Refitting An Engine: What is the “tmp_weight” ?

[lg920810]

Description

When I use refitter.get_all_weights() in engine refitting, it appears "tmp_weight".

'ctr_hidden_0.weight', 'ctr_hidden_0.bias', 'tmp_weight_9', 'tmp_weight_8'

the layer in onnx file corresponding to the tensorrt engine is "Gemm->BatchNormalizaiton->Relu", where 'ctr_hidden_0.weight', 'ctr_hidden_0.bias' in Gemm. I think (Gemm,BN,Relu) has already fused a Conv, but I do not know what is tmp_weight_9 and tmp_weight_8?

Environment

TensorRT Version: 8.5.1.7

[zerollzeng]

@pranavm-nvidia Do you know? thanks!

[pranavm-nvidia]

Those are typically created by the ONNX parser in cases where the TensorRT requires weights but the corresponding ONNX op uses an attribute instead. In this case it's most likely the scale/shift for the BN layer.

[lg920810]

Those are typically created by the ONNX parser in cases where the TensorRT requires weights but the corresponding ONNX op uses an attribute instead. In this case it's most likely the scale/shift for the BN layer.

I think you are right, I will verify it. Thanks~

[ttyio]

Closing since no activity for more than 3 weeks, thank you!