search

uqzhichen / GSMFlow

[ACM MM 2021] Mitigating Generation Shifts for Generalized Zero-Shot Learning
6 stars 3 forks source link

mask && dropout #5

Closed Xiao-lv-lol closed 2 years ago

Xiao-lv-lol commented 2 years ago

Hi Zhi，

Excuse me, why did you set the hyper-parameter dropout to 0 in your code? The output of mask is always True, mask has lost its meaning!I don't find any explanation on this aspect in your paper. Could you give me some advice?

uqzhichen commented 2 years ago

Hi Xiao,

The mask on the perturbation noise is not necessary on all datasets. You may try different dropout rate to see the impact.

Cheers, Zhi