search

NVIDIA / pix2pixHD

Synthesizing and manipulating 2048x1024 images with conditional GANs
https://tcwang0509.github.io/pix2pixHD/
Other
6.65k stars 1.39k forks source link

Transpose convolution parameters #149

Open mrgloom opened 5 years ago

mrgloom commented 5 years ago

In downsampling part nn.ConvTranspose2d with kernel_size=3, stride=2 is used but seems with these parameters at some parts 3x3 kernel will cover only one input pixel, seems this will lead to checkerboard artifacts. Is it true? or maybe it 'was fixed' somewhere in other place of network for example conv with large kernel size at output? https://github.com/NVIDIA/pix2pixHD/blob/master/models/networks.py#L203

https://github.com/vdumoulin/conv_arithmetic/blob/master/gif/padding_strides_odd_transposed.gif

vatsalag99 commented 4 years ago

Hi, I am following up on this. Currently, my images have few gridding effects, however, would this be fixed when training with the local enhancer or would I need to restart training on the global generator?

liujiaoshou17 commented 2 years ago

Hi, I am following up on this. Currently, my images have few gridding effects, however, would this be fixed when training with the local enhancer or would I need to restart training on the global generator?

I also had the same problem，did you solve it