mcabbott
commented
2 years ago FWIW the error is slightly different on master:
```julia
julia> import Yota
julia> normal_pdf(x, mean, var) = exp(-(x - mean)^2 / (2var)) / sqrt(2π * var);
julia> Yota.grad((x, mu) -> sum(log, normal_pdf.(x, mu, 1.0)), rand(10), 1.0)
┌ Error: Failed to compile rrule for broadcasted(normal_pdf, [0.8031553730805592, 0.3509560552123825, 0.032551822966513155, 0.21170603638555385, 0.2049628078398853, 0.8469464153815023, 0.4220217037413583, 0.286621858419426, 0.0338940405059448, 0.43421951685956195], 1.0, 1.0), extract details via:
│ (f, args) = Yota.RRULE_VIA_AD_STATE[]
└ @ Yota ~/.julia/packages/Yota/QGPcM/src/cr_api.jl:179
ERROR: MethodError: no method matching iterate(::Type{Val})
Closest candidates are:
iterate(::Union{LinRange, StepRangeLen})
@ Base range.jl:869
iterate(::Union{LinRange, StepRangeLen}, ::Integer)
@ Base range.jl:869
iterate(::T) where T<:Union{Base.KeySet{<:Any, <:Dict}, Base.ValueIterator{<:Dict}}
@ Base dict.jl:698
...
Stacktrace:
[1] first(itr::Type)
@ Base ./abstractarray.jl:436
[2] map(f::typeof(first), t::Tuple{UnionAll, Int64})
@ Base ./tuple.jl:274
[3] trace_call!(::Umlaut.Tracer{Yota.BcastGradCtx}, ::Function, ::Vararg{Any})
@ Yota ~/.julia/packages/Yota/QGPcM/src/grad.jl:92
[4] trace_block!(t::Umlaut.Tracer{Yota.BcastGradCtx}, ir::Core.Compiler.IRCode, bi::Int64, prev_bi::Int64, sparams::Core.SimpleVector)
@ Umlaut ~/.julia/packages/Umlaut/LH23t/src/trace.jl:312
[5] trace!(t::Umlaut.Tracer{Yota.BcastGradCtx}, v_fargs::Vector{Umlaut.Variable})
@ Umlaut ~/.julia/packages/Umlaut/LH23t/src/trace.jl:436
[6] trace(::Function, ::Vector{Float64}, ::Vararg{Any}; ctx::Yota.BcastGradCtx, deprecated_kws::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
@ Umlaut ~/.julia/packages/Umlaut/LH23t/src/trace.jl:546
[7] make_rrule(::typeof(Base.Broadcast.broadcasted), ::Function, ::Vector{Float64}, ::Vararg{Any})
@ Yota ~/.julia/packages/Yota/QGPcM/src/cr_api.jl:136
[8] rrule_via_ad(::Yota.YotaRuleConfig, ::Function, ::Function, ::Vararg{Any})
@ Yota ~/.julia/packages/Yota/QGPcM/src/cr_api.jl:172
[9] rrule(::Yota.YotaRuleConfig, ::typeof(Base.Broadcast.broadcasted), ::typeof(normal_pdf), ::Vector{Float64}, ::Float64, ::Float64)
@ Yota ~/.julia/packages/Yota/QGPcM/src/rulesets.jl:91
[10] mkcall(::Function, ::Yota.YotaRuleConfig, ::Vararg{Any}; val::Missing, line::Core.LineInfoNode, kwargs::NamedTuple{(), Tuple{}}, free_kwargs::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
@ Umlaut ~/.julia/packages/Umlaut/LH23t/src/tape.jl:202
[11] record_primitive!(::Umlaut.Tape{Yota.GradCtx}, ::Function, ::Vararg{Any})
@ Yota ~/.julia/packages/Yota/QGPcM/src/grad.jl:54
[12] trace_call!(::Umlaut.Tracer{Yota.GradCtx}, ::Function, ::Vararg{Any})
@ Umlaut ~/.julia/packages/Umlaut/LH23t/src/trace.jl:286
