Contents

데모

라이브러리 import 및 설정

%reload_ext autoreload
%autoreload 2
%matplotlib inline

from matplotlib import pyplot as plt
from matplotlib import rcParams
import numpy as np
from pathlib import Path
import pandas as pd
from sklearn.metrics import accuracy_score
from sklearn.model_selection import StratifiedKFold
from sklearn.preprocessing import StandardScaler, PolynomialFeatures
import seaborn as sns
from tensorflow import keras
from tensorflow.keras import Input, Model, Sequential
from tensorflow.keras.callbacks import EarlyStopping, ReduceLROnPlateau
from tensorflow.keras.layers import Dense
from tensorflow.keras.utils import plot_model, to_categorical
import warnings

rcParams['figure.figsize'] = (16, 8)
plt.style.use('fivethirtyeight')
pd.set_option('max_columns', 100)
pd.set_option("display.precision", 4)
warnings.simplefilter('ignore')

학습데이터 로드

03-pandas-eda.ipynb에서 생성한 feature.csv 피처파일 사용

data_dir = Path('../data/dacon-dku')
feature_dir = Path('../build/feature')
val_dir = Path('../build/val')
tst_dir = Path('../build/tst')
sub_dir = Path('../build/sub')

trn_file = data_dir / 'train.csv'
tst_file = data_dir / 'test.csv'
sample_file = data_dir / 'sample_submission.csv'

target_col = 'class'
n_fold = 5
n_class = 3
seed = 42

algo_name = 'nncv'
feature_name = 'polyfeature'
model_name = f'{algo_name}_{feature_name}'

feature_file = feature_dir / f'{feature_name}.csv'
p_val_file = val_dir / f'{model_name}.val.csv'
p_tst_file = tst_dir / f'{model_name}.tst.csv'
sub_file = sub_dir / f'{model_name}.csv'

Polynomial Feature 생성

df = pd.read_csv(feature_dir / 'feature.csv', index_col=0)
print(df.shape)
df.head()

(400000, 20)
z redshift dered_u dered_g dered_r dered_i dered_z nObserve airmass_u class d_dered_u d_dered_g d_dered_r d_dered_i d_dered_z d_dered_ig d_dered_zg d_dered_rz d_dered_iz d_obs_det
id
0 16.9396 -8.1086e-05 23.1243 20.2578 18.9551 17.6321 16.9089 2.9444 1.1898 0.0 -0.1397 -0.0790 -0.0544 -0.0403 -0.0307 -2.6257 -3.3488 2.0462 0.7232 -15.0556
1 13.1689 4.5061e-03 14.9664 14.0045 13.4114 13.2363 13.1347 0.6931 1.2533 1.0 -0.0857 -0.0574 -0.0410 -0.0322 -0.0343 -0.7683 -0.8698 0.2767 0.1016 -0.3069
2 15.3500 4.7198e-04 16.6076 15.6866 15.4400 15.3217 15.2961 1.0986 1.0225 0.0 -0.1787 -0.1388 -0.0963 -0.0718 -0.0540 -0.3649 -0.3905 0.1440 0.0257 -0.9014
3 19.6346 5.8143e-06 25.3536 20.9947 20.0873 19.7947 19.5552 1.6094 1.2054 0.0 -0.3070 -0.1941 -0.1339 -0.1003 -0.0795 -1.2000 -1.4395 0.5321 0.2395 -1.3906
4 17.9826 -3.3247e-05 23.7714 20.4338 18.8630 18.1903 17.8759 2.6391 1.1939 0.0 -0.6820 -0.2653 -0.1794 -0.1339 -0.1067 -2.2436 -2.5579 0.9871 0.3144 -9.3609 
scaler = StandardScaler()
poly = PolynomialFeatures(2)
X = poly.fit_transform(scaler.fit_transform(df.drop(target_col, axis=1)))
feature_names = poly.get_feature_names(df.columns)
feature_names

