search

facebookresearch / mae

PyTorch implementation of MAE https//arxiv.org/abs/2111.06377
Other
6.93k stars 1.17k forks source link

Loss is considerably worse on custom data set with different mean and standard deviation #179

Open bpmsilva opened 9 months ago

bpmsilva commented 9 months ago

I am using this repo to train on a custom data set. My images are grayscale, but the final reconstruction was bluish. I hypothesized that was because of the use of the Imagenet mean and standard deviation (blue is the one with the lowest mean value). I proved that correct because I changed the mean and standard deviations of the code, and the final images were gray. However, it came with the cost of lowering my performance as the loss became higher than usual. I even tried to change the mean and standard deviation of the timm package, but that was no good. Is there anyone who can lend a hand?

ChoCumsky commented 5 months ago

I am having the same problem, did you ever figure out the solution?

Mirzyaaliii commented 2 months ago

I am also working with grayscale images and my result (reconstruction) is horrible.

Can anyone suggest some strategies to get better results for grayscale images?

hugoWR commented 1 week ago

Can you provide more details ? In my experience it's possible to obtain good reconstruction results even with Gray scale images