This is a fork of original fft-conv-pytorch.
I made some modifications to support dilated and strided convolution, so it can be a drop-in-replacement of original PyTorch Conv*d
modules and conv*d
functions, with the same function parameters and behavior.
pip install git+https://github.com/yoyololicon/fft-conv-pytorch
import torch
from torch_fftconv import fft_conv1d, FFTConv1d
# Create dummy data.
# Data shape: (batch, channels, length)
# Kernel shape: (out_channels, in_channels, kernel_size)
# Bias shape: (out channels, )
# For ordinary 1D convolution, simply set batch=1.
signal = torch.randn(3, 3, 1024 * 1024)
kernel = torch.randn(2, 3, 128)
bias = torch.randn(2)
# Functional execution. (Easiest for generic use cases.)
out = fft_conv1d(signal, kernel, bias=bias)
# Object-oriented execution. (Requires some extra work, since the
# defined classes were designed for use in neural networks.)
fft_conv = FFTConv1d(3, 2, 128, bias=True)
fft_conv.weight = torch.nn.Parameter(kernel)
fft_conv.bias = torch.nn.Parameter(bias)
out = fft_conv1d(signal)
The best situation to use FFTConv
is when using large size kernel. The following image shows that when the size of input is fixed, the fft method remains an almost constant cost among all size of kernel, regardless.
For details and benchmarks on other parameters, check this notebook.
- Jittability.
- Dilated Convolution.
- Transposed Convolution.