cleinc / bts

From Big to Small: Multi-Scale Local Planar Guidance for Monocular Depth Estimation

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Stability of training

tim885 opened this issue · comments

Hi,
Thanks a lot for sharing your code! However, during training with offered code on NYUDv2 24K raw data, I found a strange instability about validation as follows:
image
Whether you have also encountered this during training and how to explain it? Thank you by advance.

@tim885 We have experienced the same. We consider it as an instability of our network when trained with low batch sizes.

Could you tell me how many hours it takes to train with offered code on NYUDv2 24K or KITTI raw data such as with 4 GPU? Thanks! @cogaplex-bts @cogaplex-bts