['1',
 'z',
 'redshift',
 'dered_u',
 'dered_g',
 'dered_r',
 'dered_i',
 'dered_z',
 'nObserve',
 'airmass_u',
 'class',
 'd_dered_u',
 'd_dered_g',
 'd_dered_r',
 'd_dered_i',
 'd_dered_z',
 'd_dered_ig',
 'd_dered_zg',
 'd_dered_rz',
 'd_dered_iz',
 'z^2',
 'z redshift',
 'z dered_u',
 'z dered_g',
 'z dered_r',
 'z dered_i',
 'z dered_z',
 'z nObserve',
 'z airmass_u',
 'z class',
 'z d_dered_u',
 'z d_dered_g',
 'z d_dered_r',
 'z d_dered_i',
 'z d_dered_z',
 'z d_dered_ig',
 'z d_dered_zg',
 'z d_dered_rz',
 'z d_dered_iz',
 'redshift^2',
 'redshift dered_u',
 'redshift dered_g',
 'redshift dered_r',
 'redshift dered_i',
 'redshift dered_z',
 'redshift nObserve',
 'redshift airmass_u',
 'redshift class',
 'redshift d_dered_u',
 'redshift d_dered_g',
 'redshift d_dered_r',
 'redshift d_dered_i',
 'redshift d_dered_z',
 'redshift d_dered_ig',
 'redshift d_dered_zg',
 'redshift d_dered_rz',
 'redshift d_dered_iz',
 'dered_u^2',
 'dered_u dered_g',
 'dered_u dered_r',
 'dered_u dered_i',
 'dered_u dered_z',
 'dered_u nObserve',
 'dered_u airmass_u',
 'dered_u class',
 'dered_u d_dered_u',
 'dered_u d_dered_g',
 'dered_u d_dered_r',
 'dered_u d_dered_i',
 'dered_u d_dered_z',
 'dered_u d_dered_ig',
 'dered_u d_dered_zg',
 'dered_u d_dered_rz',
 'dered_u d_dered_iz',
 'dered_g^2',
 'dered_g dered_r',
 'dered_g dered_i',
 'dered_g dered_z',
 'dered_g nObserve',
 'dered_g airmass_u',
 'dered_g class',
 'dered_g d_dered_u',
 'dered_g d_dered_g',
 'dered_g d_dered_r',
 'dered_g d_dered_i',
 'dered_g d_dered_z',
 'dered_g d_dered_ig',
 'dered_g d_dered_zg',
 'dered_g d_dered_rz',
 'dered_g d_dered_iz',
 'dered_r^2',
 'dered_r dered_i',
 'dered_r dered_z',
 'dered_r nObserve',
 'dered_r airmass_u',
 'dered_r class',
 'dered_r d_dered_u',
 'dered_r d_dered_g',
 'dered_r d_dered_r',
 'dered_r d_dered_i',
 'dered_r d_dered_z',
 'dered_r d_dered_ig',
 'dered_r d_dered_zg',
 'dered_r d_dered_rz',
 'dered_r d_dered_iz',
 'dered_i^2',
 'dered_i dered_z',
 'dered_i nObserve',
 'dered_i airmass_u',
 'dered_i class',
 'dered_i d_dered_u',
 'dered_i d_dered_g',
 'dered_i d_dered_r',
 'dered_i d_dered_i',
 'dered_i d_dered_z',
 'dered_i d_dered_ig',
 'dered_i d_dered_zg',
 'dered_i d_dered_rz',
 'dered_i d_dered_iz',
 'dered_z^2',
 'dered_z nObserve',
 'dered_z airmass_u',
 'dered_z class',
 'dered_z d_dered_u',
 'dered_z d_dered_g',
 'dered_z d_dered_r',
 'dered_z d_dered_i',
 'dered_z d_dered_z',
 'dered_z d_dered_ig',
 'dered_z d_dered_zg',
 'dered_z d_dered_rz',
 'dered_z d_dered_iz',
 'nObserve^2',
 'nObserve airmass_u',
 'nObserve class',
 'nObserve d_dered_u',
 'nObserve d_dered_g',
 'nObserve d_dered_r',
 'nObserve d_dered_i',
 'nObserve d_dered_z',
 'nObserve d_dered_ig',
 'nObserve d_dered_zg',
 'nObserve d_dered_rz',
 'nObserve d_dered_iz',
 'airmass_u^2',
 'airmass_u class',
 'airmass_u d_dered_u',
 'airmass_u d_dered_g',
 'airmass_u d_dered_r',
 'airmass_u d_dered_i',
 'airmass_u d_dered_z',
 'airmass_u d_dered_ig',
 'airmass_u d_dered_zg',
 'airmass_u d_dered_rz',
 'airmass_u d_dered_iz',
 'class^2',
 'class d_dered_u',
 'class d_dered_g',
 'class d_dered_r',
 'class d_dered_i',
 'class d_dered_z',
 'class d_dered_ig',
 'class d_dered_zg',
 'class d_dered_rz',
 'class d_dered_iz',
 'd_dered_u^2',
 'd_dered_u d_dered_g',
 'd_dered_u d_dered_r',
 'd_dered_u d_dered_i',
 'd_dered_u d_dered_z',
 'd_dered_u d_dered_ig',
 'd_dered_u d_dered_zg',
 'd_dered_u d_dered_rz',
 'd_dered_u d_dered_iz',
 'd_dered_g^2',
 'd_dered_g d_dered_r',
 'd_dered_g d_dered_i',
 'd_dered_g d_dered_z',
 'd_dered_g d_dered_ig',
 'd_dered_g d_dered_zg',
 'd_dered_g d_dered_rz',
 'd_dered_g d_dered_iz',
 'd_dered_r^2',
 'd_dered_r d_dered_i',
 'd_dered_r d_dered_z',
 'd_dered_r d_dered_ig',
 'd_dered_r d_dered_zg',
 'd_dered_r d_dered_rz',
 'd_dered_r d_dered_iz',
 'd_dered_i^2',
 'd_dered_i d_dered_z',
 'd_dered_i d_dered_ig',
 'd_dered_i d_dered_zg',
 'd_dered_i d_dered_rz',
 'd_dered_i d_dered_iz',
 'd_dered_z^2',
 'd_dered_z d_dered_ig',
 'd_dered_z d_dered_zg',
 'd_dered_z d_dered_rz',
 'd_dered_z d_dered_iz',
 'd_dered_ig^2',
 'd_dered_ig d_dered_zg',
 'd_dered_ig d_dered_rz',
 'd_dered_ig d_dered_iz',
 'd_dered_zg^2',
 'd_dered_zg d_dered_rz',
 'd_dered_zg d_dered_iz',
 'd_dered_rz^2',
 'd_dered_rz d_dered_iz',
 'd_dered_iz^2']

