err_t_domain and err_s_domain
InstantWindy opened this issue · comments
Hi. I refer to your network design and loss to train my dataset for this experiment. When traing the network, I try to output the total loss of err_t_domain and err_s_domain, the total loss decreases a little at first and then remain unchanged. Thanks.
If your data are similar to mnist and mnistm, you can try to adjust the lambda value.
My source data and target data are pedestrians. The value of err_domain has always been 1.3863 and does not drop.
Pedestrians are far more complex than digits, I think it may be underfitting in your experiment.
Thank for your reply. if the err_domain loss doesn't drop, it means my source data and target data are not similar?
Underfitting means the model cannot transfer the information between source and target domain, either because of the data or your model.
My task is not classification, but super resolution. I just use the super resolution network to extract features, and then input the features into the domain classification for discrimination.
Oh, I think this network can only extract class level information, pixel information may be very difficult.
Which network do you mean? super resolution network?
DANN,I mean, it is not strong enough to transfer pixel information.
Thanks.