[PYTHON] Kaggle House Prices ② ~ Model Creation ~

Create a model using the features created below. Kaggle House Prices ① ~ Feature Engineering ~

Loading the library

import numpy as np
import pandas as pd
import lightgbm as lgb
from sklearn.model_selection import KFold
from sklearn.metrics import mean_squared_error
from sklearn.externals import joblib

Data reading

def load_x_train() -> pd.DataFrame:
    """Read the features of the training data created in advance

    :return:Features of training data
    """
    return joblib.load('train_x.pkl')

def load_y_train() -> pd.Series:
    """Read the objective variable of the training data created in advance

    :return:Objective variable of training data
    """
    #Read the objective variable
    train_y = joblib.load('train_y.pkl')
    #Logarithmic conversion of the objective variable
    train_y = np.log1p(train_y)
    return train_y

Cross-validation

def load_index_fold(i_fold: int) -> np.array:
    """Returns the index of the corresponding record with fold in cross-validation

    :param i_fold:fold number
    :return:Index of records corresponding to fold
    """
    #Returns an index that separates training data and validation data
    #Here, the random number is fixed and created every time, but there is also a method to save it in a file.
    train_y = load_y_train()
    kf = KFold(n_splits=4, random_state=6, shuffle=True)
    return list(kf.split(train_y))[i_fold]

def train_fold(i_fold):
    """Perform learning / evaluation by specifying fold for cross-validation

In addition to calling from other methods, it is also used for confirmation and parameter adjustment by itself.

    :param i_fold:fold number
    :return:Tuple (model instance, record index, predicted value, score by evaluation)
    """
    #Reading training data
    train_x = load_x_train()
    print(train_x.shape)
    train_y = load_y_train()

    #Set training data and validation data
    tr_idx, va_idx = load_index_fold(i_fold)
    print(tr_idx.shape)
    print(va_idx.shape)
    tr_x, tr_y = train_x.iloc[tr_idx], train_y.iloc[tr_idx]
    va_x, va_y = train_x.iloc[va_idx], train_y.iloc[va_idx]

    #Do learning
    params_lgbm = {
        "boosting_type": "gbdt",
        "objective": "regression",
        "metric": "rmse",
        "learning_rate": 0.05,
        "max_depth": 4,
        "colsample_bytree": 0.9,
        "subsample": 0.9,
        "reg_alpha": 0.1,
        "reg_lambda": 0.0,
        "min_child_weight": 1,
        "num_leaves": 31
    }
    lgb_train = lgb.Dataset(tr_x, tr_y)
    lgb_eval = lgb.Dataset(va_x, va_y, reference=lgb_train)

    model = lgb.train(
        params_lgbm, lgb_train,
        #Pass the model evaluation data
        valid_sets=lgb_eval,
        #Learn up to 1000 rounds
        num_boost_round=1000,
        #Stop learning if performance does not improve after 10 rounds
        early_stopping_rounds=10
    )

    #Predict and evaluate validation data
    va_pred = model.predict(va_x)
    score = np.sqrt(mean_squared_error(va_y, va_pred))

    #Returns model, index, forecast, rating
    return model, va_idx, va_pred, score

Perform learning and modeling

#Perform cross-validation learning and evaluation
scores = []
va_idxes = []
preds = []
n_fold = 4

#Learn in each fold
for i_fold in range(n_fold):
    #Do learning
    print(f'fold {i_fold} - start training')
    model, va_idx, va_pred, score = train_fold(i_fold)
    print(f'fold {i_fold} - end training - score {score}')

    #Save the model
    # model.save_model()
    joblib.dump(model, f'model-{i_fold}.pkl')

    #Hold the result
    va_idxes.append(va_idx)
    scores.append(score)
    preds.append(va_pred)

#Summarize the results of each fold
va_idxes = np.concatenate(va_idxes)
order = np.argsort(va_idxes)
preds = np.concatenate(preds, axis=0)
preds = preds[order]

print(f'end training cv - score {np.mean(scores)}')

#Saving prediction results
joblib.dump(preds, 'pred-train.pkl')

#Evaluation results
print('result_scores', scores)