search

pumpikano / tf-dann

Domain-Adversarial Neural Network in Tensorflow
MIT License
627 stars 224 forks source link

please #19

Closed 116705792 closed 6 years ago

116705792 commented 6 years ago

hey,I have some questions. firstly,in your code : total_loss=domain_loss + predict_loss. but in the paper total_loss= domain_loss - lamda*predict_loss. secondly,this is because GRL layer?

jindongwang commented 6 years ago

I have these questions, too.

Engineero commented 6 years ago

@116705792 other way around. The paper is finding a saddle point of the objective function E = predict_loss - lambda*domain_loss (equation 9 in Ganin, 2016) by finding the parameters that minimize predict_loss and maximize domain_loss. Maximizing is the same as minimizing the inverse, so this is identical to finding the parameters that jointly minimize predict_loss + lambda*domain_loss. The lambda is implemented as l in the gradient reversal layer, flip_gradient.py.

116705792 commented 6 years ago

thank you very much! i m learn a lot