Skip to content

IRTools dynamos + Zygote type issues #77

@femtomc

Description

@femtomc

This particular issue is really centered around Zygote, but I'm guessing that IRTools is part of the cause.

In my probabilistic programming framework, I have learnable parameters:

function learnable_hypers()
    l = learnable(:l, Float64[4.0, 4.0])
    m = learnable(:m, 10.0)
    q = rand(:q, Normal(l[1], 1.0 + exp(m)))
    return q
end

for which gradients can be computed. But I'm having drastically different behavior depending on type annotations for parameters with Array values.

In particular, the above works correctly with my pullbacks defined here. However, if I change the type annotation:

function learnable_hypers()
    l = learnable(:l, Any[4.0, 4.0])
    m = learnable(:m, 10.0)
    q = rand(:q, Normal(l[1], 1.0 + exp(m)))
    return q
end

I get an array mutation error:

ERROR: LoadError: Mutating arrays is not supported
Stacktrace:
 [1] error(::String) at ./error.jl:33
 [2] (::Zygote.var"#1048#1049")(::Nothing) at /home/mccoy/.julia/packages/Zygote/YeCEW/src/lib/array.jl:61
 [3] (::Zygote.var"#2775#back#1050"{Zygote.var"#1048#1049"})(::Nothing) at /home/mccoy/.julia/packages/ZygoteRules/6nssF/src/adjoint.jl:49
 [4] hvcat_fill at ./abstractarray.jl:1707 [inlined]
 [5] (::typeof(∂(λ)))(::Nothing) at /home/mccoy/.julia/packages/Zygote/YeCEW/src/compiler/interface2.jl:0
 [6] typed_hvcat at ./abstractarray.jl:1729 [inlined]
 [7] (::typeof(∂(λ)))(::Nothing) at /home/mccoy/.julia/packages/Zygote/YeCEW/src/compiler/interface2.jl:0
 [8] learnable_hypers at /home/mccoy/.julia/dev/Jaynes/scratch/learnable.jl:8 [inlined]
 [9] (::typeof(∂(λ)))(::Float64) at /home/mccoy/.julia/packages/Zygote/YeCEW/src/compiler/interface2.jl:0
 [10] #174 at /home/mccoy/.julia/packages/Zygote/YeCEW/src/lib/lib.jl:182 [inlined]
 [11] #347#back at /home/mccoy/.julia/packages/ZygoteRules/6nssF/src/adjoint.jl:49 [inlined]
 [12] #84 at /home/mccoy/.julia/dev/Jaynes/src/contexts/backpropagate.jl:152 [inlined]
 [13] (::typeof(∂(λ)))(::Tuple{Float64,Float64}) at /home/mccoy/.julia/packages/Zygote/YeCEW/src/compiler/interface2.jl:0
 [14] (::Zygote.var"#36#37"{typeof(∂(λ))})(::Tuple{Float64,Float64}) at /home/mccoy/.julia/packages/Zygote/YeCEW/src/compiler/interface.jl:46
 [15] accumulate_parameter_gradients!(::Main.Learnable.Jaynes.Gradients, ::Main.Learnable.Jaynes.BlackBoxCallSite{Main.Learnable.Jaynes.HierarchicalTrace,Tuple{},Float64}, ::Float64, ::Float64) at /home/mccoy/.julia/dev/Jaynes/src/contexts/backpropagate.jl:157
 [16] get_parameter_gradients at /home/mccoy/.julia/dev/Jaynes/src/contexts/backpropagate.jl:204 [inlined]
 [17] get_parameter_gradients(::Main.Learnable.Jaynes.BlackBoxCallSite{Main.Learnable.Jaynes.HierarchicalTrace,Tuple{},Float64}, ::Float64) at /home/mccoy/.julia/dev/Jaynes/src/contexts/backpropagate.jl:203

Wierdly enough, I run into the same issue if I type annotate multi-dimensional arrays

function learnable_hypers()
    l = learnable(:l, Float64[4.0 4.0; 4.0 4.0])
    m = learnable(:m, 10.0)
    q = rand(:q, Normal(l[1], 1.0 + exp(m)))
    return q
end

So 1-dimensional Float64 arrays seem to work, but everything else will not go. Path forward might be to analyze what's going on with Cthulhu.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions