What is the best parameter set to train using ade20k dataset?

[tan-z-tan]

Thanks for useful repository!

I'm trying to train PSPNet from scratch using ade20k dataset. It works by changing train.py like this:

DATA_DIRECTORY = '../data/ADEChallengeData2016/'
DATA_LIST_PATH = './list/ade20k_train_list.txt'
#IGNORE_LABEL = 255
IGNORE_LABEL = 0

#INPUT_SIZE = '713,713'
INPUT_SIZE = '473,473'

#NUM_CLASSES = 19
NUM_CLASSES = 150

However, the model i got after 10000 steps has poor performance. (average training loss reached around 2.5 from 5 to 6) I guess the parameter set written in train.py is set to train cityscapes dataset. What is the best parameter set to train using ade20k dataset?

[TingZhang08]

I also encounter this problem. Have you resolved?

[shipengai]

Hi，@tan-z-tan can you share your training on ade20k? thanks