df_poly = pd.DataFrame(data=X, columns=feature_names, index=df.index)
df_poly[target_col] = df[target_col]
df_poly.head()
df_poly.to_csv(feature_file)

del df_poly, df

df = pd.read_csv(feature_file, index_col=0)
print(df.shape)
df.head()

(400000, 210)
1 z redshift dered_u dered_g dered_r dered_i dered_z nObserve airmass_u class d_dered_u d_dered_g d_dered_r d_dered_i d_dered_z d_dered_ig d_dered_zg d_dered_rz d_dered_iz z^2 z redshift z dered_u z dered_g z dered_r z dered_i z dered_z z nObserve z airmass_u z class z d_dered_u z d_dered_g z d_dered_r z d_dered_i z d_dered_z z d_dered_ig z d_dered_zg z d_dered_rz z d_dered_iz redshift^2 redshift dered_u redshift dered_g redshift dered_r redshift dered_i redshift dered_z redshift nObserve redshift airmass_u redshift class redshift d_dered_u redshift d_dered_g ... class d_dered_z class d_dered_ig class d_dered_zg class d_dered_rz class d_dered_iz d_dered_u^2 d_dered_u d_dered_g d_dered_u d_dered_r d_dered_u d_dered_i d_dered_u d_dered_z d_dered_u d_dered_ig d_dered_u d_dered_zg d_dered_u d_dered_rz d_dered_u d_dered_iz d_dered_g^2 d_dered_g d_dered_r d_dered_g d_dered_i d_dered_g d_dered_z d_dered_g d_dered_ig d_dered_g d_dered_zg d_dered_g d_dered_rz d_dered_g d_dered_iz d_dered_r^2 d_dered_r d_dered_i d_dered_r d_dered_z d_dered_r d_dered_ig d_dered_r d_dered_zg d_dered_r d_dered_rz d_dered_r d_dered_iz d_dered_i^2 d_dered_i d_dered_z d_dered_i d_dered_ig d_dered_i d_dered_zg d_dered_i d_dered_rz d_dered_i d_dered_iz d_dered_z^2 d_dered_z d_dered_ig d_dered_z d_dered_zg d_dered_z d_dered_rz d_dered_z d_dered_iz d_dered_ig^2 d_dered_ig d_dered_zg d_dered_ig d_dered_rz d_dered_ig d_dered_iz d_dered_zg^2 d_dered_zg d_dered_rz d_dered_zg d_dered_iz d_dered_rz^2 d_dered_rz d_dered_iz d_dered_iz^2
id
0 1.0 0.0002 -0.1093 1.8170 0.1270 0.9086 0.0208 0.0011 1.6087 0.1200 0.0 0.0063 0.1783 -0.0002 0.0008 -0.0981 -0.0626 0.0367 0.0156 -1.3511 3.4181e-08 -2.0203e-05 0.0003 2.3488e-05 0.0002 3.8455e-06 2.0594e-07 0.0003 2.2191e-05 2.2235e-05 1.1669e-06 3.2960e-05 -3.7490e-08 1.4942e-07 -1.8144e-05 -1.1575e-05 6.7832e-06 2.8794e-06 -0.0002 0.0119 -0.1986 -0.0139 -0.0993 -0.0023 -0.0001 -0.1758 -0.0131 -0.0131 -0.0007 -0.0195 ... -0.0118 -0.0075 0.0044 0.0019 -0.1625 3.9834e-05 1.1252e-03 -1.2798e-06 5.1009e-06 -6.1941e-04 -3.9513e-04 2.3156e-04 9.8297e-05 -0.0085 0.0318 -3.6150e-05 1.4408e-04 -0.0175 -1.1161e-02 0.0065 2.7766e-03 -0.2409 4.1119e-08 -1.6389e-07 1.9901e-05 1.2695e-05 -7.4398e-06 -3.1582e-06 2.7398e-04 6.5318e-07 -7.9317e-05 -5.0598e-05 2.9652e-05 1.2587e-05 -1.0920e-03 9.6316e-03 6.1442e-03 -3.6007e-03 -1.5285e-03 0.1326 3.9195e-03 -2.2970e-03 -9.7505e-04 0.0846 1.3461e-03 5.7142e-04 -0.0496 2.4256e-04 -0.0210 1.8255
1 1.0 -0.0719 -0.1007 -2.4251 -0.2656 -2.8534 -0.1751 -0.1054 -0.8761 0.6658 1.0 0.0081 0.2144 0.0001 0.0007 0.0190 0.0156 -0.0133 -0.0071 0.5650 5.1738e-03 7.2464e-03 0.1744 1.9104e-02 0.2052 1.2597e-02 7.5809e-03 0.0630 -4.7893e-02 -1.4200e-02 -5.8349e-04 -1.5419e-02 -8.7153e-06 -5.1123e-05 -1.3637e-03 -1.1239e-03 9.5460e-04 5.0978e-04 -0.0406 0.0101 0.2443 0.0268 0.2875 0.0176 0.0106 0.0883 -0.0671 -0.0199 -0.0008 -0.0216 ... 0.0037 0.0031 -0.0026 -0.0014 0.1115 6.5806e-05 1.7389e-03 9.8290e-07 5.7656e-06 1.5379e-04 1.2675e-04 -1.0766e-04 -5.7493e-05 0.0046 0.0460 2.5973e-05 1.5236e-04 0.0041 3.3493e-03 -0.0028 -1.5193e-03 0.1211 1.4681e-08 8.6118e-08 2.2971e-06 1.8932e-06 -1.6081e-06 -8.5874e-07 6.8457e-05 5.0517e-07 1.3475e-05 1.1105e-05 -9.4327e-06 -5.0373e-06 4.0156e-04 3.5943e-04 2.9622e-04 -2.5161e-04 -1.3437e-04 0.0107 2.4412e-04 -2.0736e-04 -1.1074e-04 0.0088 1.7613e-04 9.4060e-05 -0.0075 5.0230e-05 -0.0040 0.3192
2 1.0 -0.0302 -0.1082 -1.5717 -0.1600 -1.4768 -0.0822 -0.0444 -0.4286 -1.3179 0.0 0.0013 0.0650 -0.0015 0.0002 0.0444 0.0308 -0.0170 -0.0099 0.4877 9.1298e-04 3.2708e-03 0.0475 4.8338e-03 0.0446 2.4830e-03 1.3416e-03 0.0130 3.9822e-02 -1.9476e-03 -3.9487e-05 -1.9652e-03 4.4024e-05 -5.3424e-06 -1.3413e-03 -9.2915e-04 5.1426e-04 2.9778e-04 -0.0147 0.0117 0.1701 0.0173 0.1599 0.0089 0.0048 0.0464 0.1427 -0.0070 -0.0001 -0.0070 ... 0.0029 0.0020 -0.0011 -0.0006 0.0314 1.7078e-06 8.4993e-05 -1.9040e-06 2.3106e-07 5.8012e-05 4.0186e-05 -2.2242e-05 -1.2879e-05 0.0006 0.0042 -9.4759e-05 1.1499e-05 0.0029 2.0000e-03 -0.0011 -6.4096e-04 0.0317 2.1228e-06 -2.5761e-07 -6.4677e-05 -4.4803e-05 2.4797e-05 1.4359e-05 -7.1063e-04 3.1262e-08 7.8488e-06 5.4370e-06 -3.0093e-06 -1.7425e-06 8.6238e-05 1.9706e-03 1.3651e-03 -7.5553e-04 -4.3748e-04 0.0217 9.4561e-04 -5.2337e-04 -3.0305e-04 0.0150 2.8967e-04 1.6773e-04 -0.0083 9.7124e-05 -0.0048 0.2379
