I want to serve DINO model on Triton-server (after converting to tensorrt), and run object detection on custom python triton client (not in deepstream)
I want to figure out the pre and post processing required. I am looking at the following two repositories:
- tao pytorch backend : tao_pytorch_backend/nvidia_tao_pytorch/cv at main · NVIDIA/tao_pytorch_backend · GitHub
- nvidia tao deploy: tao_deploy/nvidia_tao_deploy at main · NVIDIA/tao_deploy · GitHub
I found the pre-processing methods used in both are similar. However the post processing method is not clear.
In tao pytorch backend, the post processing for DINO is done within the models’ forward() method. Therefore I assumed we dont need to apply any post processing further to the output from model.
However as given under this post, according to TAO deploy repo, specific post processing steps have been applied to the output of the tensrrot model.
Can you please clarify this and provide reference to post processing steps required for the TAO DINO model output.
1 post - 1 participant