Closed zhangchbin closed 3 years ago
And I attach the results file.
A problem in your configuration may be that you have a larger amount of GPUs than me (4 vs 2) and a lower batch size (6 vs 12). (1) the optimal learning rate may not be the same, and (2) the batch norm statistics may be poorer for you because we don't sync the statistics across GPUs.
I'm rerunning on my side those two experiments and will come back to you when I have results. But my config will be with batch_size=12 and nb_gpus=2.
A problem in your configuration may be that you have a larger amount of GPUs than me (4 vs 2) and a lower batch size (6 vs 12). (1) the optimal learning rate may not be the same, and (2) the batch norm statistics may be poorer for you because we don't sync the statistics across GPUs.
I'm rerunning on my side those two experiments and will come back to you when I have results. But my config will be with batch_size=12 and nb_gpus=2.
Hi, we conducted experiments again. We apply the standard-setting described as you, batchsize=12, gpus=2. The results are: mIoU = 34.66, 58.86, described as above. I have updated the issue.
My logs were for 1-1:
INFO:rank0:
Total samples: 695.000000
Overall Acc: 0.921154
Mean Acc: 0.526491
FreqW Acc: 0.860704
Mean IoU: 0.452352
Class IoU:
class 0: 0.96692383
class 1: 0.60595274
class 2: 0.8396912
class 3: 0.25620675
class 4: 0.15220715
class 5: 0.06744852
class 6: 0.0029509393
class 7: 0.32076913
class 8: 0.8602092
class 9: 0.4685076
class 10: 0.8987206
class 11: 0.57019883
class 12: 0.14706148
class 13: 0.8104589
class 14: 0.22377394
class 15: 0.418797
class 16: 0.5558119
class 17: 0.014650934
class 18: 0.41433895
Class Acc:
class 0: 0.9843136
class 1: 0.7646799
class 2: 0.9475544
class 3: 0.32678273
class 4: 0.16397442
class 5: 0.07432122
class 6: 0.002961872
class 7: 0.39156127
class 8: 0.9406667
class 9: 0.55477184
class 10: 0.95123327
class 11: 0.72136724
class 12: 0.1578011
class 13: 0.92616725
class 14: 0.23514673
class 15: 0.64418685
class 16: 0.59930676
class 17: 0.014668353
class 18: 0.6018618
INFO:rank0: Closing the Logger.
Last Step: 20
Final Mean IoU 45.24
Average Mean IoU 39.93
Mean IoU first 45.24
Mean IoU last 0.0
and for 11-1 (incorrectly called 10-1 in the code):
Total samples: 695.000000
Overall Acc: 0.940431
Mean Acc: 0.706060
FreqW Acc: 0.892914
Mean IoU: 0.620541
Class IoU:
class 0: 0.9772094
class 1: 0.70418805
class 2: 0.87226355
class 3: 0.43371245
class 4: 0.48649183
class 5: 0.1728502
class 6: 0.32025278
class 7: 0.4916692
class 8: 0.8817237
class 9: 0.5381132
class 10: 0.9231286
class 11: 0.66175914
class 12: 0.44145837
class 13: 0.8645326
class 14: 0.59335023
class 15: 0.69356126
class 16: 0.7649851
class 17: 0.39857766
class 18: 0.5704467
Class Acc:
class 0: 0.9894886
class 1: 0.82313305
class 2: 0.9535187
class 3: 0.55524725
class 4: 0.5560459
class 5: 0.1998001
class 6: 0.3846253
class 7: 0.5757105
class 8: 0.94679874
class 9: 0.6382271
class 10: 0.96305674
class 11: 0.77320135
class 12: 0.5875096
class 13: 0.9436415
class 14: 0.66980577
class 15: 0.7844055
class 16: 0.84219396
class 17: 0.5033585
class 18: 0.7253713
INFO:rank0: Closing the Logger.
Last Step: 10
Final Mean IoU 62.05
Average Mean IoU 62.37
Mean IoU first 62.05
Mean IoU last 0.0
Nevertheless, I'm re-running once again my code. I'll keep you updated when I have results.
Yeah, hope for your reply. By the way, we also trained the 11-5 (3 tasks), mIoU = 60.24
Hi, @arthurdouillard Is there any progress now?
I have trouble reproducing my results for Cityscapes-domain. I'm not sure why since I have logs (see my previous message) showing my reported performances.
I have updated the scripts for Cityscapes Domain with better results, sligthly above MiB and much higher than ILT, but still a bit lower than paper results.
For 11-1:
Final Mean IoU 60.48
Average Mean IoU 59.33
Mean IoU first 60.48
Mean IoU last 0.0
cityscapes_domain_11-1_PLOP_backbone-0.0001-logits-0.0001 On GPUs 0,1
Run in 17781s
For 1-1:
Final Mean IoU 43.15
Average Mean IoU 34.59
Mean IoU first 43.15
Mean IoU last 0.0
cityscapes_domain_1-1_PLOP_backbone-0.0001-logits-0.001 On GPUs 0,1
Run in 17534s
If I find a fix I'll update the code base.
Hi, @arthurdouillard , Could you please re-open this issue? If we have some questions, we can ask for your help in this page.
Hey, I prefer to have atomic issues. For more questions please open a new issue. Thanks!
Hi, @arthurdouillard We have re-cloned the repo and re-trained the model on cityscapes, but we still cannot reproduce these experiments. We have trained the model on the 1-1 setting, and got the mIoU 34.66 (45.24 in your paper). And we also re-trained the model on the 11-1 setting, and got the mIoU 58.85 (62.05 in your paper). So could you please check the training setting again?