Conditional glow network and flexible residual block

[rafaelorozco]

Simple conditional network based on glow coupling layer and the cIIN idea of tensor catting the condition to the input of the residual block.

Includes an important change to the residual block that allows for user set the size of the output of the RB with keyword argument n_out:

function ResidualBlock(n_in, n_hidden; n_out=nothing, k1=3, k2=3, p1=1, p2=1, s1=1, s2=1, fan=false, ndims=2)
.
.
.
    W3 = Parameter(glorot_uniform(k1..., n_out, n_hidden))
.
.
.

This change should be backwards compatible.

I want to use this network to showcase the MNIST non-conditional and conditional generation. This network doesn't necessarily perform better than conditional hint (same performance as far as my tests) but is much more lightweight.

[codecov[bot]]

Codecov Report

Merging #59 (25e999f) into master (a068423) will increase coverage by 0.18%. The diff coverage is 90.22%.

@@            Coverage Diff             @@
##           master      #59      +/-   ##
==========================================
+ Coverage   87.88%   88.06%   +0.18%     
==========================================
  Files          31       33       +2     
  Lines        2327     2439     +112     
==========================================
+ Hits         2045     2148     +103     
- Misses        282      291       +9     

Impacted Files	Coverage Δ
src/InvertibleNetworks.jl	60.00% <ø> (ø)
...rc/networks/invertible_network_conditional_glow.jl	82.08% <82.08%> (ø)
src/conditional_layers/conditional_layer_glow.jl	97.72% <97.72%> (ø)
src/layers/layer_residual_block.jl	98.70% <100.00%> (+0.01%)	:arrow_up:
src/utils/activation_functions.jl	89.23% <0.00%> (+6.15%)	:arrow_up:

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.