def __init__(self, n_trees=10):

self._n_trees = n_trees

self._forest = [None] * self._n_trees

self._using_features = [None] * self._n_trees

def fit(self, X, y):

self._classes = np.unique(y)

# 各決定木にわたすデータは、元データをブートストラップサンプル (復元ありの抽出) および特徴量をランダムに選択

bootstrapped_X, bootstrapped_y = self._bootstrap_sample(X, y)

for i, (i_bootstrapped_X, i_bootstrapped_y) in enumerate(zip(bootstrapped_X, bootstrapped_y)):

tree = DecisionTree()

tree.fit(i_bootstrapped_X, i_bootstrapped_y)

self._forest[i] = tree

def _bootstrap_sample(self, X, y):

#与えられたデータをブートストラップサンプル (復元抽出)

# 同時に、特徴量方向のサンプリング

n_features = X.shape[1]

n_features_forest = np.floor(np.sqrt(n_features))

bootstrapped_X = list()

bootstrapped_y = list()

for i in range(self._n_trees):

ind = np.random.choice(len(y), size=len(y)) # 用いるサンプルをランダムに選択

col = np.random.choice(n_features, size=int(n_features_forest), replace=False) # 用いる特徴量をランダムに選択

bootstrapped_X.append(X[np.ix_(ind, col)])

bootstrapped_y.append(y[ind])

self._using_features[i] = col

return bootstrapped_X, bootstrapped_y

def predict(self, X):

proba = self.predict_proba(X)

return self._classes[np.argmax(proba, axis=1)]

def predict_proba(self, X):

if self._forest[0] is None:

raise ValueError('fitしてね')

votes = [tree.predict(X[:, using_feature]) for tree, using_feature in zip(self._forest, self._using_features)] # n_trees x n_samples

counts = [Counter(row) for row in np.array(votes).transpose()] # n_samples だけの Counter オブジェクト

# 各 tree の意見の集計

counts_array = np.zeros((len(X), len(self._classes))) # n_samples x n_classes

for row_index, count in enumerate(counts):

for class_index, class_ in enumerate(self._classes):

counts_array[row_index, class_index] = count[class_]

proba = counts_array / self._n_trees # 規格化する