问题描述
如果发现我只是忽略了某些内容,我首先道歉。我对Julia编程非常陌生,并且从一个星期前就开始让模型Zoo示例在GPU上运行。
在https://github.com/FluxML/model-zoo/blob/master/vision/mnist/conv.jl的启发下,我将https://github.com/FluxML/model-zoo/blob/master/text/lang-detection/model.jl更改为:
using Flux: onehot,onehotbatch,logitcrossentropy,reset!,throttle
using Statistics: mean
using Random
using Unicode
using Parameters: @with_kw
using CUDAapi
if has_cuda()
@info "CUDA is on"
import CUDA
#CUDA.allowscalar(false)
end
@with_kw mutable struct Args
lr::Float64 = 1e-3 # learning rate
N::Int = 300 # Number of perceptrons in hidden layer
test_len::Int = 100 # length of test data
langs_len::Int = 0 # Number of different languages in Corpora
alphabet_len::Int = 0 # Total number of characters possible,in corpora
throttle::Int = 5 # throttle timeout
end
function get_processed_data(args)
corpora = Dict()
for file in readdir("corpus")
lang = Symbol(match(r"(.*)\.txt",file).captures[1])
corpus = split(String(read("corpus/$file")),".")
corpus = strip.(Unicode.normalize.(corpus,casefold=true,stripmark=true))
corpus = filter(!isempty,corpus)
corpora[lang] = corpus
end
langs = collect(keys(corpora))
args.langs_len = length(langs)
alphabet = ['a':'z'; '0':'9'; ' '; '\n'; '_']
args.alphabet_len = length(alphabet)
# See which chars will be represented as "unkNown"
unique(filter(x -> x ∉ alphabet,join(vcat(values(corpora)...))))
dataset = [(onehotbatch(s,alphabet,'_'),onehot(l,langs)) for l in langs for s in corpora[l]] |> shuffle
train,test = dataset[1:end-args.test_len],dataset[end-args.test_len+1:end]
return train,test
end
function build_model(args)
scanner = Chain(Dense(args.alphabet_len,args.N,σ),LSTM(args.N,args.N))
encoder = Dense(args.N,args.langs_len)
return scanner,encoder
end
function model(x,scanner,encoder)
state = scanner.(x.data)[end]
reset!(scanner)
encoder(state)
end
function train(; kws...)
# Initialize Hyperparameters
args = Args(; kws...)
# Load Data
train_data,test_data = get_processed_data(args)
@info("Constructing Model...")
scanner,encoder = build_model(args)
# Load model and datasets onto GPU,if enabled
train_data = gpu.(train_data)
test_data = gpu.(test_data)
scanner = gpu(scanner)
encoder = gpu(encoder)
loss(x,y) = logitcrossentropy(model(x,encoder),y)
testloss() = mean(loss(t...) for t in test_data)
opt = Adam(args.lr)
ps = params(scanner,encoder)
evalcb = () -> @show testloss()
@info("Training...")
Flux.train!(loss,ps,train_data,opt,cb = throttle(evalcb,args.throttle))
end
cd(@__DIR__)
train()
结果是:
[ Info: CUDA is on
[ Info: Constructing Model...
[ Info: Training...
