Question: How to use even sized kernels with Monai Convolution block?

Hi All, I am working on rewriting MC_GAN with the MONAI framework and have a question with the Monai Convolution block. Why are even number kernels not supported in the Monai Convolution block?

I am running into this problem with the block when my kernel size is an even number. The monai.networks.layers.convutils.same_padding() func called by Convolution().__init__ is throwing a NotImplementedError:

# from monai.networks.layers.convutils.same_padding()

    if np.any((kernel_size_np - 1) * dilation % 2 == 1):
        raise NotImplementedError(
            f"Same padding not available for kernel_size={kernel_size_np} and dilation={dilation_np}."
        )

    padding_np = (kernel_size_np - 1) / 2 * dilation_np

Is there an implementation planned for this in the future? Why is there no manual padding option in the Convolution block constructor?

I did some value testing and padding_np is an integer when k_size is odd, and ends in 0.5 when k_size is even.

For reference these are the nn.Sequential() blocks:

# Original PyTorch source I am replacing:
            torch.nn.Conv2d(in_chan, out_chan, kernel_size=4, stride=2, padding=1, bias=False)
            torch.nn.BatchNorm2d(out_chan)
            torch.nn.LeakyReLU(0.2, inplace=True)
# What I want to use: Monai Convolution Block
            Convolution(2, in_chan, out_chan, kernel_size=4, strides=1, bias=False, act=self.Act, norm=self.Norm)
# What I am using: Monai LayerFactory
            Conv["conv", 2](in_chan, out_chan, 4, 2, 1, bias=False),
            Norm["batch", 2](out_chan),
            Act["leakyrelu"](0.2, inplace=True),

Thank you for any help.

[ericspod]

The problem though is the padding for Conv2d isn’t sufficient to produce an output with the same shape given an even-sized kernel and a stride of 1:

t=torch.rand(1,1,64,64)

c=nn.Conv2d(1,1,4,1,padding=(0,0))
print(c(t).shape)  # 1x1x61x61

c=nn.Conv2d(1,1,4,1,padding=(1,1))
print(c(t).shape)  # 1x1x63x63

c=nn.Conv2d(1,1,4,1,padding=(2,2),)
print(c(t).shape)  # 1x1x65x65

Choosing to pad the input beforehand as well as pad in Conv2d can do this but when different stride or dilation values are used it becomes quite difficult to calculate what that pad should be. For example with a stride of 2 the expected output would be halved in every spatial dimension regardless of kernel size or dilation value, figuring out how to pad the input to do this isn’t clear to me.

[gagandaroach]

Thank you, this will be helpful.

The initial conv in MC GAN G/D does not use normalization but does use activation. Subsequent blocks use both.

I think this issue can be closed now.