GPU memory requirements increases when using DCNv3_pytorch rather than DCNv3

OpenGVLab / InternImage

[CVPR 2023 Highlight] InternImage: Exploring Large-Scale Vision Foundation Models with Deformable Convolutions

https://arxiv.org/abs/2211.05778

MIT License

2.47k stars 231 forks source link

GPU memory requirements increases when using DCNv3_pytorch rather than DCNv3 #260

Open jonassoebro opened 9 months ago

jonassoebro commented 9 months ago

Thank you for this awesome implementation!

I've been using the InternImage as a backbone of my Detectron2 Faster-R-CNN implementation. My trained models need to run on CPU, so I tried making them compatible using the DCNv3_pytorch operator rather than the DCNv3. However the pytorch implementation seems to require a lot more GPU memory - is this expected or am I missing something?