zhaowt61
zhaowt61
> By the way, does the batch size really matter? I train my model only ≈60 mIoU on the voc validation dataset with batch size 4. It is possible. If...
> I have tried this training settings, but it's not useful. My training dataset doesn't contain SDB. I can't get better mIoU. You have to contain SDB to achieve 77%+
By using the following settings, I managed to achieve 77.24% at 24 epochs. Setting: 1 dataset: Pascal + SDB 2 4 gpus, BatchSize = 16 3 Backbone: ResNet 4 —freeze-bn...
@SystemErrorWang why didn't I reproduce the result after 30k iterations by using exactly the same traning settings? Is it because each different initialization gives us different performance?
There are lots of stripes on human faces along with green strips on backgrounds at 30k iterations.
Hi Dr. Wang, thanks for your reply, could I possibly have you email for further communication? Thank you so much! @SystemErrorWang
@SystemErrorWang Thank you so much, Dr. Wang! 😄
The author doesn't know what "100w" means, instead, you should say "1000k". BTW, "w" means 10k in chinese.
I am also interested in the light mode, did you see any difference between the light mode and regular mode in terms of the quality?