Closed wuaalb closed 8 years ago
Hi
Looks good with the log-sum-exp and the sd->std. I'm not sure we should change the interface again for the eps->epsilon again though?
Casper
Was just for coherence with lasagne.layers.BatchNormLayer
or lasagne.updates.adagrad
for instance, but yeah, maybe it's too late now..
Lemme know if I should revert that change..
Hi again
yes I think we should revert the epsilon change (If you do not strongly object), otherwise I think it looks great!
I've changed back to eps
.
I changed the docstrings to clarify what eps
does exactly for each function, as it depends slightly on the parameterization used. There were also some T.log()
s that didn't use eps
.
Also added normal1
versions of the closed-form KL-divergences.
Extracted log-sum-exp from examples and added a log normal pdf parameterized with variance. Also changed
sd
tostd
andeps
toepsilon
for consistency with lasagne and numpy.