ablaom
commented
11 months ago Failing on GPU only. The complaint is about scalar indexing.
I've spent some time on this today but this is hard for me to debug because I don't currently have GPU access. I conjecture that the following code is failing on a GPU but not a CPU, and this contains the issue. Be good if someone can confirm this indeed fails. And if so, where is the scalar indexing?
import Flux
import MLJFlux
import StableRNGs.StableRNG
rng = StableRNG(123)
X, y = MLJFlux.make_images(rng);
typeof(X)
# Vector{Matrix{Gray{Float64}}}
data = MLJFlux.collate(ImageClassifier(), X, y);
Flux.gpu(data) # no effect on my CPU-only machine
typeof(data)
# Tuple{Vector{Array{Float32, 4}}, Vector{OneHotArrays.OneHotMatrix{UInt32, Vector{UInt32}}}}
n_channels = 1
n_classes = 3
init = Flux.glorot_uniform(rng)
chain = Flux.Chain(
Flux.Conv((2, 2), n_channels=>2, init=init),
Flux.Conv((2, 2), 2=>1, init=init),
x->reshape(x, :, size(x)[end]),
Flux.Dense(16, n_classes, init=init))
x = data[1][1]
typeof(x)
# Array{Float32, 4}
sizeof(x)
# (6, 6, 1, 1)
chain(x)
mohamed82008
ToucheSir
codecov-commenter
This PR bumps the [compat] for Metalhead to "0.8" and addresses resulting breakages.
Replaces #226