ERROR: LoadError: cuarray only supports bits types
Stacktrace:
[1] error(::String) at .\error.jl:33
[2] CUDA.cuarray{CUDA.cuarray{Float32,1},1}(::UndefInitializer,::Tuple{Int64}) at C:\Users\Fenmore\.julia\packages\CUDA\dZvbp\src\array.jl:115
[3] CUDA.cuarray{CUDA.cuarray{Float32,N} where N(::UndefInitializer,::Tuple{Int64}) at C:\Users\Fenmore\.julia\packages\CUDA\dZvbp\src\array.jl:124
[4] similar(::Type{CUDA.cuarray{CUDA.cuarray{Float32,N} where N},::Tuple{Int64}) at .\abstractarray.jl:675
[5] similar(::Type{CUDA.cuarray{CUDA.cuarray{Float32,::Tuple{Base.Oneto{Int64}}) at .\abstractarray.jl:674
[6] similar(::Base.broadcast.broadcasted{CUDA.cuarrayStyle{1},Tuple{Base.Oneto{Int64}},Zygote.var"#1177#1180"{Chain{Tuple{Dense{typeof(σ),CUDA.cuarray{Float32,2},1}},Flux.Recur{Flux.LSTMCell{CUDA.cuarray{Float32,1}}}}}},Tuple{Base.broadcast.Extruded{CUDA.cuarray{Flux.OneHotVector,Tuple{Bool},Tuple{Int64}}}},::Type{CUDA.cuarray{Float32,1}}) at C:\Users\Fenmore\.julia\packages\CUDA\dZvbp\src\broadcast.jl:11
[7] copy at .\broadcast.jl:877 [inlined]
[8] materialize at .\broadcast.jl:837 [inlined]
[9] broadcast_forward at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\lib\broadcast.jl:190 [inlined]
[10] adjoint at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\lib\broadcast.jl:202 [inlined]
[11] _pullback at C:\Users\Fenmore\.julia\packages\ZygoteRules\6nssF\src\adjoint.jl:47 [inlined]
[12] adjoint at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\lib\lib.jl:175 [inlined]
[13] _pullback at C:\Users\Fenmore\.julia\packages\ZygoteRules\6nssF\src\adjoint.jl:47 [inlined]
[14] broadcasted at .\broadcast.jl:1257 [inlined]
[15] model at D:\Users\Fenmore\git\Term-Project\module.jl:55 [inlined]
[16] _pullback(::Zygote.Context,::typeof(model),::Flux.OneHotMatrix{CUDA.cuarray{Flux.OneHotVector,::Chain{Tuple{Dense{typeof(σ),1}}}}},::Dense{typeof(identity),1}}) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\compiler\interface2.jl:0
[17] loss at D:\Users\Fenmore\git\Term-Project\module.jl:75 [inlined]
[18] _pullback(::Zygote.Context,::var"#loss#35",::Flux.OneHotVector) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\compiler\interface2.jl:0
[19] adjoint at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\lib\lib.jl:175 [inlined]
[20] _pullback at C:\Users\Fenmore\.julia\packages\ZygoteRules\6nssF\src\adjoint.jl:47 [inlined]
[21] #14 at C:\Users\Fenmore\.julia\packages\Flux\05b38\src\optimise\train.jl:83 [inlined]
[22] _pullback(::Zygote.Context,::Flux.Optimise.var"#14#20"{var"#loss#35",Tuple{Flux.OneHotMatrix{CUDA.cuarray{Flux.OneHotVector,Flux.OneHotVector}}) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\compiler\interface2.jl:0
[23] pullback(::Function,::Zygote.Params) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\compiler\interface.jl:172
[24] gradient(::Function,::Zygote.Params) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\compiler\interface.jl:53
[25] macro expansion at C:\Users\Fenmore\.julia\packages\Flux\05b38\src\optimise\train.jl:82 [inlined]
[26] macro expansion at C:\Users\Fenmore\.julia\packages\Juno\hEPx8\src\progress.jl:119 [inlined]
[27] train!(::Function,::Zygote.Params,::Array{Tuple{Flux.OneHotMatrix{CUDA.cuarray{Flux.OneHotVector,Flux.OneHotVector},::Adam; cb::Flux.var"#throttled#42"{Flux.var"#throttled#38#43"{Bool,Bool,var"#34#38"{var"#testloss#36"{var"#loss#35"}},Int64}}) at C:\Users\Fenmore\.julia\packages\Flux\05b38\src\optimise\train.jl:80
[28] train(; kws::Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}) at D:\Users\Fenmore\git\Term-Project\module.jl:82
[29] train() at D:\Users\Fenmore\git\Term-Project\module.jl:62
[30] top-level scope at D:\Users\Fenmore\git\Term-Project\module.jl:86
[31] include_string(::Function,::Module,::String,::String) at .\loading.jl:1088
in expression starting at D:\Users\Fenmore\git\Term-Project\module.jl:86
我尝试的另一个模型是https://github.com/FluxML/model-zoo/blob/master/text/char-rnn/char-rnn.jl,我将其更改为:
using Flux
using Flux: onehot,chunk,batchseq,throttle,logitcrossentropy
using StatsBase: wsample
using Base.Iterators: partition
using Parameters: @with_kw
# Hyperparameter arguments
@with_kw mutable struct Args
lr::Float64 = 1e-2 # Learning rate
seqlen::Int = 50 # Length of batchseqences
nbatch::Int = 50 # number of batches text is divided into
throttle::Int = 30 # Throttle timeout
end
function getdata(args)
# Download the data if not downloaded as 'input.txt'
isfile("input.txt") ||
download("https://cs.stanford.edu/people/karpathy/char-rnn/shakespeare_input.txt","input.txt")
text = collect(String(read("input.txt")))
# an array of all unique characters
alphabet = [unique(text)...,'_']
text = map(ch -> onehot(ch,alphabet),text)
stop = onehot('_',alphabet)
N = length(alphabet)
# Partitioning the data as sequence of batches,which are then collected as array of batches
Xs = collect(partition(batchseq(chunk(text,args.nbatch),stop),args.seqlen))
Ys = collect(partition(batchseq(chunk(text[2:end],args.seqlen))
return Xs,Ys,N,alphabet
end
# Function to construct model
function build_model(N)
return Chain(
LSTM(N,128),LSTM(128,Dense(128,N))
end
function train(; kws...)
