Closed alisiahkoohi closed 2 years ago
Merging #27 (8c2a3ad) into master (e22195d) will increase coverage by
0.01%
. The diff coverage is100.00%
.
@@ Coverage Diff @@
## master #27 +/- ##
==========================================
+ Coverage 84.16% 84.18% +0.01%
==========================================
Files 31 31
Lines 2476 2479 +3
==========================================
+ Hits 2084 2087 +3
Misses 392 392
Impacted Files | Coverage Δ | |
---|---|---|
src/conditional_layers/conditional_layer_hint.jl | 94.85% <100.00%> (ø) |
|
src/layers/invertible_layer_actnorm.jl | 87.50% <100.00%> (ø) |
|
src/layers/invertible_layer_basic.jl | 90.78% <100.00%> (ø) |
|
src/layers/invertible_layer_hint.jl | 92.79% <100.00%> (ø) |
|
...rc/networks/invertible_network_conditional_hint.jl | 89.38% <100.00%> (ø) |
|
src/utils/parameter.jl | 64.70% <100.00%> (+1.62%) |
:arrow_up: |
Continue to review full report at Codecov.
Legend - Click here to learn more
Δ = absolute <relative> (impact)
,ø = not affected
,? = missing data
Powered by Codecov. Last update e22195d...8c2a3ad. Read the comment docs.
Can we add a test for it? Except that looks fine
Can we add a test for it? Except that looks fine
Yes, I will think about a simple test.
Just saw above, it also has conflicts with master so definitely need rebase (git rebase -ii origin/master
as I showed you the other time)
This pull request adds a flag to
forward
andinverse
functions fromNetworkConditionalHINT
andConditionalLayerHINT
, which allows for conditional density estimationp(x|y)
using the change-of-variable formula. Currently, the joint densityp(x, y)
can be easily estimated, whereas for the conditional density we need thelogdet
term associated with thex
-lane only. This amounts to only accumulatinglogdet
of the Jacobian with respect tox
rather than with respect to(x, y)
.After this change, given a trained
NetworkConditionalHINT
on pairs(x, y)
, the following change-of-variable formula can be used