[13] trace_block!(t::Umlaut.Tracer{Yota.GradCtx}, ir::Core.Compiler.IRCode, bi::Int64, prev_bi::Int64, sparams::Core.SimpleVector)
@ Umlaut ~/.julia/packages/Umlaut/LH23t/src/trace.jl:312
[14] trace!(t::Umlaut.Tracer{Yota.GradCtx}, v_fargs::Vector{Umlaut.Variable})
@ Umlaut ~/.julia/packages/Umlaut/LH23t/src/trace.jl:436
[15] trace(::Function, ::Vector{Float64}, ::Vararg{Any}; ctx::Yota.GradCtx, deprecated_kws::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
@ Umlaut ~/.julia/packages/Umlaut/LH23t/src/trace.jl:546
[16] gradtape(::var"#3#4", ::Vector{Float64}, ::Vararg{Any}; ctx::Yota.GradCtx, seed::Int64)
@ Yota ~/.julia/packages/Yota/QGPcM/src/grad.jl:258
[17] grad(::var"#3#4", ::Vector{Float64}, ::Vararg{Any}; seed::Int64)
@ Yota ~/.julia/packages/Yota/QGPcM/src/grad.jl:334
[18] grad(::var"#3#4", ::Vector{Float64}, ::Vararg{Any})
@ Yota ~/.julia/packages/Yota/QGPcM/src/grad.jl:326
[19] top-level scope
@ REPL[5]:1
(jl_atLVz9) pkg> st
Status `/private/var/folders/yq/4p2zwd614y59gszh7y9ypyhh0000gn/T/jl_atLVz9/Project.toml`
[92992a2b] Umlaut v0.4.2 `https://github.com/dfdx/Umlaut.jl.git#main`
[cd998857] Yota v0.7.5 `https://github.com/dfdx/Yota.jl.git#main`
```
If I disable all rules related to broadcasting in Yota, then the error is more straightforward, and I think tells us that the rules from https://github.com/JuliaDiff/ChainRules.jl/pull/644 (which ought to cover this kind of broadcasting) are not being called:
julia> using Yota
julia> normal_pdf(x, mean, var) = exp(-(x - mean)^2 / (2var)) / sqrt(2π * var);
julia> Yota.grad((x, mu) -> sum(log, normal_pdf.(x, mu, 1.0)), rand(10), 1.0)
ERROR: No deriative rule found for op %5 = materialize(%4)::Vector{Float64} , try defining it using
ChainRulesCore.rrule(::typeof(Base.Broadcast.materialize), ::Base.Broadcast.Broadcasted{Base.Broadcast.DefaultArrayStyle{1}, Nothing, typeof(normal_pdf), Tuple{Vector{Float64}, Float64, Float64}}) = ...
Stacktrace:
[1] error(s::String)
@ Base ./error.jl:35
[2] step_back!(tape::Umlaut.Tape{Yota.GradCtx}, y::Umlaut.Variable)
@ Yota ~/.julia/dev/Yota/src/grad.jl:185
...
# same error on a much easier broadcast, should use derivatives_given_output
julia> Yota.grad((x, mu) -> sum(log, atan.(x, mu)), rand(10), 1.0)
ERROR: No deriative rule found for op %5 = materialize(%4)::Vector{Float64} , try defining it using ...
# and an even easier one, has its own rrule(broadcasted, +, ...)
julia> Yota.grad((x, mu) -> sum(log, x .+ mu), rand(10), 1.0)
ERROR: No deriative rule found for op %7 = materialize(%5)::Vector{Float64} , try defining it using ...
(jl_fINTq0) pkg> st
Status `/private/var/folders/yq/4p2zwd614y59gszh7y9ypyhh0000gn/T/jl_fINTq0/Project.toml`
[082447d4] ChainRules v1.43.2
[92992a2b] Umlaut v0.4.2 `https://github.com/dfdx/Umlaut.jl.git#main`
[cd998857] Yota v0.7.5 `~/.julia/dev/Yota`
dfdx
Code
Error message
Versions