Onnx to tensorrt conversion for input layer: cast uint8 to fp32

[max8]

My pytorch and onnx model has an uint8 to fp32 cast layer which divides by 255. This cast layer is applied to the input tensor. When i convert the onnx model to tensorrt INT8 i get the following warning:

“Missing scale and zero-point for tensor input, expect fall back to non-int8 implementation for any layer consuming or producing given tensor”

For INT8 should i remove the cast layer before exporting the onnx model or does tensorrt deal with it itself? What is the recommended approach for best INT8 performance?

[AakankshaS]

Hi @max8

@max8

“Missing scale and zero-point for tensor input, expect fall back to non-int8 implementation for any layer consuming or producing given tensor”

This is a warning from TRT due to the use of setDynamicRange API (implicit quantization instead of explicit quantization), so TRT requires you to explicitly set INT8 dynamic range of every layer to run that model in INT8, but TRT compiler on the other hand also fuses BN to conv, so this is not necessary, and this warning can then be ignored

Thanks