# YOLOv4_tensorflow

**Repository Path**: zhang-yu-1/YOLOv4_tensorflow

## Basic Information

- **Project Name**: YOLOv4_tensorflow
- **Description**: YOLOv4的tensorflow实现. implement yolov4 with tensorflow
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 1
- **Created**: 2021-01-05
- **Last Updated**: 2021-01-05

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# YOLOv4_tensorflow
yolov4的tensorflow实现. <br/>
Implement yolov4 with tensorflow<br/>
持续更新<br/>
continuous update the code<br/>

## 二战中科院计算所失败，求老师调剂收留
## I want to be a graduate student.
277118506@qq.com<br/>
rdc01234@163.com<br/>

## 使用说明
## introductions
执行命令. <br/>
run the following command.
```
python val.py
```
如果没有报错, 就没问题<br/>
if have no error, it's ok

### 转换 yolov4.weights (没有成功)
### convert yolov4.weights to fit our code(i failed )
参考[这个权重转换文件](https://github.com/wizyoung/YOLOv3_TensorFlow/blob/master/convert_weight.py), 我将 yolov4.weights 转换到了自己的代码中
执行命令<br/>
refer to [this weights convert file](https://github.com/wizyoung/YOLOv3_TensorFlow/blob/master/convert_weight.py), i converted yolov4.weights to this project.<br/>
<br/>
将下载好的 yolov4.weights 放到 yolo_weights 文件夹下, 执行命令<br/>
put yolov4.weights into the "yolo_weights" folder, and run the command.
```
python convert_weight.py
python test_yolo_weights.py
```
会在 yolo_weights 文件夹下生成权重文件<br/>
the ckpt weights file wound exits in the 'yolo_weights' folder<br/>
<br/>
权重转换是能够转换过来, 但是并不能跑出效果, 我确定全部权重都存放到了自己的项目中，排查了一天还是不知道哪里不对,我猜测是卷积层加载数据的顺序错了, 希望有同志能够帮忙解决这个问题, 谢谢了。<br/>
i'm sure that the weights file convertion to the ckpt is successful, but maybe the order of conv layer in this code is different with the yolov4, i worked for this for a day, still have no idea about it, I hope comrades can give me some help, thanks.<br/>
<br/>
weights_name.txt 文件中存放的是图模型的卷积层和bn的名字<br/>
the weights_name.txt contains all model layer's name of the network <br/>
<br/>
还有很多策略没有实现，以后会慢慢更新</br>
there are many strategies that have not been implemented, and will be updated slowly in the future.</br>
<br/>

### CIOU损失进行训练失败
### use CIOU_LOSS in our network failed
YOLO.py -> __get_CIOU_loss 函数是计算 ciou 损失的函数,在这个函数的最后面注释部分是参考[这个链接](http://bbs.cvmart.net/topics/1436)做的,函数的主体部分是参考 darknetv4 中 box.c -> dx_box_iou 函数实现的</br>
YOLO.py -> __get_CIOU_loss function is to compute ciou loss , and the comment section at the end of this function is refer to [this page](http://bbs.cvmart.net/topics/1436), others is refer to  box.c -> dx_box_iou function in darknetv4 </br>
</br>
在 YOLO.py -> __compute_loss_v4 函数中调用了损失的计算, 在第534 行是使用的yolov3的损失函数, 在第560行是使用的 CIOU_LOSS ,但是使用 CIOU_LOSS 的损失是 nan, 还在排查原因</br>
in function of YOLO.py -> __compute_loss_v4 , called the loss compute function, and in line 534 is the loss function of yolov3 used, in line 560 is the loss function of CIOU_LOSS used, but the loss of using CIOU_LOSS is NAN, and the cause of the error  is still under research.

### 训练自己的数据集
### train with own dataset
./data/JPEGImages 文件夹中存放用labelme标注json文件的jpg图片和对应的json文件, 参考我给的文件夹<br/>
The jpg image and the corresponding json file which marked with 'labelme' are stored in the folder "./data/JPEGImages", just like what I do<br/>
<br/>
然后在 ./data 文件夹下执行 python 命令, 会自动产生 label 文件和 train.txt 文件<br/>
and then, go to the folder "./data", execute the following python command, it automatically generates label files and train.txt
```
python generate_labels.py
```
继续执行命令,得到 anchor box<br/>
excute the python command, to get anchor box
```
python k_means.py
```
<br/>
打开 config.py, 将得到的 anchor box 写入到第六行，就像这样<br/>
anchors = 12,19, 19,27, 18,37, 21,38, 23,38, 26,39, 31,38, 39,44, 67,96<br/>
open config.py, write the anchor box to line 6, just like this<br/>
<br/>
所有的配置参数都在 config.py 中，你可以按照自己的实际情况来修改<br/>
all configuration parameters are in the config.py, you can modify them according to your actual situation<br/>
<br/>
配置完成,执行命令<br/>
ok, that's all, execute the command

```
python train.py
```
<br/>
训练完成后,将测试图片放到 test_pic 文件夹下,执行命令<br/>
put test images into test_pic folder, and run the following command after the model training.<br/>

```
python val.py
```
<br/>
这是我用54张图片训练了 2000 步(十分钟)的结果，效果还不错<br/>
this image is the result of training 2000 steps (10 minutes) with 54 pictures, it looks not bad. <br/>
![image](https://github.com/rrddcc/YOLOv4_tensorflow/blob/master/save/62.jpg)

### 有关 config.py 和训练的提示
### some tips with config.py and train the model
config.py 中的 width 和 height 应该是 608，显存不够才调整为 416 的<br/>
the parameters of width and height in config.py should be 608, but i have not a powerful GPU, that is why i set them as 416<br/>
学习率不宜设置太高<br/>
learning rate do not set too large<br/>
<br/>

### 自己的设备
### my device
GPU : 1660ti (华硕猛禽) 6G<br/>
CPU : i5 9400f<br/>
mem : 16GB<br/>
os  : ubuntu 18.04<br/>
cuda: 10.2<br/>
cudnn : 7<br/>
python : 3.6.9<br/>
tensorflow-gpu:1.14.0<br/>
numpy : 1.18.1<br/>
opencv-python : 4.1.2.30<br/>