1.数据格式指定VOC
/M2Det/train.py:20
parser.add_argument('-d', '--dataset', default='VOC', help='VOC or COCO dataset')
2. 删掉VOC2012
/M2Det/configs/m2det512_vgg.py:62
dataset = dict(
VOC = dict(
train_sets = [('2007', 'trainval')],
eval_sets = [('2007', 'test')],
),
# VOC = dict(
# train_sets = [('2007', 'trainval'), ('2012', 'trainval')],
# eval_sets = [('2007', 'test')],
# ),
3.更改数据的label类型
/M2Det/data/voc0712.py:27
# VOC_CLASSES = ( '__background__', # always index 0
# 'aeroplane', 'bicycle', 'bird', 'boat',
# 'bottle', 'bus', 'car', 'cat', 'chair',
# 'cow', 'diningtable', 'dog', 'horse',
# 'motorbike', 'person', 'pottedplant',
# 'sheep', 'sofa', 'train', 'tvmonitor')
4.指定图片格式
/M2Det/data/voc0712.py:170
self._imgpath = os.path.join('%s', 'JPEGImages', '%s.png')
5.指定数据路径
m2det512_vgg.py:81
VOCroot = os.path.join(home,"data/VOCdevkit/")
6.batch_size相关参数
/M2Det/configs/m2det512_vgg.py:30
7.下载 vgg16_reducedfc.pth权值文件
M2Det/configs/m2det512_vgg.py:5
地址:https://github.com/amdegroot/ssd.pytorch/issues/170
8.RTX2080 +pytorch+cuda报错问题
THCudaCheck FAIL file=/pytorch/aten/src/THC/THCGeneral.cpp line=383 error=11 : in Traceback (most recent call last)
这个问题暂时未解决(回头解决了,补充上来),换1080显卡调试,没有问题。
成功开始训练自己的voc2007数据