search

facebookresearch / ContrastiveSceneContexts

Code for CVPR 2021 oral paper "Exploring Data-Efficient 3D Scene Understanding with Contrastive Scene Contexts"
MIT License
223 stars 28 forks source link

Which epoch of pretrained model is used for downstream tasks? #40

Closed mutianxu closed 2 years ago

mutianxu commented 2 years ago

Dear authors,

Thanks for your great work!

Which epoch of models during ScanNet pertaining, would be later used for downstream tasks to get better performance?

Thanks, Mino

Sekunde commented 2 years ago

Hi, thanks for your interest! We do not ablate the pre-training epochs, and we use 100k iterations normally. There are some variances, however, we observe that training more can bring better performance in general (e.g, compare 100k iterations with 20k)