search

naver / dust3r

DUSt3R: Geometric 3D Vision Made Easy
https://dust3r.europe.naverlabs.com/
Other
4.97k stars 543 forks source link

Training and validation metrics details #140

Open recordmp3 opened 2 months ago

recordmp3 commented 2 months ago

Thank you for your wonderful work!

I'm trying to train dust3r from scratch but do not get good enough quality. Is it possible to provide with more details about your training and validation curve of three commands? Thank you in advance!

yocabon commented 1 month ago

we use our internal codebase to train our models, and use older versions of the datasets so it's very possible that there could be an issue in the release.

for reference, DUSt3R_ViTLarge_BaseDecoder_224_linear training went like this.

image image image

recordmp3 commented 1 month ago

may I know what is x?

yocabon commented 1 month ago

the x axis is the number of epochs * 1000 https://github.com/naver/dust3r/blob/main/dust3r/training.py#L327

recordmp3 commented 1 month ago

Thanks! And may I know your training time? (the Wall-clock time on x=100k)

yocabon commented 1 month ago

on 8xA100 machines, 224 linear Job Wall-clock time: 3-14:58:44 (3days 15h) 512 linear Job Wall-clock time: 1-17:47:59 512 dpt Job Wall-clock time: 2-02:26:22