3 1.0 0.0517 -0.1091 2.9763 0.1733 1.6769 0.1172 0.0758 0.1352 0.2544 0.0 -0.0033 -0.0366 -0.0026 -0.0005 -0.0083 -0.0024 -0.0061 -0.0021 0.4242 2.6757e-03 -5.6442e-03 0.1540 8.9650e-03 0.0867 6.0617e-03 3.9204e-03 0.0070 1.3162e-02 -6.1500e-03 -1.7133e-04 -1.8939e-03 -1.3410e-04 -2.6646e-05 -4.2725e-04 -1.2163e-04 -3.1351e-04 -1.0655e-04 0.0219 0.0119 -0.3248 -0.0189 -0.1830 -0.0128 -0.0083 -0.0148 -0.0278 0.0130 0.0004 0.0040 ... 0.0010 0.0003 0.0007 0.0002 -0.0504 1.0971e-05 1.2127e-04 8.5865e-06 1.7062e-06 2.7358e-05 7.7883e-06 2.0075e-05 6.8228e-06 -0.0014 0.0013 9.4914e-05 1.8860e-05 0.0003 8.6090e-05 0.0002 7.5418e-05 -0.0155 6.7204e-06 1.3354e-06 2.1412e-05 6.0957e-06 1.5712e-05 5.3400e-06 -1.0997e-03 2.6535e-07 4.2547e-06 1.2112e-06 3.1221e-06 1.0611e-06 -2.1851e-04 6.8222e-05 1.9422e-05 5.0061e-05 1.7014e-05 -0.0035 5.5290e-06 1.4251e-05 4.8436e-06 -0.0010 3.6734e-05 1.2485e-05 -0.0026 4.2431e-06 -0.0009 0.1799
4 1.0 0.0201 -0.1092 2.1535 0.1381 0.8461 0.0457 0.0284 1.2717 0.1560 0.0 -0.0093 -0.1596 -0.0039 -0.0013 -0.0741 -0.0376 0.0068 0.0007 -0.6113 4.0534e-04 -2.1983e-03 0.0434 2.7804e-03 0.0170 9.1963e-04 5.7181e-04 0.0256 3.1404e-03 -1.3188e-02 -1.8663e-04 -3.2124e-03 -7.9125e-05 -2.5252e-05 -1.4909e-03 -7.5793e-04 1.3661e-04 1.3497e-05 -0.0123 0.0119 -0.2351 -0.0151 -0.0924 -0.0050 -0.0031 -0.1389 -0.0170 0.0715 0.0010 0.0174 ... 0.0485 0.0247 -0.0044 -0.0004 0.4004 8.5934e-05 1.4791e-03 3.6433e-05 1.1627e-05 6.8645e-04 3.4898e-04 -6.2902e-05 -6.2147e-06 0.0057 0.0255 6.2708e-04 2.0013e-04 0.0118 6.0068e-03 -0.0011 -1.0697e-04 0.0975 1.5446e-05 4.9295e-06 2.9103e-04 1.4796e-04 -2.6668e-05 -2.6348e-06 2.4024e-03 1.5732e-06 9.2880e-05 4.7219e-05 -8.5109e-06 -8.4088e-07 7.6673e-04 5.4834e-03 2.7877e-03 -5.0247e-04 -4.9644e-05 0.0453 1.4173e-03 -2.5545e-04 -2.5238e-05 0.0230 4.6043e-05 4.5490e-06 -0.0041 4.4944e-07 -0.0004 0.3737

