Custom (TF) Model - Which config to set in nvinfer?

We’ve got a custom TF->ONNX->TRT model running, but the outputs are not exactly the same.

We’re not using any normalization or pixel scaling on the (grayscale) data.

We’re getting close results but still too different using net-scale-factor=0.0039215686274509803921568627451

What are we missing?

• Hardware Platform (Jetson / GPU) Jetson Xavier AGX
• DeepStream Version 6.3
• JetPack Version (valid for Jetson only) 5.1.2
• Issue Type( questions, new requirements, bugs) Question

The net-scale-factor parameter should be aligned with your training parameters. Please check the preprocess of the training.

Yes, we’re aware.

Nothing is happening in the preprocess other than scaling the frames to a smaller size.
Frames are in grayscale.
That’s really it.

We saw, internally it seems nvinfer transforms gray8 full range [0-255] to some studio range it seems [16-235] ?

What else is happening?

We tried looking into nvinfer source code, but we cannot find the exact preprocessing steps that are being taken, can you show us exactly where it’s happening?


There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

What is your picture’s format? Can you show us your pipeline?

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