search

NVIDIA / DALI

A GPU-accelerated library containing highly optimized building blocks and an execution engine for data processing to accelerate deep learning training and inference applications.
https://docs.nvidia.com/deeplearning/dali/user-guide/docs/index.html
Apache License 2.0
5.16k stars 622 forks source link

Will operations like sort, argsort or nms get implemented? #4725

Open SunXuan90 opened 1 year ago

SunXuan90 commented 1 year ago

DALI is pretty useful for postprocessing when using ensemble model in Triton Inference Server. Will the commonly used operations get implemented in the future?

awolant commented 1 year ago

Hello @SunXuan90 glad to hear you find DALI and Triton useful. With regard to the future plans the we have a Roadmap for 2023 here: https://github.com/NVIDIA/DALI/issues/4578

Could you tell more about the operators you would like to see in DALI and point us to some use cases you have for them? We are always keen on improving DALI coverage of ops.

szalpal commented 1 year ago

@SunXuan90 ,

have you considered using cupy for the operations you need? https://docs.cupy.dev/en/stable/reference/generated/cupy.sort.html# https://docs.cupy.dev/en/stable/reference/generated/cupy.argsort.html

SunXuan90 commented 1 year ago

@awolant, NMS would be great. I know there's a plugin for nms in TensorRT, but still, having NMS in DALI makes it more flexible. If the full NMS can't be inplemented, sort and argsort are needed.

SunXuan90 commented 1 year ago

@szalpal , These are helpful indeed, but I'm using DALI backend in Triton, which has it's own conda environment, and cupy is not pre-installed. I also read your comment about keeping DALI Backend dependencies to minimal. So is there a easy way that I can install dependencies during deployment now?