大規模データの機械学習におけるDaskの活用

@PyData.Tokyo One Day Conference 2018/10/20

Transcript

1. େن໛σʔλͷ ػցֶशʹ͓͚ΔDaskͷ׆༻ Masaaki Horikoshi @ ARISE analytics
2. ࣗݾ঺հ • OSS׆ಈͳͲ: • GitHub: https://github.com/sinhrks
3. ίϯςϯπ • Daskͱ͸ • ػցֶशʹ͓͚ΔDaskͷ׆༻ • Dask ML • Distributed
4. Python for Big Data • PythonͰେن໛ͳσʔλΛॲཧ͢Δͱ… • ܭࢉ͕(ݪଇ)୯ҰͷεϨουͰߦΘΕΔͨΊɺ ॲཧ଎౓͕஗͍ • σʔλ͕෺ཧతͳϝϞϦʹ͓͞·Βͳ͍
5. ฒྻॲཧͱOut-of-coreॲཧ ฒྻॲཧ 0VUPGDPSFॲཧ • ฒྻॲཧ: ෳ਺ͷλεΫΛฒྻͰॲཧ͢Δ • Out-of-coreॲཧ: ϝϞϦʹ৐Βͳ͍σʔλΛஞ࣍ॲཧ͢Δ • ഉଞͰ͸ͳ͍ (Out-of-coreॲཧΛฒྻͰߦ͏͜ͱ΋͋Δ)
6. Daskͱ͸ • ॊೈͳฒྻɾOut-of-coreॲཧύοέʔδ • σʔλॲཧɺ਺஋ܭࢉΛओ໨త • NumPy΍pandasͷαϒηοτͱͳΔσʔλ ߏ଄Λఏڙ • ॲཧΛܭࢉάϥϑͱͯ͠දݱ͠ɺಈతʹεέ δϡʔϦϯά࣮ͯ͠ߦ
7. (Incomplete) List of OSS uses Dask • (TFLearn) Deep learning library featuring a higher-level API for TensorFlow. • (Distributed Scheduler) A platform to author, schedule and monitor workflows. • Image Processing SciKit. • N-D labeled arrays and datasets in Python. • Executes end-to-end data science and analytics pipelines entirely on GPUs. Airflow
8. Daskͷσʔλߏ଄(API) %BTL
   "1* #BTF
   $MBTT %FGBVMU
   4DIFEVMFS %BTL
   "SSBZ /VN1Z
   OEBSSBZ UISFBEJOH %BTL
   %BUB'SBNF QBOEBT
   %BUB'SBNF
   UISFBEJOH %BTL
   #BH 1Z5PPM[
   MJTU
   TFU
   EJDU NVMUJQSPDFTTJOH
   DBOOPU
   SFMFBTF
   (*- %BTL
   %FMBZFE ೚ҙͷؔ਺ UISFBEJOH
9. Dask Array • ෳ਺ͷNumPy nd-arrayʹΑͬͯߏ੒ • ಺෦తʹ͸ɺ࣠ํ޲ʹԊͬͯChunkʹ෼ׂ /VN1Z
   OEBSSBZ %BTL
   "SSBZ $IVOL DIVOLTJ[F
10. array([[ 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.], [ 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.], [ 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.], [ 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.], [ 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.], [ 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.], [ 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.], [ 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.], [ 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.], [ 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.]]) Dask Array import numpy as np x = np.ones((10, 10)) x import dask.array as da dx = da.ones((10, 10), chunks=(5, 5)) dx dask.array<ones, shape=(10, 10), dtype=float64, chunksize=(5, 5)> 
    Y
    
    ͷ
    OEBSSBZ
    Λ࡞੒ 
    Y
    
    ͷ
    %BTL
    "SSBZ
    Λ࡞੒ ಺෦͸
    Y
    ͷͭͷDIVOLʹ෼ׂ
    
11. Dask Array import dask.array as da dx = da.ones((10, 10), chunks=(5, 5)) dx dask.array<ones, shape=(10, 10), dtype=float64, chunksize=(5, 5)> dx.visualize() ಺෦ͷܭࢉάϥϑΛදࣔ
    ֤ϊʔυ ༿ ͕֤DIVOLʹରԠ
    
