How use TensorRT API to add conv1d layers?

[oreo-lp]

Description

I want to use TensorRT API to rebuild a model, which contains some conv1d layers [use torch.Conv1d()]. Inputs shape is [1,1,82000], and the output of conv1d is [1,512,16399], kernel_size = 10, stride = 5, padding = 0. Some code is below: auto conv = newtork->addConvolutionNd(input, 512, DimsHW{10, 1}, weight, bias); conv->setStrideNd(DimsHW{5,1 }); But it failed. The error is below: [graphShapeAnalyzer.cpp::addVolumeCheck::610] Error Code 9: Internal Error (output: dimension 1 never exceeds -1) Error Code 4: Internal Error (Could not compute dimensions for output, because the network is not valid.) How use addConvolutionNd() to add conv layer to network?

[zerollzeng]

@nvpohanh ^ ^

[zhudongwork]

Description

I want to use TensorRT API to rebuild a model, which contains some conv1d layers [use torch.Conv1d()]. Inputs shape is [1,1,82000], and the output of conv1d is [1,512,16399], kernel_size = 10, stride = 5, padding = 0. Some code is below: auto conv = newtork->addConvolutionNd(input, 512, DimsHW{10, 1}, weight, bias); conv->setStrideNd(DimsHW{5,1 }); But it failed. The error is below: [graphShapeAnalyzer.cpp::addVolumeCheck::610] Error Code 9: Internal Error (output: dimension 1 never exceeds -1) Error Code 4: Internal Error (Could not compute dimensions for output, because the network is not valid.) How use addConvolutionNd() to add conv layer to network?

I also had the same problem

[zhudongwork]

Inputs shape

try to reshape the input to [1,1,82000, 1], the inputshape of the addConvolutionNd should be a dims4 itensor

[nvpohanh]

TRT doesn't support 1D-Conv yet. You need to first unsqueeze the tensor, add it as a 2D conv, and then squeeze the tensor

[ttyio]

Closing since no activity for more than 3 weeks, please reopen if you still have question, thanks!