julia-xai / explainableai.jl Goto Github PK

Add PassRule that implements Rₖ .= Rₖ₊₁.

• Instead of having rules modify layer parameters, avoid allocations by implementing modified forward calls that can be diff'ed.
• Pre-allocate buffers
  • for activations on forward-pass
  • for relevances on backward-pass
• For analysis of multiple output neurons, only run forward-pass once

This should speed things up a lot!

Support for nested Chains was introduced in #119, but canonization currently still requires flattening the model.

https://github.com/JuliaTesting/Aqua.jl

The types of primitives currently end with the word Rule and could be mistaken with LRP rules.

Add a Shapley analyzer when ShapML is loaded. LazyModules.jl could be used for this purpose.

Refer to Zennit implementation.

This will be the first step towards #10 by allowing nested model structures.

This requires the following changes:

• new internal representation of rules, e.g. via LRPRulesChain / LRPRulesParallel [Edit: now called ChainTuple and ParallelTuple] to support graphs
• treat chains the same way as layers by adding lrp!(Rₖ, r::AbstractLRPRule, c::Chain, aₖ, Rₖ₊₁), which can be called recursively. This will require a different approach to pre-allocating activations and relevances than the one currently used in the call to the analyzer.
• Composite might require a refactoring
• Update lrp/show.jl

Run benchmarks on a representative model, e.g. VGG16.

Add function canonize(model) which merges BatchNorm layers into Dense and Conv layers with linear activation functions.

GPU support is currently untested. In theory, GPU tests could be run on CI using the JuliaGPU Buildkite CI.

Locally, a first test of GPU support can be run by modifying the readme example to cast input to a GPU array.
It should be possible to run the following code in a fresh temp-environment:

using CUDA
using ExplainableAI
using Flux
using MLDatasets
using Downloads: download
using BSON: @load

model_url = "https://github.com/adrhill/ExplainableAI.jl/raw/master/docs/src/model.bson"
path = joinpath(@__DIR__, "model.bson")
!isfile(path) && download(model_url, path)
@load "model.bson" model

model = strip_softmax(model)
x, _ = MNIST.testdata(Float32, 10)
input = reshape(x, 28, 28, 1, :)

input_gpu = gpu(input) # cast input to GPU array
analyzer = LRP(model)
expl = analyze(input_gpu, analyzer)

This should be handled via external packages DataAugmentations.jl.

• document preprocessing with external packages

Since this is a breaking change, it should be implemented before a 1.0 release.

This issue is used to trigger TagBot; feel free to unsubscribe.

If you haven't already, you should update your TagBot.yml to include issue comment triggers.
Please see this post on Discourse for instructions and more details.

If you'd like for me to do this for you, comment TagBot fix on this issue.
I'll open a PR within a few hours, please be patient!

This issue keeps track of which Flux layers in the model reference got LRP implementations.

Basic layers

• Dense
• flatten

Convolution

• Conv
• DepthwiseConv
• ConvTranspose
• CrossCor

Pooling layers

• AdaptiveMaxPool
• MaxPool
• GlobalMaxPool
• AdaptiveMeanPool
• MeanPool
• GlobalMeanPool

General purpose

• Maxout
• SkipConnection
• Chain #119
• Parallel #10
• Bilinear
• Diagonal
• Embedding

Normalisation & regularisation

• normalise
• BatchNorm #129
• dropout
• Dropout
• AlphaDropout
• LayerNorm
• InstanceNorm
• GroupNorm

Upsampling layers

• Upsample
• PixelShuffle

Recurrent layers

• RNN
• LSTM
• GRU
• Recur

The following should be implemented before registering the package:

• #3
• #4
• #5

• Check model for non-ReLU-like activations and unknown layers.
• Display a summary as to why checks failed if they do, as well as references to the docs on how to fix these issues.
• Make checks skip-able through LRP kwarg skip_checks=true.

The goal should be to make ExplainabilityMethods transparent but extendable.

Loading VGG and preprocessing the input currently takes up a lot a space in the docs.

• Use a MNIST CNN with MLDatasets to make docs more concise.
• Show how to use custom layer functionality from #26.

