Log for val results does not save correctly when IterBasedRunner is used.

[haotian-liu]

When I am using IterBasedRunner is used, and the evaluation runs during the training (triggered by evaluation hook evaluation=dict(interval=5000)), the eval results are not logged with val mode, and are logged as train mode. This is different from EpochBasedRunner.

Specifically, for EpochBasedRunner, training losses will be logged in Tensorboard with prefix train/ (e.g., train/loss, train/loss_cls, etc.), and evaluation results will be logged with prefix val/ (e.g., val/table, val/acc, val/miou, etc.). However, in IterBasedRunner, they are all logged with prefix train/ (e.g., train/loss, train/table, train/acc, train/miou), which is weird and causes troubles for visualization.

It seems that when evaluation hook is triggered in after_iter stage, the mode will not be changed to eval and remains as train instead, which might be the cause of the issue.

Please kindly advise if there is anything that I am not doing correctly, thank you so much!

[zhouzaida]

hi @haotian-liu , thanks for you feedback, we are fixing the bug

[haotian-liu]

Thank you so much!

[zhouzaida]

closed by #1252