講座の内容を元にLightGBMモデル訓練（&振り返れる）スクリプトを実装する [CV 1.1836] [LB 1.2537]

ftnext / atmaCup10-paintings-likes

https://www.guruguru.science/competitions/16/

MIT License

0 stars 0 forks source link

params=None

$ python training.py data/preprocessed/v1/ data/datasets/train.csv submissions/

Training until validation scores don't improve for 100 rounds Did not meet early stopping. Best iteration is: [100] valid_0's l2: 1.43506 Fold 0 RMSLE: 1.1979 Training until validation scores don't improve for 100 rounds Did not meet early stopping. Best iteration is: [97] valid_0's l2: 1.37973 Fold 1 RMSLE: 1.1746 Training until validation scores don't improve for 100 rounds Did not meet early stopping. Best iteration is: [97] valid_0's l2: 1.3866 Fold 2 RMSLE: 1.1775 Training until validation scores don't improve for 100 rounds Did not meet early stopping. Best iteration is: [85] valid_0's l2: 1.44348 Fold 3 RMSLE: 1.2015 Training until validation scores don't improve for 100 rounds Did not meet early stopping. Best iteration is: [100] valid_0's l2: 1.41409 Fold 4 RMSLE: 1.1892 -------------------------------------------------- FINISHED | Whole RMSLE: 1.1882

dating_year_early 1915.0 StringLength__description 1861.0 dating_year_late 1806.0 StringLength__more_title 1803.0 CE__principal_maker 1354.0 StringLength__long_title 1194.0 StringLength__title 1061.0 StringLength__sub_title 874.0 CE__acquisition_method 517.0 CE__title 334.0 acquisition_method=transfer 188.0 acquisition_method=bequest 100.0 dating_period 89.0 acquisition_method=unknown 88.0 principal_maker=Samuel Bourne 85.0 principal_maker=Philip Henry Delamotte 68.0 acquisition_method=loan 67.0 acquisition_method=gift 60.0 principal_maker=Woodbury & Page 60.0 principal_maker=Raphaël Sadeler (I) 59.0 principal_maker=Théodore van Lelyveld 58.0 principal_maker=Charles Rochussen 55.0 principal_maker=Jan Luyken 52.0 principal_maker=Hendrick Goltzius 50.0 principal_maker=Jan van de Velde (II) 42.0 principal_maker=Jacob van der Schley 41.0 principal_maker=Atelier Kurkdjian 41.0 principal_maker=Jacques Lalaing 39.0 principal_maker=Jan Harmensz. Muller 37.0 principal_maker=Bernard Picart 31.0 principal_maker=Paulus Pontius 30.0 principal_maker=Cornelis Galle (I) 26.0 principal_maker=Henry W. Taunt 25.0 principal_maker=Romeyn de Hooghe 24.0 principal_maker=Jacob Olie jr. 19.0 principal_maker=Eberhard Cornelis Rahms 19.0 principal_maker=Giorgio Sommer 18.0 principal_maker=Johann Sadeler (I) 17.0 principal_maker=anoniem (Monumentenzorg) 17.0 principal_maker=Daniël Veelwaard (I) 17.0 principal_maker=Bonfils 17.0 principal_maker=Hermanus Jan Hendrik van Rijkelijkhuysen 14.0 principal_maker=Adolphe Burdet 14.0 principal_maker=Jan Frederik Christiaan Reckleben 13.0 principal_maker=Aegidius Sadeler 11.0 principal_maker=Juan Laurent 11.0 principal_maker=Theodoor Koning 10.0 principal_maker=Roelant Roghman 6.0 principal_maker=Richard Tepe 5.0 principal_maker=Fratelli Alinari 5.0

params=lgbm_params [CV 1.1836] [LB 1.2537]

training.pyを変更して実行 $ python training.py data/preprocessed/v1/ data/datasets/train.csv submissions/

