valeoai / esl Goto Github PK

the test pretrained code can not reach miou 43.7 ,when i test,the pretrained codes the miou is 0.47
the code runining result:
File "test.py", line 99, in
main(args.cfg, args.exp_suffix)
File "test.py", line 92, in main
evaluate_domain_adaptation(models, test_loader, cfg)
File "/home/DATA/huyujie/think/ADVENT-master/advent/domain_adaptation/eval_UDA.py", line 29, in evaluate_domain_adaptation
verbose)
File "/home/DATA/huyujie/think/ADVENT-master/advent/domain_adaptation/eval_UDA.py", line 52, in eval_single
_, pred_main, _ = model(image.cuda(device))
ValueError: not enough values to unpack (expected 3, got 2)
when i modify the code pred_main = model(image.cuda(device))[1]
can running but the miou is 0.47

File "train.py", line 27, in
import valeodata
ModuleNotFoundError: No module named 'valeodata'
CAN you give me some proposes?

I am familar with your paper ADVENT and paper BDL's SSL, so I directly read your code. If my understanding is correct, the main

difference between esl and ssl is in below code:

x = predicted_entropy[predicted_label==i] vs x = predicted_prob[predicted_label==i]

thres[thres<cfg.ESL.THRESHOLD]=cfg.ESL.THRESHOLD vs thres[thres>0.9]=0.9

label[(entropy>thres[i])(label==i)] = 255 vs label[(prob<thres[i])(label==i)] = 255

Then I do a small experiment, I trained my unsupervised domain adaptation segmentation network, the obtained mIOU is 41.31,

then I use the 41.31 mIOU 's model to run SSL.py, use the pseudo to train the target domain with a supervised way, then reach

to 43.64 mIOU. Then, I modify the SSL mainly with upper code to ESL, to obtain the pseudo and train with target, however, the

mIOU is 42.68. Besides, the SSL method mainly about 43, 43.2 mIOU, and ESL method mainly about 41.8, 42 mIOU, even 41.4

mIOU. So I think its very strange for me. Do I do some wrong in modifying the code? Can you give me some suggestions?