about loss function about depict HOT 4 CLOSED

from depict.

line827 in functions.py
https://github.com/herandy/DEPICT/blob/master/functions.py

from depict.

If you look at equation 12 in the paper, also line 3 in the algorithm, the second part of the loss function which contains z_i actually contains what we are doing at line 827, using p_i instead of z_i should be roughly equivalent.

from depict.

Considering any hyper-parameter for KL(f||u) in eq. (4) is equivalent of power in nominator of eq. (7). We set this hyper-parameter equal to 2 for all datasets. Intuitively, setting higher weight for KL(f||u) in the loss function is more helpful for the datasets with uniform class labels.

from depict.