g-pichler / knife Goto Github PK

I have a question about the table 2 in your paper. You mentioned in the "setup" section that your model was evaluated on the dev set of GLUE following VIBERT. However, in their paper section 3.1, they mentioned their results are obtained from GLUE test set. So I wonder on which split you did the experiments for Table 2, dev or test.

Hi,

I'm currently playing your KNIFE estimator, and I wonder if there is a typo or not in the computation of the log_pdf of the Conditional Kernel:

If I understand correctly, in the forward pass (or learning loss) of KNIFE estimator, we want to estimate both p(z_d) and p(z_d | z_c).
For p(z_d | z_c) we thus use the CondKernel class, in which the conditional density p(z_d | z_c) is estimated by using different FF layers to compute means, logvar, etc. from z_c samples.

Hence, I think that the initialization steps of these layers carry some typos w.r.t. the dimensions:

self.mu = FF(args, self.d, self.d, self.K * self.d)
self.logvar = FF(args, self.d, self.d, self.K * self.d)
self.weight = FF(args, self.d, self.d, self.K)
self.tri = FF(args, self.d, self.d, self.K * self.d * self.d)

with self.d = zd_dim

If we suppose now that we have different dimensions for z_c and z_d, i guess that the input dimension of these layers should rather be zc_dim, if I'm not wrong ?

Thank you in advance for your answer.