search

aim-uofa / GenPercept

GenPercept: Diffusion Models Trained with Large Data Are Transferable Visual Models
https://huggingface.co/spaces/guangkaixu/GenPercept
Creative Commons Zero v1.0 Universal
91 stars 2 forks source link

Quantitative result mismatch #4

Closed jk4011 closed 2 months ago

jk4011 commented 2 months ago

Hello, and thank you for your excellent work. I noticed that the quantitative results from tables 1, 9, 10, and 12 differ. For example, the KITTI-AbsRel scores for GenPercept (or Baseline) are 0.099, 0.140, 0.145, and 0.140, respectively. Could you help me understand what I might be missing?

guangkaixu commented 2 months ago

For table 1, the model is trained on Hypersim + Virtual KITTI. For tables 9, 10, and 12, these models are trained on Hypersim only.

For tables 9 and 12, they are trained with a batchsize of 16, but it changed to 8 for tale 10. It seems that we forgot to mention it in our paper, and we will update it in the future.

The reason for changing the batchsize to 8 is that GenPercent (PixArt-α) can not be trained for batchsize 16 because of "CUDA out of memory" on A800.