Torch::nn::functional::conv2d memory leak?

C++
1

Hello, I encountered a memory leak when writing a convolutional layer using libtorch1.7GPU version. When I execute the code of torch::nn::functional::conv2d in a loop, the memory keeps growing. Excuse me, why is this?

torch::Tensor gaussian_filter(torch::Tensor x, int channels)
{
	/*
	*Parameters:
	* tensor x [C,H,W]  int channels 
	*Return tensor x [C,H,W]
	*/

	namespace F = torch::nn::functional;
	torch::Tensor kernel = 
		torch::tensor({ {0.00078633, 0.00655965, 0.01330373, 0.00655965, 0.00078633},
						{0.00655965, 0.05472157, 0.11098164, 0.05472157, 0.00655965},
						{0.01330373, 0.11098164, 0.22508352, 0.11098164, 0.01330373},
						{0.00655965, 0.05472157, 0.11098164, 0.05472157, 0.00655965},
						{0.00078633, 0.00655965, 0.01330373, 0.00655965, 0.00078633} });


	kernel = kernel.unsqueeze(0).unsqueeze(0);
	kernel = torch::repeat_interleave(kernel, channels, 0);
	torch::Tensor out_put = F::conv2d(x.unsqueeze(0), kernel.to(x.device()), F::Conv2dFuncOptions().stride(1).padding(2).groups(channels));
	return out_put.squeeze(0);

}
2

I assume you are running gaussian_filter in a loop?
If so, are you storing the output tensor in each iteration?

1 Like
3

Execute gaussian_filter in a loop. I don’t want to store torch::Tensor out_put, I just want to get the return value and assign it to other tensors. Then destroy the memory of torch::Tensor out_put. But I found that the memory has been growing.Please tell me what should I do？

4

I’m still unsure if you are (accidentally) storing the tensors somehow or if they are attached to a computation graph, which could be stored in the assignment, so could you post an executable code snippet, which would reproduce this issue?

1 Like
5

Thank you very much, I solved this problem by canceling the gradient of the variables in the convolution operation. It may have been because of the calculation gradient that took up memory.