NVlabs / MUNIT

Multimodal Unsupervised Image-to-Image Translation
Other
2.65k stars 483 forks source link

Why 4x4 Conv in this network, and where this idea come from #44

Open imlixinyang opened 6 years ago

imlixinyang commented 6 years ago

In discriminator, style encoder and content encoder, i find 4x4 conv filters. Where this idea came from or i missed something.

ShihuaHuang95 commented 5 years ago

4x4, stride=2, padding=1, in a word, just for downsample

doantientai commented 5 years ago

From my point of view, using even-sized conv kernels is just to show that Deep networks can work with conv kernel of sizes odd or even, or none-square shapes.

chychen commented 5 years ago

In my opinion, 4x4 kernel and 2x2 stride conv might be able to alleviate the checkerboard issues.

caffeinism commented 5 years ago

I think it is just the originality of the DCGAN paper (Deconvolution and convolution with kernel size 4).

And there seems to be no reason in DCGAN's convolution layer.

See the following author's article. https://discuss.pytorch.org/t/in-dcgan-why-the-kernel-size-of-4-is-used/20616/2

In deconvolution, however, it is convenient to upscale the size of a feature map exactly twice.

I think that the intentions of the authors of DCGAN, which attempted to make the generator and discriminator equal, seem to have become a de facto standard.