# Yolov4 Setting and How to train

###### tags: `YOLOv4`

:::info
:pushpin: 參考網址 ➜ [PyTorch Yolov4配置](https://medium.com/@whchen0630/pytorch-yolov4%E9%85%8D%E7%BD%AE-8947eb17b547)

:pushpin: Pytorch_yolov4 ➜ [WongKinYiu/PyTorch_YOLOv4](https://github.com/WongKinYiu/PyTorch_YOLOv4)

:pushpin: v4流程教學 ➜ [TSR_交通號誌辨識系統 教育訓練](https://hackmd.io/@Hsiang1118/ry5b--1q5)
:::

### 創建虛擬環境（使用Anaconda）

conda create --name xxx python="your version"

* **進入虛擬環境**

conda activate xxx

### Yolov4 環境配置

* **requirements.txt 設定**

numpy == 1.21.5
opencv-python >= 4.1
torch == 1.12.1
torchvision == 0.13.1 
matplotlib
pycocotools
tqdm
pillow
tensorboard >= 1.14

> 其餘直接手動下載 ex ： (pyyaml)
* 安裝requirements.txt 指令

pip install -r requirements.txt

* Ubuntu環境之pycocotools請參照以下指令

sudo apt-get installl Cython
pip install cython
git clone https://github.com/cocodataset/cocoapi
cd PythonAPI

* 若使用過程中有套件遺失，在使用pip install指令下載

### Yolov4 Training Steps

1. git clone https://github.com/WongKinYiu/PyTorch_YOLOv4.git
2. Install requirements.txt **pip install -r PyTorch_YOLOv4/requirements.txt**
3. Edit yolov4.cfg (Change [yolo] classes numbers && [convolutional] Filters numbers)
* Filters numbers = $(classes+5)*3$
4. Create xxx.yaml in /data and input your path and nc(number of classes)

# 設定圖檔路徑
train: images/train     # 訓練用資料集（相對於 path）
val: images/train       # 驗證用資料集（相對於 path）
test:                   # 測試用資料集（相對於 path，可省略）

# 物件類別設定
nc: 1           # 類別數量
names: ['egg']  # 類別名稱

* 資料路徑擺放位置

yolov4
├── data
│   └── ├──
│       ├── train.txt/val.txt/test.txt
│       ├── xxx.names(object names)
│       ├── xxx.yaml(img's txt path)
        ├── img file(include jpg and label txt)
        ├── hyp.scratch.yaml
│       └── img.cache3
└── weights
    └── xxx.weights
│
└── cfg
    └── xxx.cfg
│
└── xxx.yml(Need to be same with your conda env)
│
└──train.py/test.py/detect.py 
│
└── requirements.txt

5. creat xxx.names in /data and put your classes name in it
6. Run python code to cahnge JPG to txt file
(it will create train.txt/test.txt)
7. **Train your data** : python train.py --batch-size 8 --epochs 100 --img 640 640 --data data/xxx.yaml --cfg cfg/yolov4.cfg --weights weights/yolov4.weights
8. **Test your model** : python test.py --img 640 --conf 0.001 --batch-size 8 --device 0 --data data/bsw.yaml --cfg cfg/bsw.cfg --weights weights/last.pt --names data/bsw.names --name bsw
9. **Detect your source** : python detect.py --weight ../weights/best.pt --source data/xxx.avi --cfg cfg/xxx.cfg --names data/xxx.names

:::info
:pushpin: PS ➜ 以上weights / cfg / data字型自行調整
:::

### Error Solved
* OSError: [WinError 1455] 分頁檔太小，無法完成操作。
Problem : batch size is too big
Solved : reduce batch size

* RuntimeError: result type Float can‘t be cast to the desired output type Long
Solved : 於loss.py裡將原本的code更改為以下的code

anchors, shape = model.module.module_list[jj].anchor_vec if multi_gpu else model.module_list[jj].anchor_vec, p[i].shape

indices.append((b, a, gj.clamp_(0, shape[2] - 1), gi.clamp_(0, shape[3] - 1)))