目标检测 | STDN配置训练笔记

最新推荐文章于 2024-09-15 14:31:32 发布

Nani_xiao

最新推荐文章于 2024-09-15 14:31:32 发布

阅读量867

点赞数

分类专栏：深度学习目标检测文章标签： STDN

本文链接：https://blog.csdn.net/xiao_lxl/article/details/85047285

版权

深度学习同时被 2 个专栏收录

34 篇文章 1 订阅

订阅专栏

目标检测

15 篇文章 2 订阅

订阅专栏

STDN配置

由于此配置运行在PyTorch环境下，首先保证配置好
PyTorch+python3

获取STDN安装文件,感谢arvention的分享。

git clone https://github.com/arvention/STDN

修改pasal_voc.py中，数据集的路径为自己电脑的存放路径

#self.data_path = data_path
self.data_path = '/media/****/****/data/VOCdevkit'

modify main.py

parser.add_argument('--voc_data_path', type=str,
                        default='/media/****/****/data/VOCdevkit/')

parser.add_argument('--coco_data_path', type=str,
                        default='/media/****/****/data/coco/')

开始训练

python3 main.py

报错：
[2000/120000]时报错 CUDA: out of memory.

将batch_size由32改为8，重新训练VOC数据

python3 main.py --batch_size 8

训练结果：

Elapsed 7:44:29.186617 -- 0:00:00.232243, Iter [120000/120000]
class_loss: 3.1876, loc_loss: 1.1356, loss: 4.3232

测试训练好的模型

 python3 main.py --mode test --voc_config 0712 --pretrained_model 2018-12-14\ 17_33_53.115576/120000

结果如下：

Reading annotation for 4801/4952
Reading annotation for 4901/4952
Saving cached annotations to ./results/2018-12-14 17_33_53.115576/120000/annotations_cache/annots.pkl
AP for aeroplane = 0.5854
AP for bicycle = 0.6386
AP for bird = 0.3809
AP for boat = 0.3736
AP for bottle = 0.1154
AP for bus = 0.6618
AP for car = 0.6590
AP for cat = 0.6445
AP for chair = 0.2785
AP for cow = 0.4271
AP for diningtable = 0.5347
AP for dog = 0.5829
AP for horse = 0.6880
AP for motorbike = 0.6726
AP for person = 0.5777
AP for pottedplant = 0.1458
AP for sheep = 0.4444
AP for sofa = 0.5690
AP for train = 0.7144
AP for tvmonitor = 0.4945
Mean AP = 0.5094
~~~~~~~~
Results:
0.585
0.639
0.381
0.374
0.115
0.662
0.659
0.645
0.279
0.427
0.535
0.583
0.688
0.673
0.578
0.146
0.444
0.569
0.714
0.495
0.509
~~~~~~~~

Reference：
[1] https://blog.csdn.net/u014380165/article/details/80602130
[2] https://github.com/arvention/STDN