using MLDatasets is very slow

[CarloLucibello]

In a fresh julia 1.7 session

julia> @time using MLDatasets
 13.485246 seconds (20.31 M allocations: 1.158 GiB, 7.56% gc time, 61.69% compilation time)

Is there a way to conditionally import packages?

julia> for pkg in [:ImageCore, :CSV, :HDF5, :JLD2, :JSON3]; print(pkg); @time @eval using $pkg; end
ImageCore  2.141235 seconds (3.02 M allocations: 200.377 MiB, 4.50% gc time, 32.08% compilation time)
CSV  3.817959 seconds (6.14 M allocations: 348.493 MiB, 9.81% gc time, 90.11% compilation time)
HDF5  0.723358 seconds (1.34 M allocations: 73.225 MiB, 1.69% gc time, 93.80% compilation time)
JLD2  1.139716 seconds (1.36 M allocations: 78.966 MiB, 3.95% gc time, 60.77% compilation time)
JSON3  0.033367 seconds (49.09 k allocations: 3.014 MiB)

julia> for pkg in [:DataFrames, :MLUtils, :Pickle, :NPZ, :MAT]; print(pkg); @time @eval using $pkg; end
DataFrames  1.789793 seconds (2.03 M allocations: 137.197 MiB, 4.63% gc time)
MLUtils  1.743072 seconds (2.07 M allocations: 117.900 MiB, 4.83% gc time, 47.32% compilation time)
Pickle  0.130685 seconds (159.17 k allocations: 9.751 MiB, 17.77% compilation time)
NPZ  0.504406 seconds (1.19 M allocations: 61.838 MiB, 4.05% gc time, 98.87% compilation time)
MAT  0.009792 seconds (22.84 k allocations: 1.044 MiB)

Related discourse thread

[johnnychen94]

julia> versioninfo()
Julia Version 1.9.0-DEV.351
Commit 385762b444 (2022-04-08 21:50 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: 24 × 12th Gen Intel(R) Core(TM) i9-12900K
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-13.0.1 (ORCJIT, goldmont)
  Threads: 1 on 24 virtual cores

