spytensor / detect_steel_bar Goto Github PK

View Code? Open in Web Editor NEW

172.0 3.0 60.0 359 KB

CCFDF AI 数钢筋大赛

Python 89.37% Shell 0.58% Cuda 3.59% C 6.27% C++ 0.18%

retinanet pytorch detection gangjin

detect_steel_bar's Introduction

提醒

由于个人版本修改较多，记不清哪个配置可以达到线上0.97。

声明

由于竞赛平台协议有规定，未经许可暂不能公开数据集，非常抱歉。

可修改的地方:

dataloader.py 中Resize函数的图像尺寸，越大效果越好。
main.py 中的模型 depth ，越大效果越好。
更有效的办法就是做数据扩充，增加数据量。

1. 比赛地址

智能盘点—钢筋数量AI识别

2. 依赖

pytorch0.4.1,opencv-python,skimage

3. 使用方法

step0: 安装

git clone https://github.com/spytensor/detect_steel_bar.git
detect_steel_bar/retinanet/lib/
bash build.sh
cd ../../

step1: 下载数据解压后，将训练数据和测试数据放到 data/images/下，效果如下：

 - data/
    - images/
        train/
        test/

step2: 将训练标签文件 train_labels.csv 复制到 data/ 下，效果如下：

- data/
    train_labels.csv

step3: 将官方提高数据转变成可供Retinanet训练格式

cd data
python convert.py
cd ..

step4: 训练

python retinanet/main.py

step5: 预测

python retinanet/predict.py

4. 效果

线上 0.97+

5. 参考

pytorch-retinanet

6. 提醒

如有疑问，请提出 issue，编码问题请自行谷歌。

detect_steel_bar's People

Contributors

Stargazers

Watchers

detect_steel_bar's Issues

请问在voc数据集上测试过这代码的性能吗？

请问在voc数据集上测试过这代码的性能吗？。。。

something about data enhancement processing

hi，big old！How to do some data enhancement processing on the training picture, I found that increasing the size of the input picture will also lead to missed detection or misdetection.There are some fake targets that are really no different from real goals. This is really difficult to deal with.

您好，这个文件是怎么生成的，怎么找不到？

Traceback (most recent call last):
File "retinanet/predict.py", line 26, in
results_file = open("./submit/larger%s.csv"%str(threshold),"w")
FileNotFoundError: [Errno 2] No such file or directory: './submit/larger0.5.csv'

撸完串数签子该怎么迁移

batch_size and collater

thanks for your work. when I try to figure out it, I have some probloms about this. could you help me to explain these?
1.why the bath_size is set to 1, it seems too small
2.it seems to spend too much time in collater process when loading train data. so the GPU is waiting this GPU-Util is very slow . do you have good idea to solve it?

CUDA error: out of memory？

我用的P100的卡，做了数据增强，出现这个error，CUDA error: out of memory。有啥解决办法不？
@spytensor

检测精度还能再提升吗

关于nms.so的问题

你好，我的Linux原来的cuda版本是10.0的。然后我在conda的虚拟环境里面安装好了cuda9.0版本、torch0.4.1版本。可是在运行main.py的时候，还是提示错误：ImportError : detect_steel_bar-master/retinanet/lib/nms/_ext/nms/_nms.so: undefined symbol: __cudaPopCallConfiguration
我也不知道是为什么了，明明cuda版本已经对应好了呀。

数据集

请问还有数据集么，可以私发我一份么，官网上已经下载不到数据集了，感激不尽！[email protected]

TypeError: dist must be a Distribution instance

when I input "bash build.sh", it occurs this error:
Compiling nms kernels by nvcc...
Including CUDA code.
/home/ljk/PycharmProjects/detect_steel_bar-master/retinanet/lib/nms
Traceback (most recent call last):
File "build.py", line 31, in
extra_compile_args=['-std=c99']
File "/home/ljk/miniconda3/envs/pt04/lib/python3.6/site-packages/setuptools/init.py", line 163, in init
_Command.init(self, dist)
File "/home/ljk/miniconda3/envs/pt04/lib/python3.6/distutils/cmd.py", line 57, in init
raise TypeError("dist must be a Distribution instance")
TypeError: dist must be a Distribution instance
what I should do?Thanks

problem about convert.py

当我运行convert.py的时候，会报错：FileNotFoundError: [Errno 2] File b'train_labels.csv' does not exist: b'train_labels.csv'。
但是train_labels.csv是在当前路径下的，作者有遇到过这个问题吗？

关于 predict.py

predict.py文件里的写结果到文件的时候str(int(box[1])) + " " + str(int(box[0])) + " " + str(int(box[3])) + " " +str(int(box[2])， box为什么不按照str(int(box[0])) + " " + str(int(box[1])) + " " + str(int(box[2])) + " " +str(int(box[3])来写，我的理解是box输出的格式不是(x_min,y_min,x_max,y_max)吗？