Visual-Attention-Network/SegNeXt

about train?

Onehundred0906 opened this issue · 7 comments

i train my own datasets, and download the pretrained_weights of 'mscan_b.pth', can you explain to put the weights_file to which place? Thanks very much! my command is : python /tools/train.py local_configs/segnext/base/segnext.base.512x512.tusimple.160k.py --load-from pretrained/mscan_b.pth

but occur some error which is showed in the follow picture!
2022-10-25 13-46-35屏幕截图

You need to put the pre-trained weight under the corresponding folder.

You need to put the pre-trained weight under the corresponding folder.

And You need to load the pre-trained weight by changing local_configs/segnext/base/segnext.base.512x512.tusimple.160k.py like other configs.

i change the weight's path in code of' local_configs/segnext/base/segnext.base.512x512.tusimple.160k.py' like the first picture,
Screenshot 2022-10-25 155929
but when i train my dataset, the result is showed like the second picture , is very Irregular. emmm, thank you very much for bothering you for my question !
2022-10-25 15-22-26屏幕截图

my data has only 2 classes(background and road_line), in my opinion, the dataset_config should be correct! thanks very much.

If you modify your dataset according to ade20k, you may ignore the label 0.

ooo, thanks, maybe this is a cause, but if i need to consider the background based the dataset ade20k, and i can modify where code ?

thanks i have found the error and it has been trained normally, the reduce_zero_label should be False in my dataset config