    Y
    
    ͷ
    %BTL
    "SSBZ
    Λ࡞੒ ಺෦͸
    Y
    ͷͭͷDIVOLʹ෼ׂ
    
12. 
    ߦ
    
    ྻͷ
    QBOEBT%BUB'SBN F
    Λ࡞੒ Dask Array dy = dx.sum(axis=0) dy dask.array<sum-aggregate, shape=(10,), dtype=float64, chunksize=(5,)> BYJTʹԊͬͯ஋Λ߹ܭ dy.visualize() dy.compute() array([10., 10., 10., 10., 10., 10., 10., 10., 10., 10.]) 4VN 4VN
13. Dask Array dy2 = dx.sum() dy2 dask.array<sum-aggregate, shape=(), dtype=float64, chunksize=()> શͯͷ஋Λ߹ܭ dy2.visualize() dy2.compute() 100.0
14. Dask Array • Dask Array ͸೚ҙͷ shape ͱ chunk sizeΛ΋ͭ • chunkؒͷॲཧ͕ൃੜ͢Δ৔߹ɺchunk size ͸ Ұகͤͨ͞ํ͕ྑ͍ da.ones((30, 20, 1, 15), chunks=(3, 7, 1, 2)) dask.array<ones, shape=(30, 20, 1, 15), dtype=float64, chunksize=(3, 7, 1, 2)>
15. Dask DataFrame QBSUJUJPO EJWJTJPO EJWJTJPO • ෳ਺ͷpandas DataFrameʹΑͬͯߏ੒ • ಺෦తʹ͸ɺindex (ߦϥϕϧ)ʹԊͬͯPartitionʹ෼ׂ QBOEBT
    %BUB'SBNF %BTL
    %BUB'SBNF
16. 
    ߦ
    
    ྻͷ
    QBOEBT%BUB'SBNF
    Λ࡞੒ Dask DataFrame import dask.dataframe as dd ddf = dd.from_pandas(df, 2) ddf QBSUJUJPO QBSUJUJPO EJWJTJPO EJWJTJPO EJWJTJPO
17. Blocked Algorithm (Mean) ddf.mean().visualize() 4VN
    PWFS
    QBSUJUJPOT $PVOU
    PWFS
    QBSUJUJPOT .FBO
    
    4VN
    
    $PVOU 4VN
    QFS
    QBSUJUJPO $PVOU
    QFS
    QBSUJUJPO
18. ܭࢉάϥϑͷ࣮ߦΠϝʔδ • ܭࢉάϥϑͷґଘؔ܎ͷղੳ • ࠷ऴ݁ՌʹෆཁͳλεΫͷ࡟আ • ಉҰλεΫͷϚʔδ • ܭࢉॱংͷ੩తɾಈతͳղੳ • ܭࢉάϥϑͷ࠷దԽ • ಛఆॲཧͷΠϯϥΠϯԽ • ࿈ଓ͢ΔλεΫΛಉҰͷϫʔΧʔʹׂΓ౰ͯ
19. Dask Internals • ͋ΒΏΔDaskͷॲཧ͸ܭࢉάϥϑͱͯ͠දݱ͞ ΕΔ • DaskͰ͸ઢܗ୅਺ͳͲͷෳࡶͳΞϧΰϦζϜ ΋࣮૷ (dask.array.linalg) • Ϣʔβࣗ਎ͷσʔλߏ଄΍ؔ਺΋ܭࢉάϥϑ Ͱ࣮૷Մೳ (dask.delayed)
20. Linear Algebra 'VODUJPO %FTDSJQUJPO MJOBMHDIPMFTLZ 3FUVSOT
    UIF
    $IPMFTLZ
    EFDPNQPTJUJPO
    PS
    PG
    B
    )FSNJUJBO
    QPTJUJWFEFpOJUF
    NBUSJY
    " MJOBMHJOW $PNQVUF
    UIF
    JOWFSTF
    PG
    B
    NBUSJY
    XJUI
    -6
    EFDPNQPTJUJPO
    BOE
    GPSXBSE
    