[LightGBM] [Warning] Accuracy may be bad since you didn't explicitly set num_leaves OR 2^max_depth > num_leaves. (num_leaves=31). Training until validation scores don't improve for 100 rounds [500] valid_0's rmse: 1.19982 Early stopping, best iteration is: [568] valid_0's rmse: 1.1983 Fold 0 RMSLE: 1.1983 [LightGBM] [Warning] Accuracy may be bad since you didn't explicitly set num_leaves OR 2^max_depth > num_leaves. (num_leaves=31). Training until validation scores don't improve for 100 rounds Early stopping, best iteration is: [324] valid_0's rmse: 1.17177 Fold 1 RMSLE: 1.1718 [LightGBM] [Warning] Accuracy may be bad since you didn't explicitly set num_leaves OR 2^max_depth > num_leaves. (num_leaves=31). Training until validation scores don't improve for 100 rounds [500] valid_0's rmse: 1.19045 Early stopping, best iteration is: [461] valid_0's rmse: 1.18878 Fold 2 RMSLE: 1.1888 [LightGBM] [Warning] Accuracy may be bad since you didn't explicitly set num_leaves OR 2^max_depth > num_leaves. (num_leaves=31). Training until validation scores don't improve for 100 rounds Early stopping, best iteration is: [378] valid_0's rmse: 1.19028 Fold 3 RMSLE: 1.1903 [LightGBM] [Warning] Accuracy may be bad since you didn't explicitly set num_leaves OR 2^max_depth > num_leaves. (num_leaves=31). Training until validation scores don't improve for 100 rounds [500] valid_0's rmse: 1.16945 Early stopping, best iteration is: [492] valid_0's rmse: 1.16845 Fold 4 RMSLE: 1.1684 -------------------------------------------------- FINISHED | Whole RMSLE: 1.1836

StringLength__sub_title 115919.8970608539 CE__principal_maker 78480.97932646889 dating_year_late 57125.873609007336 StringLength__more_title 48863.97797768813 CE__acquisition_method 46840.115645824 dating_year_early 41359.81569700455 StringLength__description 41205.97698495025 acquisition_method=transfer 28292.064049318433 StringLength__long_title 19234.030936465228 StringLength__title 18824.523672675597 dating_period 17687.047488784418 CE__title 8682.455806170823 acquisition_method=bequest 7628.177280128002 acquisition_method=purchase 2390.332360152155 acquisition_method=loan 2006.157952874899 acquisition_method=unknown 1756.5118776634336 acquisition_method=gift 1599.86732339859 principal_maker=Samuel Bourne 1444.4636361300945 principal_maker=Jacob Olie jr. 1073.2613588571548 principal_maker=anonymous 935.7564471065998 principal_maker=Philip Henry Delamotte 816.7143234610558 principal_maker=Raphaël Sadeler (I) 723.7326199412346 principal_maker=Bernard Picart 682.2669063210487 principal_maker=Woodbury & Page 666.6505894064903 principal_maker=Jacques Lalaing 652.1705242991447 principal_maker=Atelier Kurkdjian 573.5614114403725 principal_maker=Théodore van Lelyveld 553.4593561291695 principal_maker=Hendrick Goltzius 504.37551349401474 principal_maker=Richard Tepe 377.38423067331314 principal_maker=Jan Luyken 363.742679476738 principal_maker=Adolphe Burdet 360.61634880304337 principal_maker=Jacob van der Schley 346.937460064888 principal_maker=Jan Harmensz. Muller 337.7850676178932 principal_maker=Charles Rochussen 288.1016381382942 principal_maker=Hermanus Jan Hendrik van Rijkelijkhuysen 268.6380909085274 principal_maker=Romeyn de Hooghe 222.4745261669159 principal_maker=Jan van de Velde (II) 221.83499509096146 principal_maker=Giorgio Sommer 221.3071709871292 principal_maker=Johann Sadeler (I) 205.87370175123215 principal_maker=Cornelis Galle (I) 197.35874596238136 principal_maker=Jacob de Gheyn (II) 196.14085227251053 principal_maker=Aegidius Sadeler 193.35006448626518 principal_maker=Paulus Pontius 192.17733943462372 principal_maker=Waldemar Titzenthaler 190.73536175489426 principal_maker=anoniem (Monumentenzorg) 163.71789973974228 principal_maker=Daniël Veelwaard (I) 144.55989742279053 principal_maker=Bonfils 133.13036489486694 principal_maker=Johann Heinrich Maria Hubert Rennefeld 112.56654340028763 principal_maker=Jan Frederik Christiaan Reckleben 110.72550246119499 principal_maker=Roelant Roghman 105.06099009513855