# Initialize the parameters
args = Args(; kws...)
# Get Data
Xs,alphabet = getdata(args)
# Constructing Model
m = build_model(N)
Xs = gpu.(Xs)
Ys = gpu.(Ys)
m = gpu(m)
function loss(xs,ys)
l = sum(logitcrossentropy.(m.(xs),ys))
return l
end
## Training
opt = Adam(args.lr)
tx,ty = (Xs[5],Ys[5])
evalcb = () -> @show loss(tx,ty)
Flux.train!(loss,params(m),zip(Xs,Ys),args.throttle))
return m,alphabet
end
# Sampling
function sample(m,len; seed="")
m = cpu(m)
Flux.reset!(m)
buf = IOBuffer()
if seed == ""
seed = string(rand(alphabet))
end
write(buf,seed)
c = wsample(alphabet,softmax(m.(map(c -> onehot(c,collect(seed)))[end]))
for i = 1:len
write(buf,c)
c = wsample(alphabet,softmax(m(onehot(c,alphabet))))
end
return String(take!(buf))
end
cd(@__DIR__)
m,alphabet = train()
sample(m,1000) |> println
结果如下:
ERROR: LoadError: Mutating arrays is not supported
Stacktrace:
[1] error(::String) at .\error.jl:33
[2] (::Zygote.var"#455#456")(::nothing) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\lib\array.jl:68
[3] (::Zygote.var"#2384#back#457"{Zygote.var"#455#456"})(::nothing) at C:\Users\Fenmore\.julia\packages\ZygoteRules\6nssF\src\adjoint.jl:49
[4] cuarray at C:\Users\Fenmore\.julia\packages\CUDA\dZvbp\src\array.jl:206 [inlined]
[5] cuarray at C:\Users\Fenmore\.julia\packages\CUDA\dZvbp\src\array.jl:211 [inlined]
[6] (::typeof(∂(CUDA.cuarray{Float32,N} where N)))(::CUDA.cuarray{Float32,2}) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\compiler\interface2.jl:0
[7] LSTMCell at C:\Users\Fenmore\.julia\packages\Flux\05b38\src\cuda\curnn.jl:45 [inlined]
[8] (::typeof(∂(λ)))(::Tuple{nothing,2}}) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\compiler\interface2.jl:0
[9] (::Zygote.var"#177#178"{typeof(∂(λ)),Tuple{Tuple{nothing},Tuple{nothing}}})(::Tuple{nothing,2}}) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\lib\lib.jl:178
[10] (::Zygote.var"#1730#back#179"{Zygote.var"#177#178"{typeof(∂(λ)),Tuple{nothing}}}})(::Tuple{nothing,2}}) at C:\Users\Fenmore\.julia\packages\ZygoteRules\6nssF\src\adjoint.jl:49
[11] Recur at C:\Users\Fenmore\.julia\packages\Flux\05b38\src\layers\recurrent.jl:36 [inlined]
[12] (::typeof(∂(λ)))(::CUDA.cuarray{Float32,2}) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\compiler\interface2.jl:0
[13] applychain at C:\Users\Fenmore\.julia\packages\Flux\05b38\src\layers\basic.jl:36 [inlined]
[14] (::typeof(∂(applychain)))(::CUDA.cuarray{Float32,2}) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\compiler\interface2.jl:0
[15] Chain at C:\Users\Fenmore\.julia\packages\Flux\05b38\src\layers\basic.jl:38 [inlined]
[16] (::typeof(∂(λ)))(::CUDA.cuarray{Float32,2}) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\compiler\interface2.jl:0
[17] #1157 at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\lib\broadcast.jl:142 [inlined]
[18] (::Base.var"#3#4"{Zygote.var"#1157#1164"})(::Tuple{typeof(∂(λ)),2}}) at .\generator.jl:36
[19] iterate at .\generator.jl:47 [inlined]
[20] collect(::Base.Generator{Base.Iterators.Zip{Tuple{Array{typeof(∂(λ)),Array{CUDA.cuarray{Float32,1}}},Base.var"#3#4"{Zygote.var"#1157#1164"}}) at .\array.jl:686
