about QAT model

[lxzatwowone1]

I have a question, Why need qat model to convert dla model, can I use original onnx model?
Thanks! I am DL beginner

[liuanqi-libra7]

Hi @lxzatwowone1

• DLA's INT8 compute capability is much higher than float, we recommend that the model be inferred with INT8 precision on DLA.
• The accuracy of implicit quantization using TensorRT will be lower than PyTorch Quantization, so we recommend using PyTorch Quantization to quantize the model.