woshiwwwppp / ryolov3research-pytorch-master Goto Github PK

View Code? Open in Web Editor NEW

57.0 3.0 18.0 17.19 MB

An oriented bounding boxes implement of YOLOv3 with SEnet , Deformable Convolution and rotated NMS written in CUDA

Python 92.46% C++ 0.38% Cuda 7.16%

ryolov3research-pytorch-master's Introduction

An oriented bounding boxes implement of YOLOv3 with SEnet , Deformable Convolution and rotated NMS written in CUDA.

enviroment： ubuntu16.04 pytorch1.1

requirement： numpy opencv matplotlib tensorboardX shapely

install:

cd ./yolo/nms

python setup.py build_ext --inplace

usage:

1.make your own dataset by labeImage https://github.com/woshiwwwppp/ryolov3research-pytorch-master/blob/master/labelImg.exe

2.use generate_data.py to get label file

3.set the parameter in object_classes_all.txt and config.py

3.use train.py to train

ryolov3research-pytorch-master's People

Contributors

Stargazers

Watchers

Forkers

lighttoyang baltam chwwade baifanysu mentorezio chunxiaoni clw5180 sdhmdhr123 zkk931031 wj1214 sandhyacs sharpiless robertxiaobinw yueyang07 jack16888 voscar-zhang terryfic5 linyuanxs

ryolov3research-pytorch-master's Issues

跑程序的问题

很佩服您的代码能力，这么麻烦的程序写的很随性漂亮。真的很厉害！！（膜拜大佬）
最近想跑一下程序检测斜框目标，有几个疑问想请教一下：

valid的一个bug:

ryolov3research-pytorch-master/yolo/utility.py

Lines 404 to 419 in 83cb5c2

 out_box, out_score, out_class = convert_yolo_outputs(out_puts, input_shape, ratio, anchors, 

 classes, confidence, NMS, CUDA=True) 

 for k, v in enumerate(val_lines[step * batch_size:(step + 1) * batch_size]): 

 gt_boxes = [] 

 gt_classes = [] 

 for gt in v.strip().split(' ')[1:]: 

 gt_boxes.append(list(map(float, gt.split(',')[:-1]))) 

 gt_classes.append(classes[int(gt.split(',')[-1].strip())]) 

 out_classes = out_class[k] 

 out_boxes = out_box[k] 

 # 计算ap 

 for i in range(len(out_classes)): 

 for j in range(len(gt_classes)): 

 if rbox_iou(gt_boxes[j], out_boxes[i]) > matchIou and gt_classes[j] == out_classes[i]: 

 precision[out_classes[i]].append(1) 

 break

上面代码处out_box的维度是(num_boxes,5)，但是后面遍历batch内图像时out_boxes = out_box[k]只取出一个box；导致计算AP遍历时的下面这句报错：

ryolov3research-pytorch-master/yolo/utility.py

Line 417 in 83cb5c2

 if rbox_iou(gt_boxes[j], out_boxes[i]) > matchIou and gt_classes[j] == out_classes[i]: 

因为out_boxes[i]是取出上面那个box的第一个坐标信息的标量，无法传入rbox_iou计算iou。不知道是不是我哪里的理解有误...==

速度
riou实现用的是shapely，没有cuda计算overlap，会不会检测的速度很慢？我自己写的斜框检测奇慢无比，请问大佬有用cuda试着加速嘛
test检测报错
注释掉eval的部分直接训练2000epoch，让程序过拟合一张图片，得到的权重用test.py进行检测，报错如下：

Traceback (most recent call last):
  File "test.py", line 40, in <module>
    classes,confidence = config.dataSet['confidence'],NMS = config.dataSet['NMS'],CUDA= True)
  File "/py/ryolo/yolo/utility.py", line 230, in convert_yolo_outputs
    cla_id = torch.argmax(res[:,6:],dim=1)
RuntimeError: cannot perform reduction function max on tensor with no elements because the operation does not have an identity

将conf阈值设置为0.002才能跳过这个错误，似乎是2k轮都没学到东西....
然后在此基础上，继续检测得到这里的bug:

ryolov3research-pytorch-master/test.py

Lines 44 to 45 in 83cb5c2

 for i,box in enumerate(out_boxes): 

 label = out_classes[0][i]

报错是IndexError: string index out of range，因为我是单类目标检测label都是ship,out_classes[0]='ship',进而out_classes[0][i]索引的有误了，请问这个是不是我的设置有问题导致的？

检测结果
避开上述问题，在输出侧将label直接赋值为'sihp'，训练3000epoches，loss基本下降到0附近不变，但是检测的结果都丢了：（我的anchor设置就是按照gt的尺寸设的，理论上不怎么回归就能达到极小值，角度也按大佬的设置为0-pi）
模型可视化
在train.py中用您提供的tensorboardX接口记录计算图模型时报错：

  File "train.py", line 97, in <module>
    w.add_graph(yolo, (dummy_input,))
  File "/home/bit530/anaconda3/envs/torch1.1/lib/python3.7/site-packages/tensorboardX/writer.py", line 738, in add_graph
    self._get_file_writer().add_graph(graph(model, input_to_model, verbose, **kwargs))
  File "/home/bit530/anaconda3/envs/torch1.1/lib/python3.7/site-packages/tensorboardX/writer.py", line 138, in add_graph
    graph = graph_profile[0]
TypeError: 'GraphDef' object is not subscriptable

上面五个问题希望能帮忙解答，因为想先用起来再慢慢看程序，所以有的地方还不是很明白，感谢解答～～

weights path?

which weights path is used here and could you please add more information on training I am not able to train.

dalao你好，上周我在使用你推荐的命令进行安装时，cmd中出现了以下错误：
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.0\include\crt/host_config.h(143): fatal error C1189: #error: -- unsupported Microsoft Visual Studio version! Only the versions between 2013 and 2017 (inclusive) are supported!
error: command 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.0\bin\nvcc.exe' failed with exit status 2
翻译了一下大概就是yolo不支持vs2019，之前我在编译darknet的时候也出现过这个问题，通过修改2019添加vs2015的工具箱成功编译了。但是对于dalao的这个repo，不知道在命令行中如何解决，如果dalao有空的话，希望可以告诉我一下解决方案~

	out_box, out_score, out_class = convert_yolo_outputs(out_puts, input_shape, ratio, anchors,
	classes, confidence, NMS, CUDA=True)
	for k, v in enumerate(val_lines[step * batch_size:(step + 1) * batch_size]):
	gt_boxes = []
	gt_classes = []
	for gt in v.strip().split(' ')[1:]:
	gt_boxes.append(list(map(float, gt.split(',')[:-1])))
	gt_classes.append(classes[int(gt.split(',')[-1].strip())])
	out_classes = out_class[k]
	out_boxes = out_box[k]
	# 计算ap
	for i in range(len(out_classes)):
	for j in range(len(gt_classes)):
	if rbox_iou(gt_boxes[j], out_boxes[i]) > matchIou and gt_classes[j] == out_classes[i]:
	precision[out_classes[i]].append(1)
	break