Unable to retrieve saved machines #743
Comments
|
Mmm, that's frustrating. Thanks for reporting. This sounds to me like a serialization issue, rather than something specific to MLJ. So you may get more help by posting an issue at JLSO.jl or on Julia discourse. cc @oxinabox . I can tell you that:
using MLJ
X, y = @load_iris;
model = (@load DecisionTreeClassifier)()
mach = machine(model, X, y) |> fit!
MLJ.save("junk.jlso", mach)
using JLSO
jlso = JLSO.read("junk.jlso", JLSOFile);
jlso.julia
jlso.manifest
So if this works for your file, you should be able to reproduce your environment to the state at save time. |
Should load up an environment with exactly the dependencies you had loaded before. There is also an open issue, and an inlcuded hacky method and a PR for a less hacky method to also load the packages. |
casasgomezuribarri
changed the title
|
Thank you both for your attention, but none of the things suggested seems to work. Perhaps sharing one of the .jlso files I am unable to retrieve helps? GitHub doesn't support files of this type, so I got a Wetransfer link: And here is the code I am using to try to retrieve my saved machines (the code is a bit longer because I am stacking models and the learning network has to be defined and exported first), which works with a newly saved machine but not the file above. using MLJ
using EvoTrees
using DataFrames
# Define learning network
begin
# source nodes
X = source()
y = source()
# node 1
logmodel = (@load LogisticClassifier pkg=MLJLinearModels)()
log = machine(logmodel, X, y)
LOG = predict(log, X)
# node 2
evomodel = (@load EvoTreeClassifier)()
evo = machine(evomodel, X, y)
EVO = predict(evo, X)
# define new model type for node 3
mutable struct Compiler <: Static
end
# and its transform method
import MLJBase
function MLJBase.transform(c::Compiler,_,y1, y2)
prob_y1 = pdf.(y1, 1)
prob_y2 = pdf.(y2, 1)
X_stack = DataFrame(:Prob_1_LOG => prob_y1, :Prob_1_EVO => prob_y2)
return X_stack
end
# node 3
compiler = Compiler()
comp= machine(compiler)
X_stack = MLJ.transform(comp, LOG, EVO)
# node 4
metamodel = (@load LogisticClassifier pkg=MLJLinearModels)()
meta = machine(metamodel, X_stack, y)
META = predict(meta, X_stack)
end
# Export as a new model type
mach = machine(Probabilistic(), X, y; predict = META)
@from_network mach begin
mutable struct CompositeModel
model1 = logmodel
model2 = evomodel
compiler = compiler
judge = metamodel
end
end
# Fit and save a machine
begin
X = rand(100, 10)
y = coerce(rand([0, 1], 100), OrderedFactor)
mymodel = CompositeModel()
mymach = machine(mymodel, X_train, y_train) |> fit!
MLJ.save("junk.jlso", mymach)
end
# Retrieve new and old machines
success = machine("junk.jlso", X_train, y_train)
fail = machine("model1_trained_Run1.jlso", X_train, y_train)
# Trying what was suggested here
using JLSO
jlso_success = JLSO.read("junk.jlso", JLSOFile)
jlso_success.julia
jlso_success.manifest
jlso_fail = JLSO.read("model1_trained_Run1.jlso", JLSOFile)
jlso_fail.julia
jlso_fail.manifest
#note that the number of packages in each manifest is different
using Pkg
Pkg.activate(jlso_fail)
still_fails = machine("model1_trained_Run1.jlso", X, y)
Pkg.activate(jlso_success)
still_fails = machine("model1_trained_Run1.jlso", X, y)
# Trying the trick from the other issue
eval(Expr(:using, (Expr(:., n) for n in Symbol.(keys(Pkg.Types.Context().env.project.deps)))...))
