thuml / domain-adaptation-regression Goto Github PK

Hi. This is great work. I was wondering if you could also provide the code to run the baselines like DAN and DANN for regression. What changes did you make?

can you share the dataset of the dsprites again? the link is invalid. thank you!

作者你好，我在使用该代码进行图像回归任务解决域适应问题时遇到一个bug；就是代码训练二十多个epoch后就开始出现提醒：
Intel MKL ERROR: Parameter 4 was incorrect on entry to SLASCL.
然后再几个epoch后就报错：
Traceback (most recent call last): File "train_rsd.py", line 134, in <module> rsd_loss = rsd_f(feature_s, feature_t) File "/home/digisky/anaconda3/envs/torch1.2/lib/python3.6/site-packages/torch/nn/modules/module.py", line 547, in __call__ result = self.forward(*input, **kwargs) File "/home/digisky/ztj/DAR_obj/rsd.py", line 27, in forward U_s, _, _ = torch.svd(f_s.t()) RuntimeError: svd_cuda: the updating process of SBDSDC did not converge (error: 14)
我也在网上找了相关解决办法，但是都没有成功。然后我发现是在某个epoch后，梯度中出现了nan和inf，所以导致更新的权重也变为了nan，网络正向计算得到的特征中也就全部变为nan，传入到rsd loss中，svd计算就报上述的错误。请问作者有遇到过这类问题吗？有什么思路或者解决的建议吗？
谢谢！

In the training, I got this wrong as follows.

tensor([[nan, nan, nan, ..., nan, nan, nan],
[nan, nan, nan, ..., nan, nan, nan],
[nan, nan, nan, ..., nan, nan, nan],
...,
[nan, nan, nan, ..., nan, nan, nan],
[nan, nan, nan, ..., nan, nan, nan],
[nan, nan, nan, ..., nan, nan, nan]], device='cuda:0',
grad_fn=)
The tensor is the feature of source obtained by the model.

Traceback (most recent call last):
File "train_rsd.py", line 212, in
rsd_loss = RSD(feature_s,feature_t)
File "train_rsd.py", line 133, in RSD
u_s, s_s, v_s = torch.svd(Feature_s.t())
RuntimeError: svd_cuda: For batch 0: U(37,37) is zero, singular U.

Thanks for the interesting work.
I was able to reproduce your results on the two datasets (Dsprites and MPI3D).
I would like to know if the implementation details for the third dataset (Biwi Kinect) can be available?

When i use RSD loss, the following error will occur:

Intel MKL ERROR: Parameter 4 was incorrect on entry to SLASCL.
Traceback (most recent call last):
File "train_SVD.py", line 274, in
main(args)
File "train_SVD.py", line 118, in main
rsd_loss = RSD(f_s,f_t)
File "train_SVD.py", line 29, in RSD
u_s, s_s, v_s = torch.svd(Feature_s.t())
RuntimeError: svd_cuda: the updating process of SBDSDC did not converge (error: 18)

Thank you very much for the open source provided, and this is also a very good work!
However, I have the following questions, I hope the author can answer them when they are free, thank you very much!
1.The BMP loss proposed by the author is used to solve the UDA problem, but it is found that the effect of adding it is even worse. I don't know if it is for other reasons?
2.The Trade-off coefficient corresponding to BMP loss is set very small, can it really explain its benefits?
3.When running RSDloss, the code will report an error.
Best wishes to you, hope to get your reply in the near future！