Is the training procedure result normal? Masked regions do not improve and appear to be random noise.

facebookresearch / mae

PyTorch implementation of MAE https//arxiv.org/abs/2111.06377

Other

6.93k stars 1.17k forks source link

Is the training procedure result normal? Masked regions do not improve and appear to be random noise. #190

Open junzhin opened 4 months ago

junzhin commented 4 months ago

Hello,

thank you for the great work and the great repo.

I attempted to reproduce the pre-training of MAE-ViT-Large, and performed 68 epochs on a chest x -rays dataset with about 300 w medical images, and the loss stopped improving when it reached around 0.0045 loss without pixlossnorm. Additionally, the reconstruction results fail to predict the masked regions correctly.

Could you suggest a reason for that? Any idea why this is the case?

CristoJV commented 3 months ago

Hi @junzhin,

I'm facing the same errors. output Pretrained on 500k faces for 20 epochs, resuming from the authors' pretrained version MAE-ViT-Base.

Did you manage to solve them?

EDIT: I fixed it by disabling the pixlossnorm.

Thank you!

ats4869 commented 1 month ago

I would like to ask how you use finetune to train the reconstruction model on your own data set. I see that through main_finetune.py only models for classification tasks can be generated.