5 rows × 210 columns

y = df[target_col].values[:320000]
df.drop(target_col, axis=1, inplace=True)
trn = df.iloc[:320000].values
tst = df.iloc[320000:].values
feature_name = df.columns.tolist()
print(y.shape, trn.shape, tst.shape)

(320000,) (320000, 209) (80000, 209)

Stratified K-Fold Cross Validation

cv = StratifiedKFold(n_splits=n_fold, shuffle=True, random_state=seed)

Keras Model

Sequential API

def get_model():
    model = Sequential()

    model.add(Dense(units=64, activation='relu'))
    model.add(Dense(units=n_class, activation='softmax'))
    model.compile(loss='categorical_crossentropy',
                  optimizer='adam')
    return model

Functional API

def get_model():
    inputs = Input(shape=(len(feature_name),))
    x = Dense(64, activation='relu')(inputs)
    outputs = Dense(n_class, activation='softmax')(x)
    model = Model(inputs=inputs, outputs=outputs)
    model.compile(loss='categorical_crossentropy',
                  optimizer='adam')
    return model

신경망 모델 학습

p_val = np.zeros((trn.shape[0], n_class))
p_tst = np.zeros((tst.shape[0], n_class))
for i, (i_trn, i_val) in enumerate(cv.split(trn, y), 1):
    print(f'training model for CV #{i}')
    clf = get_model()
    
    es = EarlyStopping(monitor='val_loss', min_delta=0.001, patience=5,
                       verbose=1, mode='min', baseline=None, restore_best_weights=True)

    rlr = ReduceLROnPlateau(monitor='val_loss', factor=0.5,
                            patience=3, min_lr=1e-6, mode='min', verbose=1)

    clf.fit(trn[i_trn], 
            to_categorical(y[i_trn]),
            validation_data=(trn[i_val], to_categorical(y[i_val])),
            epochs=10,
            callbacks=[es, rlr])
    p_val[i_val, :] = clf.predict(trn[i_val])
    p_tst += clf.predict(tst) / n_fold

