1 
%%julia


# Load the data
println("start")
x_train, y_train,x_valid, y_valid=h5open("./data.h5", "r") do file
    read(file, "tr_X","tr_Y","va_X","va_Y")
    end|> gpu
println("read done")
x_train = reshape(x_train, (80,80,3,:))|> gpu
x_valid = reshape(x_valid, (80,80,3,:))|> gpu
println("reshape done")
# Add the channel layer
#x_train = Flux.unsqueeze(x_train, 3)
#x_valid = Flux.unsqueeze(x_valid, 3)
# Encode labels
#y_train = onehotbatch(y_train, 0:1)
#y_valid = onehotbatch(y_valid, 0:1)
# Create the full dataset
train_data = DataLoader(x_train, y_train, batchsize=32)|> gpu
println("generator done")
model = Chain(
    
    Conv((3, 3), 3=>32, stride=1, relu),    

    Conv((3, 3), 32=>64, stride=1, relu),
    MaxPool((2,2); pad = 0, stride = 1),
    Dropout(0.25),

    Conv((3, 3), 64=>64, stride=1, relu),
    MaxPool((2,2); pad = 0, stride = 1),
    Dropout(0.25),
    
    Conv((3, 3), 64=>128, pad=1, stride=1, relu),
    MaxPool((2,2); pad = 0, stride = 1),
    Dropout(0.5), 
    GlobalMeanPool(),
    flatten,
    Dense(128, 64,relu),
    Dropout(0.5),
    Dense(64, 1,sigmoid)
)|> gpu
println("chain init done")
#ŷ = model(x_valid)
#println("Prediction of first image: $(ŷ[1])")
accuracy(ŷ, y) = mean(ŷ.== y) |> gpu
loss(x, y) = Flux.crossentropy(model(x), y) |> gpu
# learning rate
lr = 0.1|> gpu
opt = Descent(lr) |> gpu 
ps = Flux.params(model) |> gpu
number_epochs = 10|> gpu
println("train start")
callback() = push!(loss_vector, loss(x_train, y_train))|> gpu
loss_vector = Vector{Float64}()|> gpu
@epochs number_epochs Flux.train!(loss, ps, train_data, opt,cb=callback) |> gpu
println("post epoch")
#accuracy(model(x_train), y_train) |> gpu
println("done!")

以上是我第一个使用flux和julia的深度学习项目。尝试使用 GPU 时出现以下错误。

RuntimeError: <PyCall.jlwrap (在从 Python 调用的 Julia 函数中) JULIA: InvalidIRError: 编译内核 getindex_kernel(CUDA.CuKernelContext, CuDeviceArray{Float32,4,1}, CuDeviceArray{Float32,4,1}, NTuple{4,Int64 }, Base.Slice{Base.OneTo{Int64}}, Base.Slice{Base.OneTo{Int64}}, Base.Slice{Base.OneTo{Int64}}, CuDeviceArray{Int64,1,1}) 导致无效LLVM IR 原因：不支持的动态函数调用（调用 pointerref(ptr::Core.LLVMPtr{T,A}, i::Int64, ::Val{align}) where {T, A, align} in LLVM.Interop at /root/.julia/packages/LLVM/F9DFY/src/interop/pointer.jl:7)

任何帮助都会很棒。我正在为这个项目使用 Kaggle 环境。

4

0 回答 0