why 'log_px = (log_pz + delta_logp - log_qh_x + log_pN)'?
qiangbo1222 opened this issue · 2 comments
qiangbo1222 commented
in the compute loss code you define log_px = (log_pz + delta_logp - log_qh_x + log_pN), but in the original paper of ffjord: 'log_pz = log_pz - delta_logp '
is there a mistake? or did I miss something in the loss computing code
vgsatorras commented
Hi,
When estimating the log-likelihood in our model we also have to include a term for the dequantization of discrete features "log_qh_x" and another one for the number of nodes "log_pN". You can find all the terms you mentioned depicted in Figure 2 of our paper. These two additional terms are explained in more detail at the end of Section 4.
Best,
Victor
qiangbo1222 commented
Thanks a lot, The explanation in the paper really helps me.
Best,
Bo