eveningdong / daosl Goto Github PK

Hello! I would like to ask something about the formulation below:

In my opinion, the highlighted part should be p(y=0|...), without the "1-". And your code in ada.py seems also prove it:

So I feel a bit confused and want to turn to you for help. Is my thought wrong?

Hi, I'd like to ask some question about the paper and the code:

1. The paper says in section 3.4 that

   If the target query image can be correctly classifi ed, the target image is "close" to the corresponding image in the projected feature space.

   Does it mean when we train ADA+RSS, we would use the target domain label? If so, I think it's a little unfair to compare to the baseline such as PN+ADDA. Did I misunderstand something?

   EDIT at 6/19: For question 1, I finally realized that the proposed method did not use the target domain label, I misunderstood the paper, sorry about that.

2. Is there typo in the train_rss.py? It seems that train_query_image and train_query_label are mismatched, maybe the first one should be corrected?

   https://github.com/leonndong/DAOSL/blob/master/train_rss.py#L215

   feed_dict={
             train_support_images: x_support_set,
             train_support_labels: y_support_set, 
             train_query_image: val_x_query, # is this typo?
             train_query_label: y_query, 
             val_support_images: val_x_support_set,
             val_support_labels: val_y_support_set,  # also, I think these rows are redundant?
             val_query_image: val_x_query,  # also, I think these rows are redundant?
             val_query_label: val_y_query # also, I think these rows are redundant?
             })

   EDIT: For question 2, I think it's not typo in that line. It matches the paper's approach.