still_fails = machine("model1_trained_Run1.jlso", X, y)Any ideas? |
|
Additional Context More information at julia-vscode/julia-vscode#1967 |
|
Did you run |
|
Sorry, I'm not sure of what you mean |
|
I went to a plain REPL and after doing Pkg.add JLSO
using JLSO
using PkgThen I defined the learning network, exported it and created the variables X and y with the code I shared in my previous post. To do this, I needed to jlso_fail = JLSO.read("model1_trained_Run1.jlso", JLSOFile) #copied file into pwd()
Pkg.activate(jlso_fail)
still_fails = machine("model1_trained_Run1.jlso", X, y)I got the following error, a bit different this time: [warn | JLSO]: KeyError: key MLJGLMInterface [caf8df21-4939-456d-ac9c-5fefbfb04c0c] not found
[warn | JLSO]: KeyError: key MLJGLMInterface [caf8df21-4939-456d-ac9c-5fefbfb04c0c] not found
[warn | JLSO]: KeyError: key MLJGLMInterface [caf8df21-4939-456d-ac9c-5fefbfb04c0c] not found
ERROR: MethodError: no method matching machine(::Array{UInt8,1}, ::Array{Float64,2}, ::CategoricalArrays.CategoricalArray{Int64,1,UInt32,Int64,CategoricalArrays.CategoricalValue{Int64,UInt32},Union{}})
Closest candidates are:
machine(::Type{#s41} where #s41<:MLJModelInterface.Model, ::Any...) at C:\Users\ivica\.julia\packages\MLJBase\fYEbE\src\machines.jl:203
machine(::Static, ::Any...) at C:\Users\ivica\.julia\packages\MLJBase\fYEbE\src\machines.jl:213
machine(::MLJModelInterface.Model, ::Any, ::Any...) at C:\Users\ivica\.julia\packages\MLJBase\fYEbE\src\machines.jl:231
...
Stacktrace:
[1] machine(::String, ::Array{Float64,2}, ::Vararg{Any,N} where N; kwargs::Base.Iterators.Pairs{Union{},Union{},Tuple{},NamedTuple{(),Tuple{}}}) at C:\Users\ivica\.julia\packages\MLJBase\fYEbE\src\machines.jl:702
[2] machine(::String, ::Array{Float64,2}, ::CategoricalArrays.CategoricalArray{Int64,1,UInt32,Int64,CategoricalArrays.CategoricalValue{Int64,UInt32},Union{}}) at C:\Users\ivica\.julia\packages\MLJBase\fYEbE\src\machines.jl:694
[3] top-level scope at REPL[26]:1I did [warn | JLSO]: KeyError: key DataFrames [a93c6f00-e57d-5684-b7b6-d8193f3e46c0] not found
Machine{ProbabilisticTunedModel{RandomSearch,…}} @602 trained 1 time.
args:
1: Source @074 ⏎ `AbstractArray{Continuous,2}`
2: Source @801 ⏎ `AbstractArray{OrderedFactor{2},1}`Which after doing Machine{ProbabilisticTunedModel{RandomSearch,…}} @171 trained 1 time.
args:
1: Source @045 ⏎ `AbstractArray{Continuous,2}`
2: Source @500 ⏎ `AbstractArray{OrderedFactor{2},1}`It looked like it had worked in the REPL, so I went to the directory where I have all my files and moved Project.toml to a different location to have an empty |
Describe the bug
I am unable to retrieve a saved trained machine.
I have a few machines that I have already tuned stored on my computer. I have been retrieving them using
machine("mymach_1.jlso", X, y)to make plots and inspect well each tuning process. However, today I haven't been able to retrieve any of them, I get the following error:Click to see error
I don't understand most of what it says there, but I saw the line
[warn | JLSO]: KeyError: key MLJGLMInterface [caf8df21-4939-456d-ac9c-5fefbfb04c0c] not foundso I tried] add MLJGLMInterfaceandusing MLJGLMInterface. However, I still got a very similar error:Click to see error
*note that the abovementioned line now says
[warn | JLSO]: EOFError: read end of fileinstead.I am not able to provide a reproducible code because newly saved machines seem to be correctly retrieved. I only have a problem accessing those that were tuned and saved in the past.
Additional context
After I tuned my last machine (the last one of those I am not able to retrieve), I wanted to add the package
StatsPlotsto make a violin plot. However, I addedStatPlotsinstead by mistake, which downgraded several packages in my environment. I rapidly did] rm StatPlots,] add StatsPlotsand] update, which caused several changes in myProject.tomlfile. At that point I wasn't having any problems and I continued working on my project. It was only the next day that I realised that I was unable to retrieve any of the machines I had saved up to that point, although newly trained and saved machines seemed to be okay to recover. I also added thePlotlyJSpackage to my environment during my last session before this problem appeared. I am unsure of the importance of all this, but everything else I did that day was in the line of what I typically would do in a normal day.Versions
VSCode: v1.53.2
OS: Windows 10
Julia: v1.4.2
output of
] statusThe text was updated successfully, but these errors were encountered: