Investigate fixing AD issues #151
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Definitely AutoZygote is required on the outside instead of AutoReverseDiff. Investigating what's going on with Enzyme in ternally
```julia
using PyCall
certifi = PyCall.pyimport("certifi")
ENV["SSL_CERT_FILE"] = certifi.where()
using ODINN
working_dir = joinpath(homedir(), "OGGM/ODINN_tests")
MB = true
fast = true
atol = 2.0
params = Parameters(OGGM = OGGMparameters(working_dir=working_dir,
multiprocessing=true),
simulation = SimulationParameters(working_dir=working_dir,
use_MB=MB,
velocities=true,
tspan=(2010.0, 2015.0),
multiprocessing=false,
workers=5,
test_mode=true),
hyper = Hyperparameters(batch_size=4,
epochs=4,
optimizer=ODINN.ADAM(0.01)),
UDE = UDEparameters(target = "A",
sensealg = ODINN.GaussAdjoint(autojacvec=ODINN.EnzymeVJP()))
)
rgi_ids = ["RGI60-11.03638"]
model = Model(iceflow = SIA2Dmodel(params),
mass_balance = mass_balance = TImodel1(params; DDF=6.0/1000.0, acc_factor=1.2/1000.0),
machine_learning = NN(params))
glaciers = initialize_glaciers(rgi_ids, params)
functional_inversion = FunctionalInversion(model, glaciers, params)
@time run!(functional_inversion)
```
This was referenced May 28, 2024
|
SciML/SciMLSensitivity.jl#1060 is required to use GaussAdjoint for this. |
|
Requires ODINN-SciML/Sleipnir.jl#55, ODINN-SciML/Huginn.jl#53, and Enzyme#main (which in turn requires SciML/OptimizationBase.jl#53). With that the callback differentiates, but it hits: ERROR: AssertionError: Base.Broadcast.Broadcasted{Base.Broadcast.DefaultArrayStyle{2}, Nothing, typeof(/), Tuple{Matrix{Float64}, Float64}} has mixed internal activity types. See https://enzyme.mit.edu/julia/stable/faq/#Mixed-activity for more informationwhich is an Enzyme bug in broadcast against a scalar. So we're close now, |
avik-pal
reviewed
May 29, 2024
| @@ -83,12 +82,12 @@ function build_D_features(H::Matrix, temp, ∇S) | |||
| ∇S_flat = ∇S[inn1(H) .!= 0.0] # flatten | |||
| H_flat = H[H .!= 0.0] # flatten | |||
| T_flat = repeat(temp,length(H_flat)) | |||
| X = Flux.normalise(hcat(H_flat,T_flat,∇S_flat))' # build feature matrix | |||
| X = Lux.normalise(hcat(H_flat,T_flat,∇S_flat))' # build feature matrix | |||
There was a problem hiding this comment.
Lux doesn't have a normalise function
avik-pal
reviewed
May 29, 2024
| - `θ`: Neural network parameters | ||
| """ | ||
| function NN(params::Sleipnir.Parameters; | ||
| architecture::Union{Flux.Chain, Nothing} = nothing, | ||
| θ::Union{Vector{F}, Nothing} = nothing) where {F <: AbstractFloat} | ||
| architecture::Union{Lux.Chain, Nothing} = nothing, |
There was a problem hiding this comment.
Does this need to be a Chain, won't AbstractExplicitLayer work?
There was a problem hiding this comment.
It will, I'm just making small changes to isolate the real AD issues though.
This was referenced May 29, 2024
|
Upstreaming issues found here:
The last issue is the real showstopper for now, it's not clear to me that there is a good workaround to that and it seems to me to just be a straight Enzyme bug to fix. The two look a bit bespoke to PyCall and TimerOutputs, the latter is easy to workaround, the former is weird because the workaround currently doesn't work seemingly due to some internals of PyCall. |
|
OK great, thanks for pushing this through @ChrisRackauckas! 👍🏻 Any idea how far is the solution to the last Enzyme issue? What is the issue with PyCall? As far as I understood, I thought we were bypassing the differentiation of all Python code for now which is present in the callback. |
|
As discussed in the thread, that issue really isn't an Enzyme issue in its own right, but rather that a type unstability exists around the broadcast which makes it illegal for certain code to be represented properly in Julia (if the type instability weren't there Enzyme can write the derivative update totally fine). recopying the backtrace here, since I think the better solution would just be fixing the type instability. |
|
@ChrisRackauckas @JordiBolibar where is the inactive function you have for pycall. Indeed I would've expected a We could also teach pycall object construction how to properly deal with allocations in an enzyme rule, but the fact you say there's an inactive marking that's not triggering seems like the bigger issue. |
ChrisRackauckas
commented
May 31, 2024
|
Another error I ran into was Tullio support. @wsmoses is that known? |
|
All tullio examples that folks have sent us are functional last we checked.
