Hi @agonzgarc , I'm very interested in your work and have read your paper carefully. In Section 2.1, you mentioned that directly calculating the L1 loss between the shared codes of two different domains may not increase the similarity between them, so you proposed to add small noise to the output of the encoder to yield better results. However, I found that you did not add such noise in your code, may I ask why? Looking forward to your reply!
Hi, please note that the noise is added right before the shared code is input in the decoder rather than for the L1 loss. This still prevents the model from relying on shared codes of very low magnitude.
Hi @agonzgarc , I'm very interested in your work and have read your paper carefully. In Section 2.1, you mentioned that directly calculating the L1 loss between the shared codes of two different domains may not increase the similarity between them, so you proposed to add small noise to the output of the encoder to yield better results. However, I found that you did not add such noise in your code, may I ask why? Looking forward to your reply!