search

unibas-gravis / Occlusion-Robust-MoFA

Other
208 stars 30 forks source link

Step4 Problem #14

Closed NewtOliver closed 1 year ago

NewtOliver commented 2 years ago

Is it normal for me to have this situation in the last step of training the model? 96aabe4cb1b9e5dae98c58510cf9880

ChunLLee commented 2 years ago

Hi, we do not expect such results. Can you check the following factors to provide more information so that we could better diagnose?

  1. Is the pre-trained model from step 1 working well?
  2. Are the masks estimated by the UNet reasonable?
  3. In the first 5000 iterations of step 4, are the faces rendered with normal human face texture?
  4. If you run step 4 again, do such results occur?
NewtOliver commented 1 year ago

Hi, we do not expect such results. Can you check the following factors to provide more information so that we could better diagnose?

  1. Is the pre-trained model from step 1 working well?
  2. Are the masks estimated by the UNet reasonable?
  3. In the first 5000 iterations of step 4, are the faces rendered with normal human face texture?
  4. If you run step 4 again, do such results occur?

1.Step 1 is fine 2.I think the Pretrain_Unet is OK 3.In the first 5000 iterations of step 4, the network can normally display face texture, but the mask photo is all black, eg fig 4.Re-running step 4 still has the same problem, and the mask photo is all black from start to finish, Idon't know how to solve the problem image

hnuzhy commented 1 year ago

I was also troubled by this problem or bug. Have you solved this problem during training Step4 ?

NewtOliver commented 1 year ago

I was also troubled by this problem or bug. Have you solved this problem during training Step4 ?

Yes, I've solved. You should compare the code in the step4 with the code in the step1 and step3, and then fix it.

DougDong666 commented 1 year ago

I was also troubled by this problem. But the loss fuctions in step4 are different from step1 and step3, I don not think it can be fixed by changing loss functions in step3 and step1.

DougDong666 commented 1 year ago

Loss functions in step4 are followed by article. But the loss function in step3 is a classic segmentation loss function. If the loss functions are changed by step3, I don not know why the article needs to mention its own loss fuctions.

Jack171516 commented 6 months ago

Hi, can I have a look at your test_mask_gt.png image generated by Pretrain_Unet in step 3, I have a feeling that this image I generated might not be correct, which is causing me to train incorrectly in step 3. Thank you very much.