Closed EiffL closed 2 years ago
Hi there!
Yes, and I just adapted the example @Justinezgh proposed in #2. It seems to work well and quit fast, and for many k>2!
Some examples with a = 1
, b = 1
and c = 0.5
:
Code here: https://colab.research.google.com/drive/1bULsH6McxKy_IqViRpx0IgvCfnnSlwGp?usp=sharing
ok, so.... almost there. I have added the corrercted couplping layer in #10 and it's training \o/ buuuut it's giving me some weird reresuts
there must be somehiing wrong, maybe in the logdet computationn....
Ok i think this is done, will be closed as soon as we merge #10.
I'm opening this issue to discuss an idea from #2 to try to use the implicit function theorem to define the inverse and jacobian of general bijections, in order to have a general way of defining smooth coupling layers.
@b-remy had some examples at some point of how to do that in JAX, before we switched to TF. Do you still have your JAX prototype Benjamin?