準備 data.yaml 文件

從roboflow 上下載 CS 游戲數據集, 因為只有CPU, 我對數據集做了瘦身, train: 689張, val: 23張, test:40張.
https://universe.roboflow.com/roboflow-100/csgo-videogame/dataset/2

train: ../train/images
val: ../valid/images
test: ../test/images

nc: 2
names: ['CT', 'T']

roboflow:
  workspace: roboflow-100
  project: csgo-videogame
  version: 2
  license: CC BY 4.0
  url: https://universe.roboflow.com/roboflow-100/csgo-videogame/dataset/2

訓練

# 使用 yolov8n.pt 預訓練模型進行 train, 不含val
.\yolo task=detect mode=train val=False data=D:\my_workspace\data.yaml model=yolov8n.pt epochs=115 workers=1 imgsz=640

# 或者使用 yolov8n.pt 預訓練模型進行 train+val
.\yolo task=detect mode=train val=True data=D:\my_workspace\data.yaml model=yolov8n.pt epochs=115 workers=1 imgsz=640

val

.\yolo task=detect mode=val split=val data=D:\my_workspace\data.yaml model=path\weights\best.pt workers=1 imgsz=640

命令行輸出:

下圖val過程的預測結果, 第一個圖片沒有識別出來, 說明train epoch還不夠.

test

.\yolo task=detect mode=val split=test data=D:\my_workspace\data.yaml model=path\weights\best.pt workers=1 imgsz=640

predict

# predict 單張圖片
.\yolo predict model=path\weights\best.pt conf=0.25 source='D:\my_workspace\my.jpg'

# predict 整個目錄
.\yolo predict model=path\weights\best.pt conf=0.25 source='D:\my_workspace' 

參考

yolov8 命令參數中文: https://blog.csdn.net/weixin_45921929/article/details/128673338