カテゴリ変数の閾値を20と講義と揃えて（v1.1）、`lgbm_params`指定

CVは改善。特徴量が多いほうがいいらしい（モデルに判断させるべきらしい）

facetを見て以下のように設定していた（v1データ）

acquisition_method 200
principal_maker 30

$ python training.py data/preprocessed/v1.1/ data/datasets/train.csv submissions/
[LightGBM] [Warning] Accuracy may be bad since you didn't explicitly set num_leaves OR 2^max_depth > num_leaves. (num_leaves=31).
Training until validation scores don't improve for 100 rounds
[500]   valid_0's rmse: 1.19089
Early stopping, best iteration is:
[521]   valid_0's rmse: 1.18976
Fold 0 RMSLE: 1.1898
[LightGBM] [Warning] Accuracy may be bad since you didn't explicitly set num_leaves OR 2^max_depth > num_leaves. (num_leaves=31).
Training until validation scores don't improve for 100 rounds
[500]   valid_0's rmse: 1.17152
Early stopping, best iteration is:
[409]   valid_0's rmse: 1.16842
Fold 1 RMSLE: 1.1684
[LightGBM] [Warning] Accuracy may be bad since you didn't explicitly set num_leaves OR 2^max_depth > num_leaves. (num_leaves=31).
Training until validation scores don't improve for 100 rounds
[500]   valid_0's rmse: 1.17761
Early stopping, best iteration is:
[428]   valid_0's rmse: 1.17497
Fold 2 RMSLE: 1.1750
[LightGBM] [Warning] Accuracy may be bad since you didn't explicitly set num_leaves OR 2^max_depth > num_leaves. (num_leaves=31).
Training until validation scores don't improve for 100 rounds
Early stopping, best iteration is:
[379]   valid_0's rmse: 1.19238
Fold 3 RMSLE: 1.1924
[LightGBM] [Warning] Accuracy may be bad since you didn't explicitly set num_leaves OR 2^max_depth > num_leaves. (num_leaves=31).
Training until validation scores don't improve for 100 rounds
Early stopping, best iteration is:
[384]   valid_0's rmse: 1.16635
Fold 4 RMSLE: 1.1664
--------------------------------------------------
FINISHED | Whole RMSLE: 1.1784

20210308-092436_data_preprocessed_v1 1

StringLength__sub_title 209930.162448043
dating_year_late    42569.24030684505
CE__principal_maker 41079.74357781559
dating_year_early   40140.34276960185
StringLength__description   39910.90862703486
StringLength__more_title    34398.36773852137
StringLength__title 18186.9582782457
acquisition_method=transfer 16716.9609423019
StringLength__long_title    15554.007925360755
CE__acquisition_method  14547.523025380448
dating_period   8276.197126097977
CE__title   8072.621644140054
principal_maker=Rembrandt van Rijn  6190.913442015648
acquisition_method=bequest  5191.203318551183
acquisition_method=purchase 4008.6148973125964
principal_maker=Herman Salzwedel    3705.6628245711327
acquisition_method=loan 2866.6571811437607
principal_maker=Jean Baptiste Vanmour   1783.4732909202576
principal_maker=George Hendrik Breitner 1737.422084748745
acquisition_method= 1643.8557840585709
acquisition_method=unknown  1628.8673606552184
acquisition_method=gift 1510.0922226384282
principal_maker=Jacob Olie jr.  1312.0474618077278
principal_maker=Samuel Bourne   1299.6625580787659
acquisition_method=nationalization 1795 1037.8393313884735
principal_maker=Raphaël Sadeler (I) 731.2425500154495
principal_maker=Philip Henry Delamotte  707.3953838348389
principal_maker=Atelier Kurkdjian   668.155992269516
principal_maker=anonymous   643.7922441363335
principal_maker=Woodbury & Page 619.2181321382523
principal_maker=Jacques Lalaing 595.9418876171112
principal_maker=Adolphe Burdet  579.9195005297661
principal_maker=Théodore van Lelyveld   578.0495392084122
principal_maker=Hermanus Jan Hendrik van Rijkelijkhuysen    517.6817879080772
principal_maker=Johann Sadeler (I)  444.7787608206272
principal_maker=Hendrick Goltzius   438.8445661664009
principal_maker=Jan Luyken  409.46383661031723
principal_maker=Bernard Picart  395.66525959968567
principal_maker=anoniem (Monumentenzorg)    376.42573696374893
principal_maker=James Higson    375.6434725522995
principal_maker=Cornelis Galle (I)  338.6184870004654
principal_maker=O. Hisgen & Co. 331.26985543966293
principal_maker=Ohannes Kurkdjian   319.7526163458824
principal_maker=Paulus Pontius  311.48908030986786
principal_maker=Charles Rochussen   306.03656232357025
principal_maker=Jacob van der Schley    305.7295599579811
principal_maker=Waldemar Titzenthaler   293.17414116859436
principal_maker=Jan Harmensz. Muller    292.3039834499359
principal_maker=Jacob de Gheyn (II) 227.67760467529297
principal_maker=Schelte Adamsz. Bolswert    204.3831462264061

ftnext / atmaCup10-paintings-likes