Open an issue?
…On Tue, Jun 4, 2024 at 7:26 PM Christopher Rackauckas < ***@***.***> wrote:
Another error I ran into was Tullio support. @wsmoses
<https://github.com/wsmoses> is that known?
—
Reply to this email directly, view it on GitHub
<#151 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAJTUXBVOCUQJRSFVLPUBLTZFX2F7AVCNFSM6AAAAABINIVDLSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCNBYGA2TCOJVHA>
.
You are receiving this because you were mentioned.Message ID:
***@***.***>
|
|
As I mentioned in the Huginn PR, Tullio should be easily avoidable here, since we only use it in the out-of-place version of the function that was needed for Zygote and ReverseDiff. Migrating to the in-place version of the function should avoid that. |
ChrisRackauckas
commented
Jun 16, 2024
Training iceflow UDE...
┌ Warning: Number of observations less than batch-size, decreasing the batch-size to 1
└ @ MLUtils ~/.julia/packages/MLUtils/LmmaQ/src/batchview.jl:95
┌ Warning: Number of observations less than batch-size, decreasing the batch-size to 1
└ @ MLUtils ~/.julia/packages/MLUtils/LmmaQ/src/batchview.jl:95
Before solving ODE problem
ODE problem solved for 1
over here for 1
simulation finished for 1
Batch 1 finished!
All batches finished
Loss computed: 13.506396213311314
┌ Warning: Automatic dt set the starting dt as NaN, causing instability. Exiting.
└ @ OrdinaryDiffEq ~/.julia/packages/OrdinaryDiffEq/F67Rp/src/solve.jl:591
┌ Warning: NaN dt detected. Likely a NaN value in the state, parameters, or derivative value caused this outcome.
└ @ SciMLBase ~/.julia/packages/SciMLBase/sakPO/src/integrator_interface.jl:593
ERROR: DimensionMismatch: array could not be broadcast to match destination
Stacktrace:
[1] check_broadcast_shape
@ ./broadcast.jl:579 [inlined]
[2] check_broadcast_axes
@ ./broadcast.jl:582 [inlined]
[3] instantiate
@ ./broadcast.jl:309 [inlined]
[4] materialize!
@ ./broadcast.jl:914 [inlined]
[5] materialize!
@ ./broadcast.jl:911 [inlined]
[6] vec_pjac!(out::ComponentArrays.ComponentVector{…}, λ::Vector{…}, y::Matrix{…}, t::Float64, S::SciMLSensitivity.AdjointSensitivityIntegrand{…})
@ SciMLSensitivity ~/.julia/dev/SciMLSensitivity/src/quadrature_adjoint.jl:288
[7] AdjointSensitivityIntegrand
@ ~/.julia/dev/SciMLSensitivity/src/quadrature_adjoint.jl:309 [inlined]
[8] (::SciMLSensitivity.AdjointSensitivityIntegrand{…})(t::Float64)
@ SciMLSensitivity ~/.julia/dev/SciMLSensitivity/src/quadrature_adjoint.jl:321
[9] evalrule(f::SciMLSensitivity.AdjointSensitivityIntegrand{…}, a::Float64, b::Float64, x::Vector{…}, w::Vector{…}, gw::Vector{…}, nrm::typeof(LinearAlgebra.norm))
@ QuadGK ~/.julia/packages/QuadGK/OtnWt/src/evalrule.jl:0
[10] #6
@ ~/.julia/packages/QuadGK/OtnWt/src/adapt.jl:15 [inlined]
[11] ntuple
@ ./ntuple.jl:48 [inlined]
[12] do_quadgk(f::SciMLSensitivity.AdjointSensitivityIntegrand{…}, s::Tuple{…}, n::Int64, atol::Float64, rtol::Float64, maxevals::Int64, nrm::typeof(LinearAlgebra.norm), segbuf::Nothing)
@ QuadGK ~/.julia/packages/QuadGK/OtnWt/src/adapt.jl:13
[13] #50
@ ~/.julia/packages/QuadGK/OtnWt/src/adapt.jl:253 [inlined]
[14] handle_infinities(workfunc::QuadGK.var"#50#51"{…}, f::SciMLSensitivity.AdjointSensitivityIntegrand{…}, s::Tuple{…})
@ QuadGK ~/.julia/packages/QuadGK/OtnWt/src/adapt.jl:145
[15] quadgk(::SciMLSensitivity.AdjointSensitivityIntegrand{…}, ::Float64, ::Vararg{…}; atol::Float64, rtol::Float64, maxevals::Int64, order::Int64, norm::Function, segbuf::Nothing)
@ QuadGK ~/.julia/packages/QuadGK/OtnWt/src/adapt.jl:252
[16] _adjoint_sensitivities(sol::SciMLBase.ODESolution{…}, sensealg::SciMLSensitivity.QuadratureAdjoint{…}, alg::OrdinaryDiffEq.RDPK3Sp35{…}; t::Vector{…}, dgdu_discrete::Function, dgdp_discrete::Nothing, dgdu_continuous::Nothing, dgdp_continuous::Nothing, g::Nothing, abstol::Float64, reltol::Float64, callback::SciMLBase.CallbackSet{…}, kwargs::@Kwargs{…})
@ SciMLSensitivity ~/.julia/dev/SciMLSensitivity/src/quadrature_adjoint.jl:382
[17] _adjoint_sensitivities
@ ~/.julia/dev/SciMLSensitivity/src/quadrature_adjoint.jl:324 [inlined]
[18] #adjoint_sensitivities#63
@ ~/.julia/dev/SciMLSensitivity/src/sensitivity_interface.jl:383 [inlined]
[19] (::SciMLSensitivity.var"#adjoint_sensitivity_backpass#314"{…})(Δ::SciMLBase.ODESolution{…})
@ SciMLSensitivity ~/.julia/dev/SciMLSensitivity/src/concrete_solve.jl:556
[20] ZBack
@ ~/.julia/packages/Zygote/nsBv0/src/compiler/chainrules.jl:211 [inlined]
[21] (::Zygote.var"#kw_zpullback#53"{…})(dy::SciMLBase.ODESolution{…})
@ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/chainrules.jl:237
[22] #291
@ ~/.julia/packages/Zygote/nsBv0/src/lib/lib.jl:206 [inlined]
[23] (::Zygote.var"#2169#back#293"{…})(Δ::SciMLBase.ODESolution{…})
@ Zygote ~/.julia/packages/ZygoteRules/M4xmc/src/adjoint.jl:72
[24] #solve#51
@ ~/.julia/packages/DiffEqBase/DS1sd/src/solve.jl:1003 [inlined]
[25] (::Zygote.Pullback{…})(Δ::SciMLBase.ODESolution{…})
@ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
[26] #291
@ ~/.julia/packages/Zygote/nsBv0/src/lib/lib.jl:206 [inlined]
[27] (::Zygote.var"#2169#back#293"{…})(Δ::SciMLBase.ODESolution{…})
@ Zygote ~/.julia/packages/ZygoteRules/M4xmc/src/adjoint.jl:72
[28] solve
@ ~/.julia/packages/DiffEqBase/DS1sd/src/solve.jl:993 [inlined]
[29] (::Zygote.Pullback{…})(Δ::SciMLBase.ODESolution{…})
@ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
[30] #simulate_iceflow_UDE!#32
@ ~/.julia/dev/ODINN/src/simulations/functional_inversions/functional_inversion_utils.jl:187 [inlined]
[31] (::Zygote.Pullback{Tuple{…}, Tuple{…}})(Δ::Nothing)
@ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
[32] simulate_iceflow_UDE!
