xinntao/dni

CVPR19 - Deep Network Interpolation for Continuous Imagery Effect Transition

Hi there 👋

Homepage: https://xinntao.github.io/

GoogleScholar: https://scholar.google.com.hk/citations?user=FQgZpQoAAAAJ

dni's Issues

day2night datasets

Thank you for sharing your awesome work.
I was very impressed by vividness of day to night example. According to Sec 4.2 of your paper, the translation algorithm base on CycleGAN or Pix2Pix, right? Could you tell me which datasets did you use?

Hi, Can I ask how to do fine-tune when you train GanEffect.

Hi, Can I ask how to do fine-tune when you train Gan Effect.
I think you use ESRGAN model for this fine-tune.
but i don't know what kinds of layer for fine-tune, and numbers of layer.

can you explain how you transfer the model PSNR Effect to Gan Effect. :)

老师，我训练esrgan网络，同时用另一个网络只有l1，发现两个生成网络参数插值后推理图片变得很暗？

老师，我训练esrgan网络，同时用另一个网络只有l1，发现两个生成网络参数插值后推理图片变得很暗？
这是什么原因呀，请求指点下

L1的结果

Gan的结果

两者插值的结果

def load_network_interpolation(load_path, network, load_path_l1, l1_a=0.5, strict=True):
    load_net = torch.load(load_path)
	load_net_l1 = torch.load(load_path_l1)
    load_net_interpolation = OrderedDict()  
    for k, v in load_net.items():
        v_l1 = load_net_l1[k]
        load_net_interpolation[k] = v*(1-l1_a) + v_l1*l1_a
    network.load_state_dict(load_net_interpolation, strict=strict)