search

alibaba / ClickSEG

Apache License 2.0
38 stars 42 forks source link

Recurrence #9

Open join-ville opened 2 years ago

join-ville commented 2 years ago

I trained COCO+LVIS SegFormerB0-S1 using the trainval_scripts/train_focalclickB0_S1_cclvs.sh with nothing changed. And the val result is as follows:

图片1

compared with the model published：

图片2

I notice that the BS is 32 in paper, and it's 64 in the bash script. Is it the reason for the accuracy gap？

Also, there are two ways of data augmentation in the code and the paper says "During training, we only use flip and random resize with the scale from 0.75 to 1.4 as data augmentation." which one is used in the model published?

XavierCHEN34 commented 2 years ago

Hi, have you tried training multiple times? 8.53 is not a typical result. If you keep getting results like that, there might be something wrong, maybe the version of some libs or envs.

we only use flip and random resize