Thanks for your impressive work! When I run your code in my workstation, it always reports "[easydl] tensorflow not available!" And I satisfy the requirements which you mentioned in "readme". What should I do?

您好，最近拜读Universal Domain Adaptation，对UAN算法非常感兴趣，谢谢你们分享代码！在调试代码的过程中，我遇到了几个问题，想请教一下： 1.pretrained_model可以采用离线下载的pytorch预训练模型吗？例如，通过‘https://download.pytorch.org/models/resnet50-19c8e357.pth’下载ResNet-50的预训练模型进行实验 2.在Google Colab上运行“！python main.py --config office-train-config.yaml”时，遇到了图中的报错信息，该如何解决？ 感谢你的阅读，期待回复！

Hello, after reading the paper and the code, I have a few questions about them. Wish you could answer them for me :) The UAN model proposed in the paper looks like this

. In my view, the feature z extracted from F is loaded into G, D, D' separately. However, the model in the code is defined as follow:

def forward(self, x):
        f = self.feature_extractor(x)
        f, _, __, y = self.classifier(f)
        d = self.discriminator(_)
        d_0 = self.discriminator_separate(_)
        return y, d, d_0

The feature f goes into classifier first but the output of the first layer of the classifier goes into the discriminator? Why?

Besides, the training of the network can be seen as a minimax game of equation (4). However, the code just added all the loss up loss = ce + adv_loss + adv_loss_separate. How to understand this?

Thanks

The paper and the released code miss some training details about the mentioned dataset. Could you provide the yaml files of office-31, visda 2017 and imagenet-caltech? In addition, the usage of dataset imagenet-caltech is quite different in paper SAN and PADA. Could you provide the label files of imagenet-caltech used in the paper? Thanks!

self.grl = GradientReverseModule(lambda step: aToBSheduler(step, 0.0, 1.0, gamma=10, max_iter=10000))

It seems that max_iter is set as 10000.

From the DANN paper, I thought it depends on the total iteration number during training. Is there any reason to use a fixed number?

Thank you.

Hello, I have read the paper and code, and not found the details about how to calculate the thresholdw_0 and hyper-parameter λ.

Specifically, w_0 is seemed to be a fixed value for every dataset in the code, which confused me.

For different datasets, should calculate different w_0 and λ? and how to define a w_0 to help us distinguish unknown samples from others in the target domain? Or I missed some details, please help give me some suggestions. Thank you

def normalize_weight(x): min_val = x.min() max_val = x.max() x = (x - min_val) / (max_val - min_val) x = x / torch.mean(x) return x.detach()

according to paper, x in (0,1), why normalize_weight need div torch.mean(x)?

Interesting work! I have read this paper and agreed with the assumption about Domain similarity and Prediction uncertainty, but I am confused about the calculation method of w^s(x) in eq.7 and w^t(x) in eq.8. What is the exact meaning of uncertainty minus similarity (eq.7)? And why are w^s(x) and w^t(x) calculated in opposite ways?

Look forward to your reply! Thanks

You had mentioned that the backbone network is ResNet-50 pretrained on Imagenet.

https://github.com/thuml/Universal-Domain-Adaptation/blob/5d7caa95af7e3675305c542253c4e372801897d2/net.py#L37

But in many experiments in the paper, the labels in the unsupervised target overlap with that of the supervised source labels in the ImageNet. Is it justified that you pretrain on supervised ImageNet labels?

Shouldn't there be non-linearity between the bottleneck layers in the classifier?

https://github.com/thuml/Universal-Domain-Adaptation/blob/5d7caa95af7e3675305c542253c4e372801897d2/net.py#L127