how to use beit to train if the input size is 768x768

open-mmlab / mmsegmentation

OpenMMLab Semantic Segmentation Toolbox and Benchmark.

https://mmsegmentation.readthedocs.io/en/main/

Apache License 2.0

8.29k stars 2.62k forks source link

how to use beit to train if the input size is 768x768 #1973

Closed haoxurt closed 2 years ago

haoxurt commented 2 years ago

I want to use beit train ade20k with input size 768 x 768，can I only change the input size according to official configs? Do I need to change anything else？

fingertap commented 2 years ago

Hi @haoxurt ! I have a few questions on this training configs:

Will training with 768x768 lead to large performance gain?
Which paper uses 768x768 as input?
What is the memory consumption of 768x768 inputs on ADE20k?

Thanks!

haoxurt commented 2 years ago

There is no paper using 768x768 as input. I only want to try. I have A100 GPU with 40G memory and it can train with 768x768 .

fingertap commented 2 years ago

There is no paper using 768x768 as input. I only want to try. I have A100 GPU with 40G memory and it can train with 768x768 .

Can you please record the memory usage for a single GPU when you start to train? I wonder whether my 3090 (24G) can handle 768x768 inputs. Thanks!

haoxurt commented 2 years ago

3090(24G) can’t train. It need 39.6G .

fingertap commented 2 years ago

Thanks!

xiexinch commented 2 years ago

Hi @haoxurt, Might modify the crop size like https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/datasets/ade20k_640x640.py.