search

sjenni / temporal-ssl

Video Representation Learning by Recognizing Temporal Transformations. In ECCV, 2020.
https://sjenni.github.io/temporal-ssl/
GNU General Public License v3.0
48 stars 1 forks source link

About two softmax outputs #3

Closed 321hallelujah closed 3 years ago

321hallelujah commented 3 years ago

Hi, When SSLtraining In your paper, you claim to use two softmax for the pseudo-task. However, in this code, it seems you use tf.split to split the prediction（eg. 8 class for all, and 4 for skip ,the other 4 for transforms）, which lead to only one softmax. what is the difference about this, or just my misunderstanding. Looking forward to your reply.

sjenni commented 3 years ago

Hi, Thanks for the question. The softmax is applied in the loss function, so in the end, two softmax functions are applied. The code with splitting should be equivalent to two separate heads. Hope that clarifies the issue? :) Cheers

321hallelujah commented 3 years ago

I got it! the split variable is just fully_connected output of C3D. So it can be split to any pseudo-task. The softmax is applied in every loss function. Finally, the learned temporal-spatial representation is a concatenation style of two pseudo-task. Is it right? Thanks