Follow up on unpruned resnet10 model for Deepstream on Orin Nano/NX / Alternatives

Continuing the discussion from How to build resnet10 equivalent with TLT for deepstream:

  1. Is there an update as to whether Nvidia will provide a unpruned resnet-10 model for to retrain/prune in TAO.
  2. Can you provide real-time inference performance of TrafficCamNet (TrafficCamNet | NVIDIA NGC) for Orin Nano and NX.


Trafficcamnet unpruned model has only resnet18 version.
For inference performance, please refer to Overview - NVIDIA Docs.

TrafficCamNet (Based on Detecnet_v2-ResNet18)
Input: 960x544x3
Mode: INT8

GPU FPS is as below.
Xaiver NX: bs8, 251fps
AGX Xavier:bs16, 458fps
Orin NX: bs16, 457fps
Orin: bs32, 1268fps
Orin Nano: N/A
T4: bs64, 1725fps
A100: bs256, 9717fps
A30: bs64, 5082fps
A10: bs64, 4754fps
A2: bs32, 916fps

1 Like

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.