dc1d 0.0.2

A PyTorch implementation of 1D deformable convolution

dc1d (DeformConv1d)

An 1D implementation of a deformable convolutional layer implemented in pure Python in PyTorch. The code style is designed to imitate similar classes in PyTorch such as torch.nn.Conv1D and torchvision.ops.DeformConv2D.

Requirements

You must install PyTorch. Follow the details on the website to install properly: https://pytorch.org/get-started/locally/

This package was most thoroughly tested on PyTorch 1.12.1 running CUDA 11.6 on Windows with Python version 3.8.

Installation

pip install dc1d

Usage

Example of how to use the deformable convolutional layer DeformConv1d() with timing information.

DeformConv1d is the deformable convolution layer designed to imitate torch.nn.Conv1d. Note: DeformConv1d does not compute the offset values used in its forward(...) call. These most be computed outside the layer.

import time

# Import layer
from dc1d.nn import DeformConv1d

# Hyperparameters
batch_size = 16
in_channels = 512
out_channels = 512
kernel_size = 16
stride = 1
padding = "same"
dilation = 3
groups = 1
bias = True
length = 128

# Construct layer
model = DeformConv1d(
    in_channels = in_channels,
    out_channels = out_channels,
    kernel_size = kernel_size,
    stride = stride,
    padding = "same",
    dilation = dilation,
    groups = groups,
    bias = True,
    device="cuda"
)

# Generate input sequence
x = torch.rand(batch_size, in_channels, length,requires_grad=True).cuda()
print(x.shape)

# Generate offsets by first computing the desired output length
output_length = x.shape[-1]-dilation*(kernel_size-1)
offsets = nn.Parameter(torch.ones(batch_size, 1, output_length, kernel_size, requires_grad=True, device="cuda"))

# Process the input sequence and time it
start = time.time()
y = model(x, offsets)
end = time.time()

# Print output shape and time taken
print(y.shape)
print("Deformable runtime =",end-start)

---

For more detailed examples, the nn and ops modules have example usage scripts appended to the bottom of the file inside their if __name__ == "__main__": clauses. For example one could run

python dc1d/nn.py

to compare the runtime of our DeformConv1d layer against torch.nn.Conv1d.

A class called PackedConv1d also exists in dc1d.nn which computes the offsets using a depthwise separable convolutionoperation as detailed in our paper below

Papers

Please cite the following if you use this package4

@article{ravenscroft2022dtcn,
  title={Deformable Temporal Convolutional Networks for Monaural Noisy Reverberant Speech Separation},
  author={Ravenscroft, William and Goetze, Stefan and Hain, Thomas},
  year={2022}
}