search

hustvl / CrossVIS

[ICCV 2021] Crossover Learning for Fast Online Video Instance Segmentation
https://arxiv.org/abs/2104.05970
Other
85 stars 4 forks source link

Question about codes of loss #10

Closed samsara-ku closed 2 years ago

samsara-ku commented 2 years ago

Hi, I'm wondering the meaning of codes in this figure. These codes are in the python file adet/modeling/condinst/crossvis.py.

image

When I actually run your training code, there is no update with sem_losses_0, proposal_losses_0 and mask_losses_0 due to updating with losses dict, not the losses_0.

So based on this situation, I want to ask you two questions:

1) Did you intend this kind of logic, updating with only one frame? 2) Although I didn't see all iteration of training, based on some iteration, sem_losses was empty dict. Is it correct?

vealocia commented 2 years ago

Hi, @samsara-ku. Thanks for your interest in our work. For the first question, this is absolutely a mistake in our core rearrangement. Thanks for Thanks for pointing this out and I will update the correct code now. For the second question, the semantic loss is introduced by CondInst and I suggest you to follow the config here.