search

WongKinYiu / yolov7

Implementation of paper - YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors
GNU General Public License v3.0
13.21k stars 4.17k forks source link

Can anyone tell me why I spend so much time on v7 training and get nothing #206

Open lixiangyussss opened 2 years ago

lixiangyussss commented 2 years ago

i do not know why，in my dataset, V7 is not comparable with any model of YOLO series V3-V6. The convergence rate of V7 and V7-tiny I used is very slow over 300 epochs, only 0.1map after 100 rounds, and the accuracy will vary a lot even at the end of training，whether I use pretrained weights or not

M15-3080 commented 2 years ago

What is your run command?

WongKinYiu commented 2 years ago

If your GPU has no tensor core, you have to set half to False.

lixiangyussss commented 2 years ago

如果您的 GPU 没有张量核心，则必须将 half 设置为 False。

You are referring to when training, or in the test.py file, I set half_precision=False in my test.py.

lixiangyussss commented 2 years ago

你的运行命令是什么？

![Uploading image.png…]()，i use python train.py

M15-3080 commented 2 years ago

You should have the same mistake as me, please use train_aux.py for large models

sontung2310 commented 2 years ago

@lixiangyussss Did you solve this problem? I am training YOLOv7 with custom dataset but it takes so much time and the convergence is extremely slow.

way2016 commented 2 years ago

I encountered same issues as @sontung2310 . It took around 1 hour per epoch，and the evaluation mAP=0.293 after 143 epochs (still training slowly...). I used to train with same custom dataset using YOLOv5x，which took around 2 days completing 300 epochs with mAP ~ 0.4. I wonder if there are some settings got wrong.