How can you generate a Convolution Operator with UINT8/INT8 output tensor?

[markcrowley1]

Hi

I need to generate an ONNX model containing a single convolution operation with INT8 precision for the input, weight and output tensors.

I can't do this by simply using the ONNX Conv or ConvInteger operations, as Conv only supports float precision, and the output tensor of a ConvInteger operation is INT32.

Is there a simple way of achieving what I want without having to quantize the model? If quantizing the model is the only option, please let me know the simplest way of doing this.

Thanks