Summ net #82
Summ net #82
Conversation
Codecov ReportPatch coverage:
Additional details and impacted files@@ Coverage Diff @@
## master #82 +/- ##
==========================================
+ Coverage 88.47% 88.52% +0.05%
==========================================
Files 33 34 +1
Lines 2577 2589 +12
==========================================
+ Hits 2280 2292 +12
Misses 297 297
☔ View full report in Codecov by Sentry. |
| end | ||
|
|
||
| @Flux.functor NetworkConditionalGlow | ||
|
|
||
| # Constructor | ||
| function NetworkConditionalGlow(n_in, n_cond, n_hidden, L, K;freeze_conv=false, split_scales=false, rb_activation::ActivationFunction=ReLUlayer(), k1=3, k2=1, p1=1, p2=0, s1=1, s2=1, ndims=2, squeezer::Squeezer=ShuffleLayer(), activation::ActivationFunction=SigmoidLayer()) | ||
| function NetworkConditionalGlow(n_in, n_cond, n_hidden, L, K; freeze_conv=false, split_scales=false, summary_net = nothing, rb_activation::ActivationFunction=ReLUlayer(), k1=3, k2=1, p1=1, p2=0, s1=1, s2=1, ndims=2, squeezer::Squeezer=ShuffleLayer(), activation::ActivationFunction=SigmoidLayer()) |
There was a problem hiding this comment.
How is this different than just the composition sumarry \odot G ?
The whole point of Chain/odot is to easily combine/chain networks and have it work well together, I don't particularly like having networks as input of another like that.
There was a problem hiding this comment.
The main difference is that composition is not defined for conditional networks.
Practically:
I don't think it will work without some extra work because only one of the inputs to G goes into summary.
(x,y) -> G(x, summary(y))
And summary isnt necessarily invertible so the backward needs to be provide the input to the summary for the gradient calculation (C_save here)
Zx, Zy, lgdet = G.forward(X, Y)
dx, x, dy = G.backward(Zx, Zx, Zy; C_save = Y)
So it will not play nicely with the composition operator that was built for invertible networks.
That said, I agree that it isnt pretty and if these things are easy to solve I am all for it.
|
How about something like this previous commit @mloubout ? And the main operations are the same as before |
First try at adding a summary network to conditional glow.
I could use some extra eyes on how to implement this cleanly. Although it gets the job done and works with arbitrary flux layers, it feel a little bit clunky.