bit-da/gdcan

[TPAMI 2021] Code release for "Generalized Domain Conditioned Adaptation Network" https://arxiv.org/abs/2103.12339

gdcan's Issues

About the channel attention

Thank you for your excellent job about the condition domain. I am very insteresed in the novel channel attetion. In this project, images of source domain and target domain are concated. If images of different domains are feed into different parts of the FC layers separatively, does it still work? Is there any difference? Thank you in advance for your kind help.

The freezing BN setting during training.

Hello, thank you for the excellent job and the code. However, I'm confused about the training setting in train_dcan.py.

# freeze BN layers for m in base_network.modules(): if isinstance(m, nn.BatchNorm2d): m.training = False m.weight.requires_grad = False m.bias.requires_grad = False

Why freeze the BN layers during training?

About the paper "Generalized Domain Conditioned Adaptation Network" (TPAMI-2020-06-0743)

Thank you very much for your excellent job. So sorry to bother you. However, I can't find this paper from the internet, Can you share this excellent paper "Generalized Domain Conditioned Adaptation Network" (TPAMI-2020-06-0743)。