makefile NVCC = /usr/local/cuda-10.2/bin/nvcc

遇见一个ubuntu16.04网络错误

解决方案

vim /etc/network/interfaces

ubuntu16.04 插入Ｕ盘报错：

解决方案：

sudo apt-get install exfat-fuse exfat-utils