Flux: no method matching ndims(::ChainRulesCore.NoTangent)

Specific Domains Machine Learning
,
1

Consider the MWE below.

using Flux 

model = Chain(
    RNN(1, 128),
    Dense(128, 1)
);

# one feature, sequence length = 3, batch size = 6
x = [rand(Float32, 1, 6) for _ = 1:3]
y = rand(Float32, 1, 6)

function loss(x, y)
    Flux.reset!(model)
    # xi is the input at one step
    # many-to-one: we only care about the last output
    ŷ = [model(xi) for xi in x][end]
    return Flux.mse(ŷ, y)
end

# compute the gradient corresponding to (x, y)
ps = Flux.params(model)
gs = gradient(ps) do
    loss(x, y)
end

The above gradient line causes the following error:

MethodError: no method matching ndims(::ChainRulesCore.NoTangent)
Closest candidates are:
  ndims(::CUDA.CUSPARSE.CuSparseMatrix{Tv, Ti} where {Tv, Ti}) at /home/shuhua/.julia/packages/CUDA/YpW0k/lib/cusparse/array.jl:204
  ndims(::CUDA.CuSparseDeviceVector) at /home/shuhua/.julia/packages/CUDA/YpW0k/src/device/sparse.jl:23
  ndims(::AbstractChar) at char.jl:191
  ...
2

The weird behaviour disappears and everything works after I update Flux from v0.12.7 to v0.12.8. :joy:

1 Like
3

Not much changed in Flux for that point release, so most likely a change in Zygote or ChainRules(Core) that also updated when you ran ] up fixed things.

1 Like
4

Yes :smiley:. It is probably due to the update of ChainRulesCore: its recent releases are mostly concerned with the Tangent stuff.