julia> @time_imports using MLDatasets
     18.5 ms        ┌ Preferences
     26.9 ms      ┌ JLLWrappers
     30.0 ms    ┌ OpenSSL_jll
      0.3 ms    ┌ Zlib_jll
     31.5 ms  ┌ HDF5_jll
      9.2 ms    ┌ MacroTools
      9.7 ms  ┌ ZygoteRules
      1.6 ms  ┌ GZip
      1.0 ms  ┌ ZipFile
      0.1 ms      ┌ IteratorInterfaceExtensions
      0.3 ms    ┌ TableTraits
    101.7 ms    ┌ SentinelArrays
     25.3 ms    ┌ Parsers
      0.9 ms    ┌ Compat
      4.9 ms    ┌ OrderedCollections
      1.3 ms      ┌ TranscodingStreams
      2.5 ms    ┌ CodecZlib
      0.1 ms    ┌ DataValueInterfaces
      9.5 ms    ┌ FilePathsBase
      5.0 ms      ┌ InlineStrings
      0.8 ms      ┌ DataAPI
     16.9 ms    ┌ WeakRefStrings
      8.5 ms    ┌ Tables
     10.7 ms    ┌ PooledArrays
   1348.2 ms  ┌ CSV
      0.4 ms  ┌ DefineSingletons
      0.4 ms    ┌ NaNMath
      0.9 ms      ┌ Adapt
     36.4 ms      ┌ OffsetArrays
     39.2 ms    ┌ PaddedViews
     24.4 ms    ┌ FixedPointNumbers
     48.3 ms      ┌ ChainRulesCore
     49.2 ms    ┌ ChangesOfVariables
      5.8 ms    ┌ AbstractFFTs
      0.3 ms    ┌ OpenLibm_jll
      1.6 ms      ┌ StackViews
      1.4 ms      ┌ MappedArrays
      4.1 ms    ┌ MosaicViews
      0.6 ms    ┌ InverseFunctions
      0.1 ms    ┌ Reexport
      3.7 ms    ┌ DocStringExtensions
    107.3 ms      ┌ ColorTypes
     76.2 ms      ┌ Colors
    185.5 ms    ┌ Graphics
      3.4 ms    ┌ IrrationalConstants
      0.4 ms    ┌ TensorCore
      1.4 ms      ┌ LogExpFunctions
    215.0 ms      ┌ OpenSpecFun_jll
    252.4 ms    ┌ SpecialFunctions
     80.6 ms    ┌ ColorVectorSpace
    888.7 ms  ┌ ImageCore
      0.5 ms    ┌ Requires
      0.9 ms    ┌ ConstructionBase
     24.4 ms    ┌ Setfield
     14.2 ms    ┌ InitialValues
     81.5 ms  ┌ BangBang
      0.3 ms    ┌ MbedTLS_jll
     32.6 ms  ┌ MbedTLS
     11.6 ms  ┌ FunctionWrappers
     57.0 ms  ┌ DataStructures
      0.2 ms    ┌ CompositionsBase
     21.0 ms  ┌ Accessors
      0.3 ms  ┌ InternedStrings
     18.4 ms    ┌ URIParser
    113.0 ms  ┌ BinDeps
     10.4 ms  ┌ StructTypes
      1.3 ms  ┌ ContextVariablesX
      0.3 ms    ┌ StatsAPI
      0.4 ms    ┌ SortingAlgorithms
      7.9 ms    ┌ Missings
     27.0 ms  ┌ StatsBase
      7.7 ms    ┌ MicroCollections
      0.4 ms    ┌ ArgCheck
      8.8 ms    ┌ SplittablesBase
     23.7 ms    ┌ Baselet
     87.6 ms  ┌ Transducers
      0.5 ms    ┌ Libiconv_jll
      3.7 ms  ┌ StringEncodings
      9.6 ms  ┌ JSON3
      2.5 ms  ┌ InvertedIndices
     97.1 ms    ┌ FileIO
    352.8 ms  ┌ NPZ
      2.4 ms    ┌ BufferedStreams
    715.8 ms    ┌ HDF5
    719.9 ms  ┌ MAT
    203.6 ms  ┌ MLStyle
      0.5 ms  ┌ IniFile
      0.9 ms  ┌ Glob
      6.0 ms    ┌ URIs
     54.3 ms  ┌ HTTP
      4.3 ms  ┌ DataDeps
    161.8 ms  ┌ JLD2
      1.7 ms  ┌ PrettyPrint
     29.3 ms    ┌ Crayons
      0.8 ms      ┌ Formatting
     88.3 ms    ┌ PrettyTables
    790.5 ms  ┌ DataFrames
      7.8 ms  ┌ ShowCases
    393.5 ms  ┌ FoldsThreads
      0.3 ms  ┌ FLoopsBase
      0.6 ms  ┌ NameResolution
      3.5 ms  ┌ JuliaVariables
      1.4 ms    ┌ TupleTools
     33.8 ms  ┌ Strided
      7.0 ms  ┌ FLoops
     17.2 ms  ┌ Pickle
      7.0 ms  ┌ MLUtils
   6173.8 ms  MLDatasets

[johnnychen94]

Is there a way to conditionally import packages?

One way is to let FileIO handle the IO of various formats. If you use the load/save interface, then you don't need to explicitly call using XXX in src/MLDatasets.jl

For NPZ and CSV we can delegate to CSVFiles.jl and NPZ.jl. Is JLD2 currently used in this package?

Might need to ask MAT to provide FileIO support

[johnnychen94]

If you want to build our own version of lazy loading, then we need to use invokelatest function. An small example can be found in https://github.com/JuliaIO/ImageIO.jl/blob/d8bbac7bb9c4367b4bf145b9cca5d49abc9e42ab/src/ImageIO.jl#L51-L61

[CarloLucibello]

I'll try to explore both the FileIO and the custom lazy loading.

Filed an issue for .mat JuliaIO/FileIO.jl#361