    CBDLXBSE
    TVCTUJUVUJPOT MJOBMHMTUTR 3FUVSO
    UIF
    MFBTUTRVBSFT
    TPMVUJPO
    UP
    B
    MJOFBS
    NBUSJY
    FRVBUJPO
    VTJOH
    23
    EFDPNQPTJUJPO MJOBMHMV $PNQVUF
    UIF
    MV
    EFDPNQPTJUJPO
    PG
    B
    NBUSJY MJOBMHOPSN .BUSJY
    PS
    WFDUPS
    OPSN MJOBMHRS $PNQVUF
    UIF
    RS
    GBDUPSJ[BUJPO
    PG
    B
    NBUSJY MJOBMHTPMWF 4PMWF
    UIF
    FRVBUJPO
    B
    Y
    
    C
    GPS
    Y MJOBMHTPMWF@USJBOHVMBS 4PMWF
    UIF
    FRVBUJPO
    B
    Y
    
    C
    GPS
    Y
    BTTVNJOH
    B
    JT
    B
    USJBOHVMBS
    NBUSJY MJOBMHTWE $PNQVUF
    UIF
    TJOHVMBS
    WBMVF
    EFDPNQPTJUJPO
    PG
    B
    NBUSJY MJOBMHTWE@DPNQSFTTFE 3BOEPNMZ
    DPNQSFTTFE
    SBOLL
    UIJO
    4JOHVMBS
    7BMVF
    %FDPNQPTJUJPO MJOBMHTGRS %JSFDU
    4IPSUBOE
    'BU
    23 MJOBMHUTRS %JSFDU
    5BMMBOE4LJOOZ
    23
    BMHPSJUIN
21. Example: LU decomposition • LU෼ղ • LU෼ղ͸νϟϯΫ(ϒϩοΫ)୯ҐͰͷܭࢉʹ෼ׂ Ͱ͖Δ "ɹɹɹ
    ɹ
    
    
    
    
    
    
    
    
    
    
    
    
    
    -
    
    
    
    
    
    
    
    
    
    
    Y
    
    
    
    
    
    
    
    
    6
22. Blocked LU Decomposition • Diagonal Block • Row-direction(i < j) • Columns direction (i < j) ∴ ∴ * LU: Function to solve LU decomposition * Solve: Function to solve equation
23. Blocked LU Decomposition arr = da.random.random((9, 9), chunks=(3, 3)) arr dask.array<random_sample, shape=(9, 9), dtype=float64, chunksize=(3, 3)> from dask import compute t, l, u = da.linalg.lu(arr) t, l, u = compute(t, l, u) ͭͷܭࢉάϥϑΛ
    Ϛʔδ࣮ͯ͠ߦ
24. Blocked LU Decomposition from dask import visualize visualize(t, l, u)
25. DaskʹΑΔॲཧͷશମ૾ ஞ࣍ॲཧ /VN1Z QBOEBT TDJLJUMFBSO %BTL %BTL
    %JTUSJCVUFE ฒྻॲཧ
    ϊʔυ಺ 0VUPGDPSF
    ॲཧ ϊʔυ಺ ෼ࢄॲཧ
    ϊʔυؒ %BTL
    .- ߦྻܭࢉ ςʔϒϧ
    σʔλॲཧ ػցֶश
26. scikit-learnͷฒྻॲཧ • “n_jobs” Ҿ਺Ͱฒྻ࣮ߦ਺Λࢦఆ • ಺෦తʹ͸joblibΛར༻ • scikit-learnίϛολத৺ʹ։ൃ • ϊʔυ಺ฒྻ (threading, multiprocessing) • Out-of-coreॲཧ΍ϊʔυؒ෼ࢄॲཧ͸Ͱ͖ͳ͍ from sklearn.model_selection import GridSearchCV grid = GridSearchCV(pipe, cv=3, n_jobs=12, param_grid=param_grid)
27. DaskʹΑΔॲཧͷશମ૾ ஞ࣍ॲཧ /VN1Z QBOEBT TDJLJUMFBSO %BTL %BTL
    %JTUSJCVUFE ฒྻॲཧ
    ϊʔυ಺ 0VUPGDPSF
    ॲཧ ϊʔυ಺ ෼ࢄॲཧ
    ϊʔυؒ %BTL
    .- ߦྻܭࢉ ςʔϒϧ
    σʔλॲཧ ػցֶश
28. ຊ౰ʹશͯͷσʔλ͕ඞཁ͔ʁ • αϯϓϦϯάͰे෼Ͱ͸ʁ IUUQTDJLJUMFBSOPSHTUBCMFBVUP@FYBNQMFTNPEFM@TFMFDUJPOQMPU@MFBSOJOH@DVSWFIUNM
29. Dask ML • ػցֶशͰDaskΛ׆༻͢ΔͨΊͷύοέʔδ • Daskͷσʔλߏ଄ʹରͯ͠ػցֶशΛߦ͏ • ػցֶशΞϧΰϦζϜΛDaskΛ༻͍ͯฒྻɾ Out-of-coreԽ͢Δ
30. Dask ML • ػցֶशͷͨΊͷαϒύοέʔδΛఏڙ • Preprocessing • Model Selection • Cross validation • Hyper parameter search • GLM • Clustering • Incremental • ParallelPostFit • XGBoost • TensorFlow %BTLʹରԠͨ͠
    TDJLJUMFBSOޓ׵ͷ
    ֶशثΛఏڙ TDJLJUMFBSOΛϥοϓ͠
    ฒྻɾ0VUPGDPSFԽ TDJLJUMFBSOҎ֎ͷ
    ύοέʔδରԠ
31. • scikit-learnޓ׵ͷֶशثͰDaskͷσʔλߏ଄Λ ѻ͑Δ Dask ML TDJLJU MFBSO %BTL
    .- /VN1Z
    OEBSSBZ %BTL
    "SSBZ /VN1Z
    OEBSSBZ /VN1Z
    OEBSSBZ %BTL
    "SSBZ %BTL
    "SSBZ "SSBZ
    *OUFSGBDFʹΑΓɺ /VN1Z
    OEBSSBZʹม׵ ܭࢉॲཧ ܭࢉॲཧ %BTLͷσʔλߏ଄Λҡ࣋ͯ͠
    ॲཧΛ࣮ߦ
32. k-means IUUQTFOXJLJQFEJBPSHXJLJ,NFBOT@DMVTUFSJOH
33. k-means (Dask ML) ʜ $IVOL ʜ $IVOL ʜ $IVOL 4BNQMJOH σʔληοτ͔Βη ϯτϩΠυͷॳظ஋ Λܾఆ ηϯτϩΠυ͔Βͷ ڑ཭Λܭࢉ͠ɺΫϥ ελʹ෼ྨ
    $IVOL͝ͱʹ࣮ߦ Ϋϥελ͝ͱʹϨίʔ υΛू໿͠ɺηϯτ ϩΠυΛߋ৽
    $IVOL͝ͱʹ࣮ߦˠ ू໿ 4VN 4VN
    $PVOU ʜ طఆͰ͸
    4DBMBCMF
    ,NFBOT
    #BINBOJ
    FU
    BM
    
    ͷ,NFBOTcc
    ͰηϯτϩΠυΛॳظԽ
34. Incremental • partial_fitΛॱʹద༻͢Δ 
    ߦ
    
    ྻͷ
    QBOEBT%BUB'SBNF
    Λ࡞੒ pU %BTL
    "SSBZ *ODSFNFOUBM QBSUJBM@pU $IVOL $IVOL ʜ ʜ QBSUJBM@pUΛஞ࣮࣍ߦ
    ฒྻॲཧ͞Εͳ͍
35. Incremental • IncrementalͰֶशثΛϥοϓ • ෼ྨ໰୊ͷ৔߹ɺΫϥεΛclassesҾ਺Ͱ౉͢ • ෳ਺ΤϙοΫͰֶश͢Δ৔߹͸ Incremental.partial_fit • for _ in range(10): inc.partial_fit(X_train, y_train, classes=classes) print('Score:', inc.score(X_test, y_test)) from sklearn.linear_model import SGDClassifier from dask_ml.wrappers import Inclemental clf = SGDClassifier(loss='log', penalty='l2', tol=1e-3)) inc = Incremental(clf, scoring='accuracy') inc.fit(X_train, y_train, classes=classes)
36. 
    ߦ
    
