sarthak268 / discont Goto Github PK

Dear Sarthak,

Many thanks for the excellent paper and publishing your code on GitHub. Much appreciated!

I have a question regarding the collection of the transformed chunks in line 164 of the training loop contained in the train.py module. Shouldn't it read:

transformed_chunks[i * FLAGS.z_num_chunks : (i+1) * FLAGS.z_num_chunks] = curr_cv_full_view

instead of:

transformed_chunks[i*FLAGS.z_num_chunks : (i+1)*FLAGS.z_num_chunks]

to collect the distinct transformed chunks and to subsequently compute the center loss? I fear that otherwise, the information contained in the transformed_chunks tensor won't contain any updated data other than the initialisation.

Thanks again,
Marco

Hello Sarthak268,

Thanks for sharing your code. The idea is great.

I checked the code and I have a question,

You create variable cv_full_view:

If I'm not wrong, I think the idea is to use the stored cv_full_view (centers) to get the center_loss, update it, then keep doing this procedure, right? Otherwise, there is no point to do:

Thanks again! Great paper!