@ ~/.julia/dev/ODINN/src/simulations/functional_inversions/functional_inversion_utils.jl:171 [inlined]
[33] (::Zygote.Pullback{Tuple{…}, Any})(Δ::Nothing)
@ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
[34] batch_iceflow_UDE
@ ~/.julia/dev/ODINN/src/simulations/functional_inversions/functional_inversion_utils.jl:146 [inlined]
[35] (::Zygote.Pullback{…})(Δ::@NamedTuple{…})
@ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
[36] #28
@ ~/.julia/dev/ODINN/src/simulations/functional_inversions/functional_inversion_utils.jl:114 [inlined]
[37] (::Zygote.Pullback{…})(Δ::@NamedTuple{…})
@ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
[38] #680
@ ~/.julia/packages/Zygote/nsBv0/src/lib/array.jl:201 [inlined]
[39] #235
@ ~/.julia/juliaup/julia-1.10.3+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/Distributed/src/pmap.jl:157 [inlined]
[40] (::Base.var"#1023#1028"{Distributed.var"#235#236"{…}})(r::Base.RefValue{Any}, args::Tuple{Tuple{…}})
@ Base ./asyncmap.jl:94
[41] (::Base.var"#1039#1040"{Base.var"#1023#1028"{Distributed.var"#235#236"{Zygote.var"#680#685"}}, Channel{Any}, Nothing})()
@ Base ./asyncmap.jl:228
Stacktrace:
[1] (::Base.var"#1033#1035")(x::Task)
@ Base ./asyncmap.jl:171
[2] foreach(f::Base.var"#1033#1035", itr::Vector{Any})
@ Base ./abstractarray.jl:3097
[3] maptwice(wrapped_f::Function, chnl::Channel{Any}, worker_tasks::Vector{Any}, c::Base.Iterators.Zip{Tuple{…}})
@ Base ./asyncmap.jl:171
[4] wrap_n_exec_twice
@ ./asyncmap.jl:147 [inlined]
[5] #async_usemap#1018
@ ./asyncmap.jl:97 [inlined]
[6] async_usemap
@ ./asyncmap.jl:78 [inlined]
[7] #asyncmap#1017
@ ./asyncmap.jl:75 [inlined]
[8] asyncmap
@ ./asyncmap.jl:74 [inlined]
[9] pmap(f::Function, p::Distributed.WorkerPool, c::Base.Iterators.Zip{…}; distributed::Bool, batch_size::Int64, on_error::Nothing, retry_delays::Vector{…}, retry_check::Nothing)
@ Distributed ~/.julia/juliaup/julia-1.10.3+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/Distributed/src/pmap.jl:126
[10] pmap(f::Function, p::Distributed.WorkerPool, c::Base.Iterators.Zip{Tuple{Vector{Tuple{…}}, Vector{@NamedTuple{…}}}})
@ Distributed ~/.julia/juliaup/julia-1.10.3+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/Distributed/src/pmap.jl:99
[11] pmap(f::Function, c::Base.Iterators.Zip{Tuple{Vector{Tuple{…}}, Vector{@NamedTuple{…}}}}; kwargs::@Kwargs{})
@ Distributed ~/.julia/juliaup/julia-1.10.3+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/Distributed/src/pmap.jl:156
[12] pmap
@ ~/.julia/juliaup/julia-1.10.3+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/Distributed/src/pmap.jl:156 [inlined]
[13] pmap(f::Function, c1::Vector{Tuple{…}}, c::Vector{@NamedTuple{…}})
@ Distributed ~/.julia/juliaup/julia-1.10.3+0.aarch64.apple.darwin14/share/julia/stdlib/v1.10/Distributed/src/pmap.jl:157
[14] (::Zygote.var"#map_back#682"{ODINN.var"#28#29"{…}, 1, Tuple{…}, Tuple{…}, Vector{…}})(Δ::Vector{@NamedTuple{…}})
@ Zygote ~/.julia/packages/Zygote/nsBv0/src/lib/array.jl:201
[15] (::Zygote.var"#2861#back#688"{Zygote.var"#map_back#682"{…}})(Δ::Vector{@NamedTuple{…}})
@ Zygote ~/.julia/packages/ZygoteRules/M4xmc/src/adjoint.jl:72
[16] predict_iceflow!
