如何使用 nn.conv2d 做与 nn.conv1d 一样的 function？

Question

I'm working on deploying python CNNs on FPGAs, and i'm facing an issue.

I have a 1d CNN which uses torch.conv1d layer function, and it's not supported by the software I'm using: vitis-ai 1.3 .

I'm figuring out if there is some way to use nn.conv2d instead to do the same job as nn.conv1d ?

for example:

Conv1d(1, 32, 3, bias = True) => Conv2d(... 

Is there a way to do that? (without loosing too much performance?)

Thank you.

Answer 1

Given 1d input, a 2d convolution of kernel-size (n,1) is equivalent to a 1d convolution of kernel-size (n,) :

nn.Conv1d(in_channels, out_channels, n)

# for 1d input, is equivalent to:
nn.Conv2d(in_channels, out_channels, (n,1))

---

Note you will have to unsqueeze(-1) the input to get it of the correct shape for a 2d convolution, but you can return the output to the original shape by applying squeeze(-1) :

F.conv1d(inputs, filters)

# is equivalent to:
F.conv2d(inputs.unsqueeze(-1), filters.unsqueeze(-1)).squeeze(-1)