slimgroup / InvertibleNetworks.jl

A Julia framework for invertible neural networks
MIT License
152 stars 23 forks source link

make NetworkMultiScaleConditionalHINT reversible and use logdet correctly #24

Closed rafaelorozco closed 2 years ago

rafaelorozco commented 3 years ago

For NetworkMultiScaleConditionalHINT -Added backward_inv so that reverse() works -Add logic so that logdet can be turned off for mse based training.

Other changes: -exposed activation function to user when making Glow network. -Change Z_dims to Array{Array} so that we can initialize it and can get passed to gpu without a error inducing undef

codecov[bot] commented 3 years ago

Codecov Report

Merging #24 (1c566fc) into master (7d21194) will decrease coverage by 0.73%. The diff coverage is 80.00%.

Impacted file tree graph

@@            Coverage Diff             @@
##           master      #24      +/-   ##
==========================================
- Coverage   86.50%   85.77%   -0.74%     
==========================================
  Files          31       31              
  Lines        2460     2509      +49     
==========================================
+ Hits         2128     2152      +24     
- Misses        332      357      +25     
Impacted Files Coverage Δ
.../invertible_network_conditional_hint_multiscale.jl 75.93% <70.65%> (-11.49%) :arrow_down:
src/layers/invertible_layer_glow.jl 97.61% <93.75%> (ø)
src/networks/invertible_network_glow.jl 99.09% <100.00%> (ø)
src/utils/activation_functions.jl 83.33% <100.00%> (+1.28%) :arrow_up:
src/utils/dimensionality_operations.jl 94.57% <100.00%> (+0.23%) :arrow_up:
src/utils/neuralnet.jl 67.69% <0.00%> (+1.53%) :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 7d21194...1c566fc. Read the comment docs.

mloubout commented 3 years ago

Any update?