training model for CV #1
Epoch 1/10
8000/8000 [==============================] - 6s 691us/step - loss: 11.6517 - val_loss: 0.3027
Epoch 2/10
8000/8000 [==============================] - 5s 603us/step - loss: 1.4335 - val_loss: 0.2639
Epoch 3/10
8000/8000 [==============================] - 5s 604us/step - loss: 0.2691 - val_loss: 0.2678
Epoch 4/10
8000/8000 [==============================] - 5s 621us/step - loss: 1.4684 - val_loss: 0.2482
Epoch 5/10
8000/8000 [==============================] - 5s 594us/step - loss: 0.4254 - val_loss: 0.2944
Epoch 6/10
8000/8000 [==============================] - 5s 643us/step - loss: 0.2451 - val_loss: 0.2431
Epoch 7/10
8000/8000 [==============================] - 5s 674us/step - loss: 2.2416 - val_loss: 0.2343
Epoch 8/10
8000/8000 [==============================] - 5s 653us/step - loss: 0.2418 - val_loss: 0.2345
Epoch 9/10
8000/8000 [==============================] - 5s 631us/step - loss: 0.2449 - val_loss: 0.2294
Epoch 10/10
8000/8000 [==============================] - 5s 641us/step - loss: 4.3611 - val_loss: 0.2274
training model for CV #2
Epoch 1/10
8000/8000 [==============================] - 5s 650us/step - loss: 26.2863 - val_loss: 0.2878
Epoch 2/10
8000/8000 [==============================] - 5s 640us/step - loss: 7.0914 - val_loss: 5.9831
Epoch 3/10
8000/8000 [==============================] - 5s 680us/step - loss: 0.2641 - val_loss: 9.2831
Epoch 4/10
7941/8000 [============================>.] - ETA: 0s - loss: 0.7149
Epoch 00004: ReduceLROnPlateau reducing learning rate to 0.0005000000237487257.
8000/8000 [==============================] - 5s 658us/step - loss: 0.7113 - val_loss: 14.3258
Epoch 5/10
8000/8000 [==============================] - 5s 659us/step - loss: 0.2424 - val_loss: 17.0050
Epoch 6/10
7960/8000 [============================>.] - ETA: 0s - loss: 0.2377Restoring model weights from the end of the best epoch.
8000/8000 [==============================] - 5s 623us/step - loss: 0.2376 - val_loss: 18.2475
Epoch 00006: early stopping
training model for CV #3
Epoch 1/10
8000/8000 [==============================] - 5s 670us/step - loss: 16.6418 - val_loss: 0.2737
Epoch 2/10
8000/8000 [==============================] - 5s 637us/step - loss: 4.4442 - val_loss: 0.2609
Epoch 3/10
8000/8000 [==============================] - 5s 597us/step - loss: 0.3640 - val_loss: 0.2484
Epoch 4/10
8000/8000 [==============================] - 5s 618us/step - loss: 0.2563 - val_loss: 0.2314
Epoch 5/10
8000/8000 [==============================] - 5s 596us/step - loss: 0.2565 - val_loss: 0.2314
Epoch 6/10
8000/8000 [==============================] - 5s 597us/step - loss: 0.4754 - val_loss: 0.2374
Epoch 7/10
7959/8000 [============================>.] - ETA: 0s - loss: 0.8006
Epoch 00007: ReduceLROnPlateau reducing learning rate to 0.0005000000237487257.
8000/8000 [==============================] - 5s 596us/step - loss: 0.7975 - val_loss: 0.2322
Epoch 8/10
8000/8000 [==============================] - 5s 598us/step - loss: 0.2830 - val_loss: 0.2204
Epoch 9/10
8000/8000 [==============================] - 5s 620us/step - loss: 0.2286 - val_loss: 0.2141
Epoch 10/10
8000/8000 [==============================] - 5s 677us/step - loss: 0.4495 - val_loss: 0.2163
training model for CV #4
Epoch 1/10
8000/8000 [==============================] - 6s 720us/step - loss: 12.5209 - val_loss: 104.1083
Epoch 2/10
8000/8000 [==============================] - 5s 600us/step - loss: 7.3628 - val_loss: 47.8241
Epoch 3/10
