Description
I try to build a siamRPN Network with TensorRT C++ API, with a AlexNet Backbone. What I want is to do the cross correlation between two ITensor* of shape [1, 126, 24, 24] and [256, 1, 4, 4]. Those ITensor are output of ILayers. I have the network working with pytorch in python. This cross Correlation layer is Implemented as followed in python:
torch.nn.fonctionnal.conv2D(tensor1, tensor 2, groups=nb_channel_of_tensor1).
I don’t know if I need to do a plugin layer or if I can use network->addConvolution layer to do it.
I tried this solution :
ILayer* siamRPN::crossCorrelation(INetworkDefinition network, ITensor input1, ITensor* input2) {
Weights emptywts{ DataType::kFLOAT, nullptr, 0 };
Weights emptybiais{ DataType::kFLOAT, nullptr, 0};
int16_t Nb_groups = 256;
IConvolutionLayer* crossConv = network->addConvolutionNd(*input1, 256, DimsHW{4, 4}, emptywts, emptybiais);
crossConv->setNbGroups(Nb_groups);
crossConv->setInput(1, *input2);
return crossConv;
}
But when building engine this error is returned
[06/06/2023-13:43:39] [I] [TRT] Graph optimization time: 0.0184208 seconds.
[06/06/2023-13:43:39] [I] [TRT] Local timing cache in use. Profiling results in this builder pass will not be stored.
[06/06/2023-13:43:51] [E] [TRT] 10: Could not find any implementation for node (Unnamed Layer* 46) [Convolution].
[06/06/2023-13:43:51] [E] [TRT] 10: [optimizer.cpp::computeCosts::3869] Error Code 10: Internal Error (Could not find any implementation for node (Unnamed Layer* 46) [Convolution].)
Environment
TensorRT Version: 8.6.0.1
GPU Type: RTX3060
Nvidia Driver Version: 530.30.02
CUDA Version: 12.1
CUDNN Version:
Operating System + Version: Linux Ubuntu 20.04
Thanks in advance for the help