Add LRP composites

[adrhill]

Closes #82.
Adds Composite, which sequentially applies "Composite Primitives":

To apply a single rule, use:
- LayerRule to apply a rule to the n-th layer of a model
- GlobalRule to apply a rule to all layers of a model
- RangeRule to apply a rule to a positional range of layers of a model
- FirstRule to apply a rule to the first layer of a model
- LastRule to apply a rule to the last layer of a model

To apply a set of rules to multiple layers, use:
- GlobalRuleMap to apply a dictionary that maps layer types to LRP-rules
- FirstNRuleMap for a RuleMap on the first n layers of a model
- LastNRuleMap for a RuleMap on the last n layers
- RangeRuleMap for a RuleMap on generalized ranges

Example

Using a flattened VGG11 model:

julia> composite = Composite(
           GlobalRuleMap(
               ConvLayer => AlphaBetaRule(),
               Dense => EpsilonRule(),
               PoolingLayer => EpsilonRule(),
               DropoutLayer => PassRule(),
               ReshapingLayer => PassRule(),
           ),
           FirstNRuleMap(7, Conv => FlatRule()),
       );

julia> analyzer = LRP(model, composite);

julia> analyzer.rules
19-element Vector{AbstractLRPRule}:
 FlatRule()
 EpsilonRule{Float32}(1.0f-6)
 FlatRule()
 EpsilonRule{Float32}(1.0f-6)
 FlatRule()
 FlatRule()
 EpsilonRule{Float32}(1.0f-6)
 AlphaBetaRule{Float32}(2.0f0, 1.0f0)
 AlphaBetaRule{Float32}(2.0f0, 1.0f0)
 EpsilonRule{Float32}(1.0f-6)
 AlphaBetaRule{Float32}(2.0f0, 1.0f0)
 AlphaBetaRule{Float32}(2.0f0, 1.0f0)
 EpsilonRule{Float32}(1.0f-6)
 PassRule()
 EpsilonRule{Float32}(1.0f-6)
 PassRule()
 EpsilonRule{Float32}(1.0f-6)
 PassRule()
 EpsilonRule{Float32}(1.0f-6)

TODO

• Pretty-printing of Composites
• Document Composites in advanced LRP docs

Future work

Moved to PR #85:

• Add presets similar to those in Zennit, e.g.:
  • EpsilonGammaBox
  • EpsilonPlus
  • EpsilonAlpha2Beta1
  • EpsilonPlusFlat
  • EpsilonAlpha2Beta1Flat
• Update README with presets

[codecov-commenter]

Codecov Report

Merging #84 (271709d) into master (084c962) will increase coverage by 0.29%.
The diff coverage is 97.50%.

@@            Coverage Diff             @@
##           master      #84      +/-   ##
==========================================
+ Coverage   94.23%   94.53%   +0.29%     
==========================================
  Files          13       14       +1     
  Lines         399      439      +40     
==========================================
+ Hits          376      415      +39     
- Misses         23       24       +1     

Impacted Files	Coverage Δ
src/lrp/composite.jl	97.22% <97.22%> (ø)
src/lrp/lrp.jl	96.42% <100.00%> (+0.59%)	⬆️

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