Visual ChangeNet-Segmentation Discrepancy in Inference Results Between Custom TensorRT Code and NVIDIA TAO Toolkit

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Dear NVIDIA Support Team,

I successfully converted my model （Visual ChangeNet segmentation) to TensorRT using the NVIDIA TAO Toolkit on an RTX 3060. However, I’ve noticed that the inference results from my custom TensorRT code differ from those generated using the TAO Toolkit’s pipeline.

I’ve ensured that preprocessing (e.g., normalization, resizing) matches the TAO configuration, but the outputs still don’t align. Could you advise on potential causes or debugging steps?
By the way , my TensorRT version is 8.4.1.5.

Thank you for your help!

The following are training experiment.yaml and TensorRT inference code.
experiment.txt (5.2 KB)

code.txt (2.6 KB)

3

Please the TensorRT version when you run inference with TAO.
Then use the same TensorRT version to test with your code.
More, the TAO code is open sourced. See tao_deploy/nvidia_tao_deploy/cv/visual_changenet/scripts/inference.py at main · NVIDIA/tao_deploy · GitHub. You can leverage it.