8000/8000 [==============================] - 5s 588us/step - loss: 3.8502 - val_loss: 32.1715
Epoch 4/10
8000/8000 [==============================] - 5s 645us/step - loss: 1.4737 - val_loss: 21.5491
Epoch 5/10
8000/8000 [==============================] - 5s 646us/step - loss: 1.8249 - val_loss: 23.6221
Epoch 6/10
8000/8000 [==============================] - 5s 629us/step - loss: 0.2341 - val_loss: 23.6290
Epoch 7/10
8000/8000 [==============================] - 5s 685us/step - loss: 0.2341 - val_loss: 21.0535
Epoch 8/10
8000/8000 [==============================] - 5s 663us/step - loss: 0.2381 - val_loss: 24.0096
Epoch 9/10
8000/8000 [==============================] - 5s 633us/step - loss: 0.2232 - val_loss: 25.9969
Epoch 10/10
7918/8000 [============================>.] - ETA: 0s - loss: 0.2201
Epoch 00010: ReduceLROnPlateau reducing learning rate to 0.0005000000237487257.
8000/8000 [==============================] - 5s 641us/step - loss: 0.2200 - val_loss: 26.4051
training model for CV #5
Epoch 1/10
8000/8000 [==============================] - 5s 630us/step - loss: 28.4105 - val_loss: 9.3231
Epoch 2/10
8000/8000 [==============================] - 5s 669us/step - loss: 9.5780 - val_loss: 4.7251
Epoch 3/10
8000/8000 [==============================] - 5s 654us/step - loss: 5.6948 - val_loss: 0.2427
Epoch 4/10
8000/8000 [==============================] - 5s 671us/step - loss: 0.2686 - val_loss: 0.2458
Epoch 5/10
8000/8000 [==============================] - 5s 636us/step - loss: 1.5572 - val_loss: 0.2441
Epoch 6/10
8000/8000 [==============================] - 5s 659us/step - loss: 0.6911 - val_loss: 0.2274
Epoch 7/10
8000/8000 [==============================] - 5s 627us/step - loss: 0.2479 - val_loss: 0.2302
Epoch 8/10
8000/8000 [==============================] - 5s 664us/step - loss: 1.1990 - val_loss: 0.2404
Epoch 9/10
7989/8000 [============================>.] - ETA: 0s - loss: 0.2406
Epoch 00009: ReduceLROnPlateau reducing learning rate to 0.0005000000237487257.
8000/8000 [==============================] - 5s 648us/step - loss: 0.2405 - val_loss: 0.2336
Epoch 10/10
8000/8000 [==============================] - 5s 634us/step - loss: 0.2212 - val_loss: 0.2235

print(f'{accuracy_score(y, np.argmax(p_val, axis=1)) * 100:.4f}%')

91.5106%

print(p_val.shape, p_tst.shape)

(320000, 3) (80000, 3)

np.savetxt(p_val_file, p_val, fmt='%.6f', delimiter=',')
np.savetxt(p_tst_file, p_tst, fmt='%.6f', delimiter=',')

신경망 모델 시각화

clf.summary()

Model: "functional_9"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_5 (InputLayer)         [(None, 209)]             0         
_________________________________________________________________
dense_10 (Dense)             (None, 64)                13440     
_________________________________________________________________
dense_11 (Dense)             (None, 3)                 195       
=================================================================
Total params: 13,635
Trainable params: 13,635
Non-trainable params: 0
_________________________________________________________________

plot_model(clf)

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
<ipython-input-25-e41951ebbf56> in <module>
----> 1 plot_model(clf)

NameError: name 'plot_model' is not defined

제출 파일 생성

sub = pd.read_csv(sample_file, index_col=0)
print(sub.shape)
sub.head()

sub[target_col] = np.argmax(p_tst, axis=1)
sub.head()

sub[target_col].value_counts()

sub.to_csv(sub_file)
12. 신경망, Keras 13. 문자열변수 가공