@ ~/.julia/dev/ODINN/src/simulations/functional_inversions/functional_inversion_utils.jl:114 [inlined]
[17] (::Zygote.Pullback{Tuple{…}, Tuple{…}})(Δ::Nothing)
@ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
[18] loss_iceflow
@ ~/.julia/dev/ODINN/src/simulations/functional_inversions/functional_inversion_utils.jl:58 [inlined]
[19] (::Zygote.Pullback{Tuple{…}, Any})(Δ::Float64)
@ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
[20] #25
@ ~/.julia/dev/ODINN/src/simulations/functional_inversions/functional_inversion_utils.jl:31 [inlined]
[21] (::Zygote.Pullback{Tuple{…}, Tuple{…}})(Δ::Float64)
@ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
[22] #291
@ ~/.julia/packages/Zygote/nsBv0/src/lib/lib.jl:206 [inlined]
[23] #2169#back
@ ~/.julia/packages/ZygoteRules/M4xmc/src/adjoint.jl:72 [inlined]
[24] OptimizationFunction
@ ~/.julia/packages/SciMLBase/sakPO/src/scimlfunctions.jl:3762 [inlined]
[25] (::Zygote.Pullback{Tuple{…}, Tuple{…}})(Δ::Float64)
@ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
[26] #291
@ ~/.julia/packages/Zygote/nsBv0/src/lib/lib.jl:206 [inlined]
[27] (::Zygote.var"#2169#back#293"{Zygote.var"#291#292"{Tuple{…}, Zygote.Pullback{…}}})(Δ::Float64)
@ Zygote ~/.julia/packages/ZygoteRules/M4xmc/src/adjoint.jl:72
[28] #37
@ ~/.julia/dev/OptimizationBase/ext/OptimizationZygoteExt.jl:90 [inlined]
[29] (::Zygote.Pullback{Tuple{…}, Tuple{…}})(Δ::Float64)
@ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
[30] #291
@ ~/.julia/packages/Zygote/nsBv0/src/lib/lib.jl:206 [inlined]
[31] #2169#back
@ ~/.julia/packages/ZygoteRules/M4xmc/src/adjoint.jl:72 [inlined]
[32] #39
@ ~/.julia/dev/OptimizationBase/ext/OptimizationZygoteExt.jl:93 [inlined]
[33] (::Zygote.Pullback{Tuple{…}, Tuple{…}})(Δ::Float64)
@ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface2.jl:0
[34] (::Zygote.var"#75#76"{Zygote.Pullback{Tuple{…}, Tuple{…}}})(Δ::Float64)
@ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface.jl:91
[35] gradient(f::Function, args::ComponentArrays.ComponentVector{Float64, Vector{Float64}, Tuple{ComponentArrays.Axis{…}}})
@ Zygote ~/.julia/packages/Zygote/nsBv0/src/compiler/interface.jl:148
[36] (::OptimizationZygoteExt.var"#38#56"{…})(::ComponentArrays.ComponentVector{…}, ::ComponentArrays.ComponentVector{…}, ::Vector{…}, ::Vararg{…})
@ OptimizationZygoteExt ~/.julia/dev/OptimizationBase/ext/OptimizationZygoteExt.jl:93
[37] macro expansion
@ ~/.julia/packages/OptimizationOptimisers/AOkbT/src/OptimizationOptimisers.jl:68 [inlined]
[38] macro expansion
@ ~/.julia/packages/ProgressLogging/6KXlp/src/ProgressLogging.jl:328 [inlined]
[39] (::OptimizationOptimisers.var"#12#13"{OptimizationBase.OptimizationCache{…}, ComponentArrays.ComponentVector{…}})()
@ OptimizationOptimisers ~/.julia/packages/Optimization/jWtfU/src/utils.jl:29
[40] maybe_with_logger(f::OptimizationOptimisers.var"#12#13"{…}, logger::Nothing)
@ Optimization ~/.julia/packages/Optimization/jWtfU/src/utils.jl:7
[41] macro expansion
@ ~/.julia/packages/Optimization/jWtfU/src/utils.jl:28 [inlined]
[42] __solve(cache::OptimizationBase.OptimizationCache{…})
@ OptimizationOptimisers ~/.julia/packages/OptimizationOptimisers/AOkbT/src/OptimizationOptimisers.jl:66
[43] solve!(cache::OptimizationBase.OptimizationCache{…})
@ SciMLBase ~/.julia/packages/SciMLBase/sakPO/src/solve.jl:188
[44] solve(prob::SciMLBase.OptimizationProblem{…}, alg::Optimisers.Adam, args::IterTools.NCycle{…}; kwargs::@Kwargs{…})
@ SciMLBase ~/.julia/packages/SciMLBase/sakPO/src/solve.jl:96
[45] train_UDE!(simulation::FunctionalInversion)
@ ODINN ~/.julia/dev/ODINN/src/simulations/functional_inversions/functional_inversion_utils.jl:43
[46] run!(simulation::FunctionalInversion)
@ ODINN ~/.julia/dev/ODINN/src/simulations/functional_inversions/functional_inversion_utils.jl:11
[47] macro expansion
@ ./timing.jl:279 [inlined]
[48] top-level scope
@ ~/Desktop/test.jl:76
Some type information was truncated. Use `show(err)` to see complete types.Hokay the autodiff is working using the SIA2D!, but in the adjoint the dt goes to NaN which crashes it and I'll need to investigate that. |
|
oh if the issue is a nan, enzyme (on cpu only presently) has a nan checker which will throw a backtrace the first time a nan is generated for a derivative. set Clearly this needs more docs [PRs welcome ofc!]. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Definitely AutoZygote is required on the outside instead of AutoReverseDiff. Investigating what's going on with Enzyme in ternally