Closed yorkerlin closed 9 years ago
What is more,
I think the adadelta
with momentum correction in http://climin.readthedocs.org/en/latest/adadelta.html, is not correct.
The formula about Δθ
should be similar to v
in http://climin.readthedocs.org/en/latest/rmsprop.html
You indeed found a bug there, thanks. I fixed it.
I could not find an error in the docs. Note that the use of \theta_{t + 1 \over 2} can be confusing, as it is described differently for rmsprop.
@bayerj It seems there is a bug in
adadelta.py
whenmomentum
is used. Themomentum
correction can be applied toadadelta
,rmrprop
and others stochastic updates. The potential bug is at the110-th
line ofadadelta.py
I think it should be
step1 = step_m1 * m
instead ofstep1 = step_m1 * m * self.step_rate
. Correct me if I am wrong.Note that in
rmsprop.py
, the160-th
line is, which is correct.