rmsrosa / chainplots.jl Goto Github PK

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!

Currently, the method neuron_connections(m::Flux.Chain, input_data::Union{Nothing,Array} = nothing) traverses the neural network layers and builds a Vector{Dict{Tuple, Vector{Tuple}}} with all the connections between consecutive layers. In a second step, chaingraph takes this connections and builds a metagraph, which internally is stored as Vector{Vector{Tuple}} (joining every key-value pair into a vector. This is orders of magnitude more efficient.

So the idea is to either build a Vector{Vector{Tuple}} instead of Vector{Dict{Tuple, Vector{Tuple}}}, as an intermediate step, or build the graph directly, fusing the two methods.

I couldn't find a better way to do this, but: I took examples.jl and turned it into a Pluto notebook. Enjoy!
ChainPlotsExamples.zip

By the way: awesome package! Thank you!

Many thanks for the nice package!

I am encountering the following issue when trying to plot a NN with a custom layer. While the following code works as expected

custom(x) = [x[1]^2, x[2]^2, x[3]]
nnx = Chain(custom, Dense(3,4), Dense(4,2))
plot(nnx, rand(3), title="$nna")

the following small modification of it does not:

custom(x) = [x[1]^2, x[2]^2, sin(x[3])]
nnx = Chain(custom, Dense(3,4), Dense(4,2))
plot(nnx, rand(3), title="$nna")

The latter code fails with error

ERROR: MethodError: no method matching sin(::ChainPlots.NeuralNumbers.NeuralNumber)

I tried to extend the function sin to NeuralNumbers, but the issue persisted, so any help would be much appreciated.

EDIT: Actually, the following line

Base.sin(x::ChainPlots.NeuralNumbers.NeuralNumber) = x

fixes the issue. It would be nice to have native support for sin and cos, though.

It would be nice to have Plots as an optional dependency instead

1. This is a nice package. Will you be registering it with the package manager soon?
2. Going with the trend "Plots" and "StatsPlots", adding an "s" at the end ("ChainPlots") might be nice.
3. There is some issue using your package with CUDA and Flux:

using CUDA, Flux, ChainPlot
arr = [1,2,3] |> cpu

leads to:

MethodError: adapt_storage(::Flux.FluxCPUAdaptor, ::Vector{Int64}) is ambiguous. Candidates:
  adapt_storage(NeuronState, xs::AbstractArray{<:Number}) in ChainPlot.NeuronNumbers at C:\Users\matth\.julia\packages\ChainPlot\XVMxo\src\NeuronNumbers.jl:137
  adapt_storage(to::Flux.FluxCPUAdaptor, x::AbstractArray) in Flux at C:\Users\matth\.julia\packages\Flux\qAdFM\src\functor.jl:106
Possible fix, define
  adapt_storage(::Flux.FluxCPUAdaptor, ::AbstractArray{<:Number})

The same issue with GPU, e.g., [1,2,3] |> gpu...

Connections between neurons are now inferred using the submodule NeuronNumbers, which contains coldneurons and hotneurons. Problem is that relu(hotneuron) should return hotneuron but is returning coldneuron.

Hi,

If I have custom layers, that are based on Flux layers, can this be used or do I need to chain everything?

https://github.com/rmsrosa/ChainPlot.jl#examples
with julia 1.7.1

julia> plot(nnr, title="With theme solarized_light", titlefontsize=10)
ERROR: MethodError: no method matching add_edge!(::MetaGraphs.MetaGraph{Int64, Float64}, ::Int64, ::Int64)
Closest candidates are:
  add_edge!(::LightGraphs.SimpleGraphs.AbstractSimpleGraph, ::Any, ::Any) at ~/.julia/packages/LightGraphs/IgJif/src/SimpleGraphs/SimpleGraphs.jl:90
  add_edge!(::LightGraphs.SimpleGraphs.AbstractSimpleGraph, ::Any) at ~/.julia/packages/LightGraphs/IgJif/src/SimpleGraphs/SimpleGraphs.jl:86
Stacktrace:
 [1] chaingraph(m::Chain{Tuple{Dense{typeof(σ), Matrix{Float32}, Vector{Float32}}, Flux.Recur{Flux.RNNCell{typeof(relu), Matrix{Float32}, Vector{Float32}, Matrix{Float32}}, Matrix{Float32}}, Flux.Recur{Flux.LSTMCell{Matrix{Float32}, Vector{Float32}, Tuple{Matrix{Float32}, Matrix{Float32}}}, Tuple{Matrix{Float32}, Matrix{Float32}}}, Flux.Recur{Flux.GRUCell{Matrix{Float32}, Vector{Float32}, Matrix{Float32}}, Matrix{Float32}}, Dense{typeof(identity), Matrix{Float32}, Vector{Float32}}}}, input_data::Nothing)
   @ ChainPlot ~/.julia/packages/ChainPlot/Ua2jN/src/chaingraph.jl:77
 [2] macro expansion
   @ ~/.julia/packages/ChainPlot/Ua2jN/src/chainplotrecipe.jl:56 [inlined]
 [3] apply_recipe(plotattributes::AbstractDict{Symbol, Any}, m::Chain, input_data::Union{Nothing, Array})
   @ ChainPlot ~/.julia/packages/RecipesBase/qpxEX/src/RecipesBase.jl:289
 [4] apply_recipe(plotattributes::AbstractDict{Symbol, Any}, m::Chain)
   @ ChainPlot ~/.julia/packages/RecipesBase/qpxEX/src/RecipesBase.jl:283
 [5] _process_userrecipes!(plt::Any, plotattributes::Any, args::Any)
   @ RecipesPipeline ~/.julia/packages/RecipesPipeline/oVorB/src/user_recipe.jl:36
 [6] recipe_pipeline!(plt::Any, plotattributes::Any, args::Any)
   @ RecipesPipeline ~/.julia/packages/RecipesPipeline/oVorB/src/RecipesPipeline.jl:70
 [7] _plot!(plt::Plots.Plot, plotattributes::Any, args::Any)
   @ Plots ~/.julia/packages/Plots/cPJQu/src/plot.jl:208
 [8] #plot#135
   @ ~/.julia/packages/Plots/cPJQu/src/plot.jl:91 [inlined]
 [9] top-level scope

I have defined a new layer which contains a SparseMatrixCSC. I would like to plot this layer to show how if differs from Dense.

Is this possible already? It seems not, but I may have missed something.

If not, what functions must be defined to plot such a layer?

Currently, in convolutional layers, all neurons show up as connected to every neuron in the previous layer.