relbers/ada-conv-pytorch

ada-conv-pytorch's Issues

ValueError: test_size=8 should be either positive and smaller than the number of samples 0 or a float in the (0, 1) range

Hello, I am glad to see your article, I was very inspired, but when I run the code you provided, it appeared
ValueError: test_size=8 should be either positive and smaller than the number of samples 0 or a float in the (0, 1) range

Can you help me solve this problem?

Improved speed with batch adaconv?

Thanks for publishing this code! I've been tinkering around with AdaConv for months, and it's super helpful seeing somebody else's interpretation!

Something I've implemented in my code you might consider is applying the kernel convolutions in one go, by stacking all the channels together. Example:

    def forward(self, style_encoding: torch.Tensor, predicted: torch.Tensor, thumb_stats=None):
        N, c, h, w = predicted.size()
        depthwise = self.depthwise_kernel_conv(style_encoding)
        depthwise = depthwise.view(N*self.c_out, self.c_in // self.n_groups, 5, 5)
        s_d = self.pointwise_avg_pool(style_encoding)
        pointwise_kn = self.pw_cn_kn(s_d).view(N*self.c_out, self.c_out // self.n_groups, 1, 1)
        pointwise_bias = self.pw_cn_bias(s_d).view(N*self.c_out)

        if self.norm:
            predicted = F.instance_norm(predicted)

        predicted = predicted.view(1,N*C,h,w)
        content_out = nn.functional.conv2d(
                nn.functional.conv2d(self.pad(predicted),
                                     weight=depthwise,
                                     stride=1,
                                     groups=self.batch_groups
                                     ),
                stride=1,
                weight=pointwise_kn,
                bias=pointwise_bias,
                groups=self.batch_groups)
        content_out = content_out.permute([1, 0, 2, 3]).view(N,C,h,w)
        return content_out

I believe this achieves the same results without iteration.

Bias before pointwise conv

ada-conv-pytorch/lib/nn/adaconv/adaconv.py

Line 44 in bb3bab5

x = F.conv2d(x, w_spatial, groups=self.n_groups, bias=bias)

In the paper the bias is added after the pointwise cov. Is there a reason why we do it before here?

train issue

TypeError: transfer_batch_to_device() takes 3 positional arguments but 4 were given

AdaConv with StyleGAN

Thanks for the authors' of this repo hard work.

I have a question about implementation of styleGAN2 with adaConv. Hope, somebody can help me understand how it is supposed to be implemented. StyleGAN2 takes style as a 1d vector, but adaconv requires the style to be an image (2d vector). Does the reshape of 1d style to an image is done at the beginning with replacement of Linear layers with Conv2d or layers aren't changed except for demodulation part, before which reshape is done every time?

On the implementation of pointwise convolution

Thanks for the authors' hard work of this code repository.

I have noticed that in current implementation, the pointwise convolution is also generated and viewed as group convolution (

ada-conv-pytorch/lib/adaconv/kernel_predictor.py

Line 23 in 5d0578e

nn.Conv2d(style_channels,

). For example, with input channel C_in and output channel C_out, the generated pointwise convolution is with size C_out x C_in/N_g x 1 x 1. However, according my understandig I wonder if the pointwise convolution should have the same size as a normal convolution kernel with size C_out x C_in x 1 x1 ?

Looking forward to your respose. Thanks a lot.