[21] map at .\abstractarray.jl:2248 [inlined]
[22] (::Zygote.var"#1156#1163"{Tuple{Array{Flux.OneHotMatrix{CUDA.cuarray{Flux.OneHotVector,Val{2},Array{typeof(∂(λ)),1}})(::Array{CUDA.cuarray{Float32,1}) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\lib\broadcast.jl:142
[23] #3985#back at C:\Users\Fenmore\.julia\packages\ZygoteRules\6nssF\src\adjoint.jl:49 [inlined]
[24] (::Zygote.var"#177#178"{Zygote.var"#3985#back#1167"{Zygote.var"#1156#1163"{Tuple{Array{Flux.OneHotMatrix{CUDA.cuarray{Flux.OneHotVector,Tuple{Tuple{nothing,nothing,nothing},Tuple{}}})(::Array{CUDA.cuarray{Float32,1}) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\lib\lib.jl:178
[25] #1730#back at C:\Users\Fenmore\.julia\packages\ZygoteRules\6nssF\src\adjoint.jl:49 [inlined]
[26] broadcasted at .\broadcast.jl:1257 [inlined]
[27] (::typeof(∂(broadcasted)))(::Array{CUDA.cuarray{Float32,1}) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\compiler\interface2.jl:0
[28] loss at C:\Users\Fenmore\github\model-zoo\text\char-rnn\char-rnn.jl:60 [inlined]
[29] (::typeof(∂(λ)))(::Float32) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\compiler\interface2.jl:0
[30] #177 at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\lib\lib.jl:178 [inlined]
[31] #1730#back at C:\Users\Fenmore\.julia\packages\ZygoteRules\6nssF\src\adjoint.jl:49 [inlined]
[32] #14 at C:\Users\Fenmore\.julia\packages\Flux\05b38\src\optimise\train.jl:83 [inlined]
[33] (::typeof(∂(λ)))(::Float32) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\compiler\interface2.jl:0
[34] (::Zygote.var"#54#55"{Zygote.Params,Zygote.Context,typeof(∂(λ))})(::Float32) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\compiler\interface.jl:177
[35] gradient(::Function,::Zygote.Params) at C:\Users\Fenmore\.julia\packages\Zygote\rqvFi\src\compiler\interface.jl:54
[36] macro expansion at C:\Users\Fenmore\.julia\packages\Flux\05b38\src\optimise\train.jl:82 [inlined]
[37] macro expansion at C:\Users\Fenmore\.julia\packages\Juno\hEPx8\src\progress.jl:119 [inlined]
[38] train!(::Function,::Base.Iterators.Zip{Tuple{Array{Array{Flux.OneHotMatrix{CUDA.cuarray{Flux.OneHotVector,Array{Array{Flux.OneHotMatrix{CUDA.cuarray{Flux.OneHotVector,var"#25#27"{var"#loss#26",Array{Flux.OneHotMatrix{CUDA.cuarray{Flux.OneHotVector,Int64}}) at C:\Users\Fenmore\.julia\packages\Flux\05b38\src\optimise\train.jl:80
[39] train(; kws::Base.Iterators.Pairs{Union{},Tuple{}}}) at C:\Users\Fenmore\github\model-zoo\text\char-rnn\char-rnn.jl:69
[40] train() at C:\Users\Fenmore\github\model-zoo\text\char-rnn\char-rnn.jl:47
[41] top-level scope at C:\Users\Fenmore\github\model-zoo\text\char-rnn\char-rnn.jl:91
[42] include_string(::Function,::String) at .\loading.jl:1088
in expression starting at C:\Users\Fenmore\github\model-zoo\text\char-rnn\char-rnn.jl:91
我想我不能仅仅使用这些gpu()操作,但是我不知道从哪里开始弄清楚我还要做些什么。 感谢您的帮助。 提前致谢。 =)
解决方法
暂无找到可以解决该程序问题的有效方法,小编努力寻找整理中!
如果你已经找到好的解决方法,欢迎将解决方案带上本链接一起发送给小编。
小编邮箱:dio#foxmail.com (将#修改为@)