Conv1d just like numpy convolve

1

Hi

I am trying to mimic numpy.convovle jusing torch.nn.functional.conv1d. The results are not the same given my dimensions. I appreciate if someone can correct it. The script is below

import torch.nn.functional as F
import numpy as np

inputs = torch.randn(2,240,60)
filters = torch.randn(240,240,60)
filters_flip=filters.flip(2)
F.conv1d(inputs[:,:,:], filters[:,:,:]).shape
conv_tens=F.conv1d(inputs, filters_flip, padding=30)[1,1,:]

inputs_np=inputs.numpy()
filters_np=filters.numpy()
conv_np=np.convolve(filters_np[1,1,:],inputs_np[1,1,:], mode=‘same’)

Thanks

2

Since I am interested in 1d convolution. Using torchaudio.functional helped. Below is small sample

import torchaudio.functional as F2

inputs2 = torch.randn(60)
filters2 = torch.randn(60)
filters2_flip=filters2.flip(0)
conv_tens=F2.convolve(inputs2, filters2,‘same’)

conv_np=np.convolve(filters2,inputs2, mode=‘same’)