search

lhoyer / HRDA

[ECCV22] Official Implementation of HRDA: Context-Aware High-Resolution Domain-Adaptive Semantic Segmentation
Other
231 stars 32 forks source link

Using 2 GPUs to train the model. #30

Closed Caillen-W closed 1 year ago

Caillen-W commented 1 year ago

I want to know if I use 2 GPUs to train the model, where I should modify in the codes. Thank you very much.

lhoyer commented 1 year ago

Thank you for your interest in HRDA. Multi-GPU training is currently not supported by this repository.

Caillen-W commented 1 year ago

Thank you for your reply. Does this repository support the mixed precision training?

lhoyer commented 1 year ago

This repository is based on mmsegmentation v0.16.0, which supports mixed precision training according to its README.md (https://github.com/open-mmlab/mmsegmentation/tree/v0.16.0). However, the HRDA was not developed considering this aspect. So, you have to check if the relevant changes compared to the basic mmsegmentation framework are compatible with mixed precision training. Also, from experience with a previous project, mixed precision training can cause training instabilities and the original results might be not reproducable.