I have a Tensor that represents a set of 1D signals, that are concatenated along the column axis. So say I have 300 1D signals that are of size 64. So [64x300]
I want to apply a smooth convolution / moving average kernel on it [0.2 0.2 0.2 0.2 0.2] on the GPU, but I am not sure exactly what is the API to do it. Can I be provided an example?
You can use regular torch.nn.Conv1d to do this.
In your case you have 1 channel (1D) with 300 timesteps (please refer to documentation those values will be appropriately C_in and L_in).
So, for your input it would be (you need 1 there, it cannot be squeezed!):
import torch
inputs = torch.randn(64, 1, 300)
You need torch.nn.Conv1d with kernel_size equal to 5 (as indicated by your elements: [0.2 0.2 0.2 0.2 0.2]) and no bias. I assume your output has to be of the same size (300) so 2 elements have to be padded at the beginning and end. All of this gives us this module:
module = torch.nn.Conv1d(
in_channels=1, out_channels=1, kernel_size=5, padding=2, bias=False
)
Weights of this module (0.2 values) can be specified like this:
module.weight.data = torch.full_like(module.weight.data, 0.2)
torch.full_like will work for kernel of any size in case you want other size than 5.
Finally run it to average steps and you're done:
out = module(inputs)
If you want to use GPU just cast your module and inputs like this:
inputs = inputs.cuda()
module = module.cuda()
See CUDA documentation for more information.
