hytseng0509 / drit_hr Goto Github PK

some mistakes
It may be :
git clone https://github.com/hytseng0509/DRIT_hr.git
cd DRIT_hr

Hello,

what is the difference to the original code ? Why is this repository well suited for high resolution images? Which resolution of GTA and Cityscapes can you recommend for training ?

Best Regards

Hello!
I found a following thing in LeakyReLUConv2d:

class LeakyReLUConv2d(nn.Module):
  def __init__(self, ..., norm='None', ...):
    ....
    if 'norm' == 'Instance': 
      model += [nn.InstanceNorm2d(n_out, affine=False)]
    ...

https://github.com/hytseng0509/DRIT_hr/blob/master/networks.py#L496

It seems that normalization is never applied in LeakyReLUConv2d block.

I'm curious if it affect the model performance, as LeakyReLUConv2d present in all Encoders and Discriminators?

Are the best results reported in paper are gained with turned on InstanceNormalization?

Best Regards,
Aleksei Silvestrov

Thanks for your great work.
I want to reproduce the experiment of high resolution (GTA to Cityscapes). How many images are in the two datasets, and are they from training set or testing set? And I also confuse about the training setting, such as which training epoch is better?
Hope for your reply. Many thanks!