qiexing / adaptive_weighted_attribute Goto Github PK

As mentioned in the paper, trend = abs(cur_mean - pre_mean) / cur_mean. So for one particular attribute, if cur_mean equals pre_mean, then trend = 0, norm_trend = 0 and λ = 0. Does this mean in the next k iterations the weight of this attribute is 0 when calculating loss? If so, how does this make sense?

Besides, if cur_mean >> pre_mean, then trend ≈ 1. If cur_mean << pre_mean, then trend is a large number. Thus if the val loss of the latest k iterations is small and the val loss of the last 2k-k iterations is large, trend is more likely to be a large number. Does this mean λ is more likely to be a large number? What is the logic behind this or how this variation is related to weight?

Many thanks to you.

waht does data attributes or net outputs of 40 mean??thank you for your answer

Can you share the test code? I test the model in CelebA dataset. But only get the 89% acc in the eyeglasses attributes.

hello, when i test the trained model, i get the error, can you help to solve the problem

I training the model as your expressed and only got an average acc 74%, Is there any trick in training and testing?