NVIDIA
GridSample 5D input support
Dear team, I notice currently tensorrt only supports 4D input GirdSample(see #3207), is there any chance to support 5D input in the future?
My Team also need this feature to be implemented urgently . Would it be considered ?
https://github.com/pytorch/pytorch/blob/v1.11.0/aten/src/ATen/native/cuda/GridSampler.cu
then define a trt plugin
@lix19937 Thanks, I will try your hints. But I hope this feature will be officially supported.
No efforts planned for this right now.
@1059692261 @tianlinzx , can you please let us know what kind of networks you're working with that are blocked by this? Will help us prioritize.
No efforts planned for this right now.
@1059692261 @tianlinzx , can you please let us know what kind of networks you're working with that are blocked by this? Will help us prioritize.
No efforts planned for this right now.
@1059692261 @tianlinzx , can you please let us know what kind of networks you're working with that are blocked by this? Will help us prioritize.
@brb-nv Dinet,a open source github project, Used to support talking face .
https://github.com/SeanWangJS/grid-sample3d-trt-plugin
https://github.com/pytorch/pytorch/blob/v1.11.0/aten/src/ATen/native/cuda/GridSampler.cu then define a trt plugin
Also commenting for more visibility. I'm creating a custom NeRF-like model. I sample multiple 3D feature grids with shape [N,C,D,H,W] given query points of shape [N,B,3]. A for loop iterating over N is a temporary solution but slows down training in PyTorch significantly (especially as N grows).