    ྻͷ
    QBOEBT%BUB'SBNF
    Λ࡞੒ ParallelPostFit • ֶशࡁΈEstimatorͷtransform΍predictΛฒྻͰ ద༻͢Δ QSFEJDU %BTL
    "SSBZ 1BSBMMFM1PTU'JU QSFEJDU $IVOL QSFEJDU $IVOL ʜ ʜ
37. • ParallelPostFitͰֶशثΛϥοϓ • fit࣌͸Կ΋ߦΘͳ͍ (scikit-learnͷॲཧͱಉҰ) • ΑΓେ͖ͳσʔλʹରͯ͠ predict ΛฒྻͰߦ͏ ParallelPostFit y_pred = clf.predict(X_large) y_pred from sklearn.linear_model import LogisticRegressionCV from dask_ml.wrappers import ParallelPostFit clf = ParallelPostFit(LogisticRegressionCV(cv=3)) clf.fit(X_train, y_train)
38. DaskʹΑΔॲཧͷશମ૾ ஞ࣍ॲཧ /VN1Z QBOEBT TDJLJUMFBSO %BTL %BTL
    %JTUSJCVUFE ฒྻॲཧ
    ϊʔυ಺ 0VUPGDPSF
    ॲཧ ϊʔυ಺ ෼ࢄॲཧ
    ϊʔυؒ %BTL
    .- ߦྻܭࢉ ςʔϒϧ
    σʔλॲཧ ػցֶश
39. Dask Distributed • Daskຊମ͸ҎԼ2छྨͷεέδϡʔϥΛαϙʔ τ: • threading • multiprocessing • Dask Distributedύοέʔδ͸ϊʔυؒ෼ࢄΛߦ ͏εέδϡʔϥΛఏڙ
40. Dask Distributed • εέδϡʔϥͰͷܭࢉ࣮ߦΛෳ਺ϊʔυͰ෼ࢄͰ͖Δ • ௿ϨΠςϯγ: λεΫຖͷΦʔόʔϔου͸1msఔ౓ • WorkerؒͰͷσʔλڞ༗: σʔλసૹ͸WorkerؒͰ௚઀࣮ࢪ • ෳࡶͳεέδϡʔϦϯά: ೚ҙͷܭࢉάϥϑΛ࣮ߦՄ • ہॴੑ: WorkerؒͷσʔλసૹΛͳΔ΂͘ߦΘͳ͍ %JTUSJCVUFE
    8PSLFS %JTUSJCVUFE
    8PSLFS %JTUSJCVUFE
    4DIFEVMFS %JTUSJCVUFE
    $MJFOU
41. Distributed joblib • ϓϥΨϒϧAPI • with ϒϩοΫͰ joblib.Parallel ͷطఆόοΫΤϯυΛมߋՄ • ஫ҙ఺ • scikit-learnʹόϯυϧ͞Ε͍ͯΔjoblibΛ࢖͏ (sklearn.externals.joblib) • ෼ࢄͰ͖ͳ͍৔߹΋͋Δ • backendͱͯ͠threading / multiprocessing͕໌ࣔ͞Ε͍ͯΔ΋ͷ import distributed.joblib from sklearn.externals.joblib import parallel_backend with parallel_backend('dask.distributed', scheduler_host=‘scheduler-addr:8786’): grid.fit(digits.data, digits.target)
42. ·ͱΊ • Daskͱ͸ • NumPy΍ pandas ޓ׵ͷσʔλߏ଄Λఏڙ • ػցֶशʹ͓͚ΔDaskͷ׆༻ • Dask ML: ػցֶशͰDaskͷσʔλߏ଄Λѻ͑ Δ • Distributed: scikit-learnͷॲཧΛϊʔυؒͰ෼ࢄ ॲཧͰ͖Δ