Update README to reflect #157.

CIFAR10 in the docs and readme should be more interesting than MNIST.

Replace Zygote dependency with AbstractDifferentiation.jl for backend agnostic AD.

Add the Threads.@threads macro to the for-loop in gradients_wrt_batch:
https://github.com/adrhill/ExplainableAI.jl/blob/f1b89ab9c784ff2a86de9997efa904580b2af6dd/src/gradient.jl#L5-L16

Using loops over pre-allocated arrays or StackViews.jl should speed up things and help with type stability.

All gradient analyzers currently use mapreduce:
https://github.com/adrhill/ExplainableAI.jl/blob/8641dfb101b63ed5a9876b32988c25e0c6b2191d/src/gradient.jl#L5-L13

Methods using InputAugmentation would also benefit from refactoring:
https://github.com/adrhill/ExplainableAI.jl/blob/8641dfb101b63ed5a9876b32988c25e0c6b2191d/src/input_augmentation.jl#L74-L83

Refer to quickstart example: https://lorenzoh.github.io/DataAugmentation.jl/dev/docs/literate/quickstart.md.html

This would remove the dependency on Flux and make the package lighter for users that don't require LRP.

Since this requires a breaking release, this is a milestone for a future 1.0 release.

Update field names.
The field layerwise_relevances of the Explanation struct is too specific to LRP.
An extras field of type Union{Nothing, Dict} would be more flexible.

Could be implemented as the one-liner ZPlusRule() = AlphaBetaRule(1.0f0, 0.0f0),
but a large part of the computation could be skipped since β=0.

for more performant use of Tullio.jl.

Zygote's gradient appears to be non-deterministic on Metalhead's VGG19:

julia> a = gradient((in) -> model(in)[1], imgp)[1];

julia> b = gradient((in) -> model(in)[1], imgp)[1];

julia> isapprox(a, b; atol=1e-3)
false

julia> isapprox(a, b; atol=1e-2)
true

Check whether this is due to Dropout layers or Zygote.

ColorSchemes 3.18 adds :centered option to ColorSchemes.get that makes _normalize redundant.

A visualisation method such as plot_text_heatmap from this iNNvestigate notebook would be useful for language tasks such as sentiment analysis.

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

This would close #79 and be a major milestone towards a 1.0 release.

Wrap attribution in struct containing metadata such as used analyzer to dispatch on heatmap.
This struct could also contain the neuron selection and possibly the classifier output.

Similar to Flux's printing of Chains.

and add tests on LRP rules.

Flux.Parallel docs.

Using keys obtained by chainkey. A user-friendly version of chainkey must be exported and documented.

For a more efficient implementation of LRP rules.

Metalhead disabled pretrained weights in 0.6.0 due to model inaccuracies.
These can technically still be loaded while they are being fixed:

model = VGG19()
Flux.loadparams!(model.layers, weights("vgg19"))

However, the VGG19 weights are a 548 MB download every time CI is run. It might therefore be more reasonable to use a smaller model. Currently, MetalheadWeights contains (in ascending size):

• SqueezeNet (5 MB) -> requires Parallel for "fire" modules
• GoogLeNet (27 MB) -> requires Parallel
• Densenet121 (31 MB) -> requires SkipConnection
• ResNet-50 (98 MB) -> requires Parallel, skip_identity
• VGG-19 (548 MB)

An easy workaround would be to run the methods on randomly initialized parameters (with fixed seed). The explanations w.r.t. to this model should still stay constant.

Add doc pages for:

• Basic example on VGG model
• Combining LRP rules
• Custom LRP rules

Also:

• Add images to readme

Currently, the following things can be improved or are missing documentation:

• input augmentations: NoiseAugmentation, InterpolationAugmentation
• usage of LayerMap and show_layer_indices introduced in #131
• LRP keyword flatten and performance benefits
• LRP model canonization
• Update section "How it works internally" for #119
• Update "Model checks for humans" for #119
• Use DocumenterCitations.jl
• #64

This should return a Vector{Explanation}, which should also be supported by heatmap.

Automatic differentiation can be skipped for most known layers.
Add tests for a Dense wrapper to still test AD fallback.