UNetGenerator

[xuwanpeng12]

Hi:
I found that your deep generation network abandoned the traditional resnet and adopted a structure similar to resnet+inception as the backbone network. What is the point of doing this? Is it owing to use of resnet directly not effective?

[lyndonzheng]

Hi @xuwanpeng12, as the generative adversarial learning is sensitive to the network structure, our generation structure follows the work of CycleGAN. At that time, the ResNet performs worse on the generation task. While the latest work, e.g. stylegan, stargan, tried to design a deeper generation network, they still use a simple way instead of a very deep Resnet.

[xuwanpeng12]

Is it due to the depth of the network that resnet does not perform well in the process of generative adversarial learning? So add inceptionblock at the end to expand the width of the network?