search

open-mmlab / mmtracking

OpenMMLab Video Perception Toolbox. It supports Video Object Detection (VID), Multiple Object Tracking (MOT), Single Object Tracking (SOT), Video Instance Segmentation (VIS) with a unified framework.
https://mmtracking.readthedocs.io/en/latest/
Apache License 2.0
3.52k stars 591 forks source link

a question of training STARK_st1 #515

Open yuwenyouhaoran opened 2 years ago

yuwenyouhaoran commented 2 years ago

image when i train stark_st1 using config/sot/stark/stark_st1_r50_500e_got10k.py. why is Loss_iou always 0.0000?

JingweiZhang12 commented 2 years ago

Please provide your configs and the modifications of the code.

yuwenyouhaoran commented 2 years ago

only modified some path in offical configs/sot/stark/stark_st1_r50_500e_got10k.py such as:

line 84    data_root = 'D:/
line 157  'GOT-10k/annotations/got10k_train_infos.txt'

Other parameters are not modified

JingweiZhang12 commented 2 years ago

It's normal. The box prediction is invalid in the first few dozen iterations and can't be calculated in the iou loss, so we ignore them and set the loss_iou=0. Details can be seen here. Generally speaking, the loss_iou after 50 iterations will be normal.

yuwenyouhaoran commented 2 years ago

but after 10 epochs(>>50 iterations ) ,its loss_iou still equals 0

JingweiZhang12 commented 2 years ago

I just trained the model with batchsize=8 on a single GPU to make the number of iterations per epoch equal to 7500. The loss_iou after 50 iterations is greater than 0. It's OK.

yuwenyouhaoran commented 2 years ago

I also trained the model with batchsize=8 on a single GPU to make the number of iterations per epoch equal to 7500.

image the loss_iou always 0

parameters are not modified in configs

JingweiZhang12 commented 2 years ago

Maybe you can pull the latest repo and rerun it.