search

facebookresearch / hiera

Hiera: A fast, powerful, and simple hierarchical vision transformer.
Apache License 2.0
870 stars 39 forks source link

Training scripts #28

Open VGrondin opened 7 months ago

VGrondin commented 7 months ago

Hello great paper,

Any idea on when the training files will be available?

dbolya commented 7 months ago

Training code kinda fell by the wayside (we have to rewrite it from scratch because the original project used internal training code). I do however plan to get it out by the end of this month alongside the model / code release for the Window Attention is Bugged paper.

VGrondin commented 7 months ago

Thanks! Looking forward to it :)

tkhe commented 5 months ago

hi, is there any new developments？

dbolya commented 5 months ago

hi, is there any new developments？

The in progress training code is available here: https://github.com/dbolya/hiera/tree/v0.2.0

See examples/train.py to see how to use it. However, I'm still waiting on some full training runs to fully test it.

tapohongchen commented 5 months ago

Hello, thanks for your work. I saw the train in the picture. Any idea on When the train in the video will be available, Especially in the ava.

qianyizhang commented 2 months ago

Training code kinda fell by the wayside (we have to rewrite it from scratch because the original project used internal training code). I do however plan to get it out by the end of this month alongside the model / code release for the Window Attention is Bugged paper.

nice work, i would also like too read about HieraDet release :P

guillembraso commented 2 weeks ago

Hi! Is there any update on the release of the training code? Or could someone please share some script with the preliminary release? Thanks in advance! 🙏

andyqmongo commented 1 week ago

hi, is there any new developments？

The in progress training code is available here: https://github.com/dbolya/hiera/tree/v0.2.0

See examples/train.py to see how to use it. However, I'm still waiting on some full training runs to fully test it.

Thanks for sharing the training script! However, I wonder why increasing masking ratio doesn't reduce training time. I thought that the work introduced a way to speed up MAE pretrained without compromising quality. Or did I miss something? Thanks! 🙏

guokeqianhg commented 6 days ago

Training code kinda fell by the wayside (we have to rewrite it from scratch because the original project used internal training code). I do however plan to get it out by the end of this month alongside the model / code release for the Window Attention is Bugged paper.

Hi, Thanks for your works. Is the code for Window Attention is Bugged paper released now?