search

geopavlakos / hamer

HaMeR: Reconstructing Hands in 3D with Transformers
https://geopavlakos.github.io/hamer/
MIT License
326 stars 28 forks source link

Training loss does not converge #24

Closed linjiangya closed 6 months ago

linjiangya commented 6 months ago

Originally posted by @linjiangya in https://github.com/geopavlakos/hamer/issues/17#issuecomment-1901318513

Here are examples of using official checkpoints and checkpoints of mine:

image (using epoch=0-step=1000000.ckpt) image (using the officially released model hamer.ckpt)

Here's the tensorboard log: image

geopavlakos commented 6 months ago

Thanks for pointing this out. I'm still investigating this, but I believe it happened due to the incorrect initialization of the backbone. I have fixed this, and the training losses are decreasing as expected. Could you pull the latest updates and give it another shot?

xungeer29 commented 5 months ago

Thanks for pointing this out. I'm still investigating this, but I believe it happened due to the incorrect initialization of the backbone. I have fixed this, and the training losses are decreasing as expected. Could you pull the latest updates and give it another shot?

I want to train hamer with vitpose_base as the backbone, but how to get the pretrained vitpose_backbone_base.pth?