UpSample optional kernel_size for deconv mode

[myron]

Description

Adds (optional) kernel_size parameter to UpSample, used for deconv (convolution transpose up-sampling).

This allows to upsample, e.g to upscale to 2x with a kernel_size 3. (currently the default is to upscale to 2x with a kernel size 2)

if this parameter is not set, the behavior is the same as before

Types of changes

• Non-breaking change (fix or new feature that would not break existing functionality).
• Breaking change (fix or new feature that would cause existing functionality to change).
• New tests added to cover the changes.
• Integration tests passed locally by running ./runtests.sh -f -u --net --coverage.
• Quick tests passed locally by running ./runtests.sh --quick --unittests --disttests.
• In-line docstrings updated.
• Documentation updated, tested make html command in the docs/ folder.

[KumoLiu]

Hi @myron, you can set up padding and output_padding separately like this, which might solve the test error of mismatched test output sizes.

output_padding = [0, ] * spatial_dims
for i, _scale_factor_ in enumerate(scale_factor_):
    if kernel_size_[i] % 2 == 0:
        output_padding[i] = _scale_factor_ - 2
    else:
        output_padding[i] = _scale_factor_ - 1

padding = tuple((i-1)//2 for i in kernel_size_)

self.add_module(
    "deconv",
    Conv[Conv.CONVTRANS, spatial_dims](
        in_channels=in_channels,
        out_channels=out_channels or in_channels,
        kernel_size=kernel_size_,
        stride=scale_factor_,
        padding=padding,
        output_padding=tuple(output_padding),
        bias=bias,
    ),
)

Thanks!

[myron]

ok, I believe I fixed that, and added another unit test. thanks you

[wyli]

looks good to me, just to note that when the kernel size is not divisible by the scale_factor, this will introduce some checkerboard artefacts.

[wyli]

/build

[wyli]

/build

[myron]

looks good to me, just to note that when the kernel size is not divisible by the scale_factor, this will introduce some checkerboard artefacts.

Thanks. is it a known fact? I use it with scale 2, kernel 3, I don't observe any artifacts.

[wyli]

the idea is presented here https://distill.pub/2016/deconv-checkerboard/, I guess it'll be less noticeable for segmentation tasks..