Giters

Leaf48 / YOLO-Models-For-Valorant

Valorant Models in Yolov5 and Yolov7

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

No detections Found

AStrangePotato opened this issue · comments

commented

Hello, I trained 400 more epochs with the valorant-v10.pt weights using --img 736 --batch 24 --epochs 400 --data valorant.yaml --weights valorant-v10.pt --device 0 however, after training it seems to work worse than your weights. It does not detect anything when I load a torch model and show the results. The data I used was downloaded from your repository, so I don't know why this is. Do I need a higher batch size? Or more epochs? My GPU is a 1660ti so it cannot handle batch size more than 24. What settings did you use when training the model? Thanks so much!
image
image

commented

1.You had better increase batch-size. If possible, it would be better to set batch-size to 80 or more.
I encourage you to try yolov5s as weight.

2.Try using dataset labeled only "enemy".
Your dataset is included some images labeled not only "enemy" but also "head".
It may have confused your machine learning.

My env:
Weight: yolov5s or valorant-v8.pt(It's the best precise of models)
Epoch: 300 to 500
Batch-size: 80 or more
Google Colab Pro

Have fun with your machine learning 😄

commented

Thanks so much! I will try the dataset with only enemy labelled and valorant-v8.pt. 🥰 Just curious: How did you manage to have the v10 model detect enemy_head as well as enemy? Would I be able to label 'head' on your dataset and merge it with my current to have it detect both? Right now my class names are ['body', 'head'] so I am not sure if I have to name them enemy, enemy_head as long as the class values [0, 1] stay the same. (I can also give you the newly labelled data if it works to merge with valorant-v11.pt)