search

Epiphqny / VisTR

[CVPR2021 Oral] End-to-End Video Instance Segmentation with Transformers
https://arxiv.org/abs/2011.14503
Apache License 2.0
740 stars 96 forks source link

CUDA out of memory on 8xv100 32G GPUs for r101 training #59

Closed rambleramble closed 3 years ago

rambleramble commented 3 years ago

CUDA out of memory with the current source code on 8xv100 32G

However, it seems the following code, need to use r101 resolution, instead of r50 to fit into 32G of V100. https://github.com/Epiphqny/VisTR/blob/554e711602dff135b15afb36da0efec46e13f648/datasets/ytvos.py#L167

My question is that the claimed result of r101, is it achieved under the r101 resolution, aka T.RandomResize([280], max_size=504),#for r101

or the default source code one: T.RandomResize([300], max_size=540),#for r50

thanks in advance!!

Epiphqny commented 3 years ago

Hi @rambleramble, the result is achieved by the resolution of r101