Conv Transpose 2d for Pytorch initialized with bilinear filter / kernel weights

pytorch_bilinear_conv_transpose.py

"""
Copyright 2018 Matt Stevens. See license at end of file.
Defines the BilinearConvTranspose2d class.
"""
import torch
import torch.nn as nn


class BilinearConvTranspose2d(nn.ConvTranspose2d):
    """A conv transpose initialized to bilinear interpolation."""

    def __init__(self, channels, stride, groups=1):
        """Set up the layer.

        Parameters
        ----------
        channels: int
            The number of input and output channels

        stride: int or tuple
            The amount of upsampling to do

        groups: int
            Set to 1 for a standard convolution. Set equal to channels to
            make sure there is no cross-talk between channels.
        """
        if isinstance(stride, int):
            stride = (stride, stride)

        assert groups in (1, channels), "Must use no grouping, " + \
            "or one group per channel"

        kernel_size = (2 * stride[0] - 1, 2 * stride[1] - 1)
        padding = (stride[0] - 1, stride[1] - 1)
        super().__init__(
            channels, channels,
            kernel_size=kernel_size,
            stride=stride,
            padding=padding,
            groups=groups)

    def reset_parameters(self):
        """Reset the weight and bias."""
        nn.init.constant(self.bias, 0)
        nn.init.constant(self.weight, 0)
        bilinear_kernel = self.bilinear_kernel(self.stride)
        for i in range(self.in_channels):
            if self.groups == 1:
                j = i
            else:
                j = 0
            self.weight.data[i, j] = bilinear_kernel

    @staticmethod
    def bilinear_kernel(stride):
        """Generate a bilinear upsampling kernel."""
        num_dims = len(stride)

        shape = (1,) * num_dims
        bilinear_kernel = torch.ones(*shape)

        # The bilinear kernel is separable in its spatial dimensions
        # Build up the kernel channel by channel
        for channel in range(num_dims):
            channel_stride = stride[channel]
            kernel_size = 2 * channel_stride - 1
            # e.g. with stride = 4
            # delta = [-3, -2, -1, 0, 1, 2, 3]
            # channel_filter = [0.25, 0.5, 0.75, 1.0, 0.75, 0.5, 0.25]
            delta = torch.arange(1 - channel_stride, channel_stride)
            channel_filter = (1 - torch.abs(delta / channel_stride))
            # Apply the channel filter to the current channel
            shape = [1] * num_dims
            shape[channel] = kernel_size
            bilinear_kernel = bilinear_kernel * channel_filter.view(shape)
        return bilinear_kernel

"""
This file is subject to the Zero-Clause BSD license.

Permission to use, copy, modify, and/or distribute this software for any
purpose with or without fee is hereby granted.

THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
"""

@mjstevens777: I am very grateful for this work of yours in enabling transposed convolutions with bilinear interpolation in PyTorch. Would you be interested in facilitating the reuse of this code by including a license of your choice (MIT License, BSD License, Apache License 2.0 etc.)?

Done! I chose the most permissive license I could find. Thanks for your interest!