search

imlixinyang / HiSD

Code for "Image-to-image Translation via Hierarchical Style Disentanglement" (CVPR 2021 Oral).
Other
391 stars 49 forks source link

A few questions about training tricks #33

Closed zhushuqi2333 closed 2 years ago

zhushuqi2333 commented 2 years ago

Sorry to bother you, I'm doing some work on your model and would like to ask a few questions about training skills

  1. As for the number of training times, I combine the loss function with the process graph generated in the training process. When the loss function declines to be stable and the generated graph looks ok, it means that the training is done. Do you have any additional comments?

  2. Regarding the problem of model collapse, can you give some examples in this task?

  3. Do you have any skills in setting loss function coefficients?

Looking forward to your reply~

imlixinyang commented 2 years ago

Hi there.

  1. Due to the particularity of generation task, the visual quality should be the first one to concern about when you choose your best checkpoint.
  2. For example, in HiSD, the reference-guided extracted by the encoder sometimes can be only applied to the image itself.
  3. I think keeping the values of each loss in the same magnitude is fine enough in most situations.
zhushuqi2333 commented 2 years ago

Thank you for your reply~ On the second point, you mean, for example, there are two pictures x1 and x2. The s extracted from x1 can only act on x1, but not on x2，is that right？

imlixinyang commented 2 years ago

Yes, the s extracted from x1 can only act on x1 and its translated image ~x1, but not on x2. It satisfies the loss objective but not what we want.

zhushuqi2333 commented 2 years ago

I'm sorry to see the message so late. Thank you for your reply. May I ask what ALI is? Is there a paper on this?

imlixinyang commented 2 years ago

ALI is short for Adversarially Learned Inference from https://arxiv.org/abs/1606.00704.

zhushuqi2333 commented 2 years ago

Tnanks for your reply! I wish you every success in your scientific research!