Gushwell's Dev Notes

   Tweet

79. 適合率-再現率グラフの描画
ロジスティック回帰モデルの分類の閾値を変化させることで，適合率-再現率グラフを描画せよ．

precision, recall, threshold = precision_recall_curve(y_test, pred_positive)

np.argmin(np.abs(thresholds - (i * 0.05)))

import re
from nltk import stem
import numpy as np
from sklearn.metrics import precision_recall_curve
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.externals import joblib
from sklearn.model_selection import train_test_split
import matplotlib.pyplot as plt

class Stopwords:
    words = [ \
        'a', 'about', 'all', 'an', 'and', 'any', 'are', 'as', \
        'at', 'be', 'been', 'but', 'by', 'can', 'could', 'do', \
        'does', 'for', 'from', 'has', 'have', 'he', 'her', 'his', \
        'how', 'i', 'if', 'in', 'into', 'is', 'it', 'its', 'made', \
        'make', 'may', 'me', 'my', 'no', 'not', 'of', 'on', 'one', \
        'or', 'out', 'she', 'should', 'so', 'some', 'than', 'that', \
        'the', 'their', 'them', 'there', 'then', 'they', 'this', \
        'those', 'to', 'too', 'us', 'was', 'we', 'what', 'when',\
        'which', 'who', 'with', 'would', 'you', 'your', ''
        ]

    @staticmethod
    def exists(word):
        return word in  Stopwords.words

class SentimentFeatures:
    def __init__(self):
        self.stemmer = stem.PorterStemmer()
        self.validreg = re.compile(r'^[-=!@#$%^&*()_+|;";,.<>/?]+$')
        self.splitreg = re.compile(r'\s|,|\.|\(|\)|\'|/|\'|\[|\]|-')

    def isValid(self, word):
        if word == '' or len(word) <= 2:
            return False
        if self.validreg.match(word):
            return False
        return not Stopwords.exists(word)

    def getFromLine(self, line):
        array = self.splitreg.split(line)
        # こういう時はlambda キーワードいらないんですね。
        words = filter(self.isValid, array)
        xs = map(self.stemmer.stem, words)
        return xs

    def enumerate(self, filename, encoding):
        with open(filename, 'r', encoding=encoding) as fin:
            for line in fin:
                sentiment = line[:3]
                yield sentiment, self.getFromLine(line[3:])

class SentimentAnalyser:
    def __init__(self):
        self.cv = CountVectorizer(encoding='utf-8')
        self.lr = LogisticRegression(solver='sag', max_iter=10000)

    # LogisticRegression を使い学習する
    def fit(self, X_train, y_train):
        X_train_cv = self.cv.fit_transform(X_train)
        self.lr.fit(X_train_cv, y_train)

    # LogisticRegression を使い予測する
    def predict(self, X_test):
        x = self.cv.transform(X_test)
        return self.lr.predict(x)

    # 予測し、分類毎に確率を得る
    def predict_proba(self, X_test):
        x = self.cv.transform(X_test)
        return self.lr.predict_proba(x)

    # 学習済みデータをロードする
    def load(self):
        self.cv = joblib.load('chapter08/cv73.learn')
        self.lr = joblib.load('chapter08/lr73.learn')

    # 学習済みデータを保存する
    def save(self):
        # 学習したデータを保存する
        joblib.dump(self.cv, 'chapter08/cv73.learn')
        joblib.dump(self.lr, 'chapter08/lr73.learn')

    # 学習に利用するデータを取り出す
    # y[] は、センチメント
    # X[] は、素性データ
    @staticmethod
    def getFeatureData(filename):
        X = []
        y = []
        sf = SentimentFeatures()
        for sentiment, features in sf.enumerate(filename, 'utf-8'):
            y.append(1.0 if sentiment[0] == '+' else 0.0)
            X.append(' '.join(features))
        return X, y

# 参考URL http://ohke.hateblo.jp/entry/2017/08/25/230000

def main():
    sa = SentimentAnalyser()
    X, y = sa.getFeatureData('chapter08/sentiment.txt')
    # 5分割交差検定を行うためにデータを分割する
    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

    # ベクターに変換
    #X_train_cv = sa.trainToVector(X_train)
    #X_test_cv = sa.testToVector(X_test)

    # 学習
    sa.fit(X_train, y_train)

    # 予測
    pp = sa.predict_proba(X_test)
    print(pp[:, 1])
    print()
    print(pp[:5, 1])

    # +1の予測確率を取り出す  [:, 1] は、1列目のすべてのデータ
    pred_positive = sa.predict_proba(X_test)[:, 1]

    # ある閾値の時の適合率、再現率, 閾値の値を取得
    precisions, recalls, thresholds = precision_recall_curve(y_test, pred_positive)

    # 0から1まで0.05刻みで○をプロット
    for i in range(21):
        close_point = np.argmin(np.abs(thresholds - (i * 0.05)))
        plt.plot(precisions[close_point], recalls[close_point], 'o')

    # 適合率-再現率曲線
    plt.plot(precisions, recalls)
    plt.xlabel('Precision')
    plt.ylabel('Recall')

    plt.show()
    input()

if __name__ == '__main__':
    main()

   Tweet

78. 5分割交差検定
76-77の実験では，学習に用いた事例を評価にも用いたため，正当な評価とは言えない．すなわち，分類器が訓練事例を丸暗記する際の性能を評価しており，モデルの汎化性能を測定していない．そこで，5分割交差検定により，極性分類の正解率，適合率，再現率，F1スコアを求めよ．

import re
from nltk import stem
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.externals import joblib
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score

class Stopwords:
    words = [ \
        'a', 'about', 'all', 'an', 'and', 'any', 'are', 'as', \
        'at', 'be', 'been', 'but', 'by', 'can', 'could', 'do', \
        'does', 'for', 'from', 'has', 'have', 'he', 'her', 'his', \
        'how', 'i', 'if', 'in', 'into', 'is', 'it', 'its', 'made', \
        'make', 'may', 'me', 'my', 'no', 'not', 'of', 'on', 'one', \
        'or', 'out', 'she', 'should', 'so', 'some', 'than', 'that', \
        'the', 'their', 'them', 'there', 'then', 'they', 'this', \
        'those', 'to', 'too', 'us', 'was', 'we', 'what', 'when',\
        'which', 'who', 'with', 'would', 'you', 'your', ''
        ]

    @staticmethod
    def exists(word):
        return word in  Stopwords.words

class SentimentFeatures:
    def __init__(self):
        self.stemmer = stem.PorterStemmer()
        self.validreg = re.compile(r'^[-=!@#$%^&*()_+|;";,.<>/?]+$')
        self.splitreg = re.compile(r'\s|,|\.|\(|\)|\'|/|\'|\[|\]|-')

    def isValid(self, word):
        if word == '' or len(word) <= 2:
            return False
        if self.validreg.match(word):
            return False
        return not Stopwords.exists(word)

    def getFromLine(self, line):
        array = self.splitreg.split(line)
        # こういう時はlambda キーワードいらないんですね。
        words = filter(self.isValid, array)
        xs = map(self.stemmer.stem, words)
        return xs

    def enumerate(self, filename, encoding):
        with open(filename, 'r', encoding=encoding) as fin:
            for line in fin:
                sentiment = line[:3]
                yield sentiment, self.getFromLine(line[3:])

class SentimentAnalyser:
    def __init__(self):
        self.cv = CountVectorizer(encoding='utf-8')
        self.lr = LogisticRegression(solver='sag', max_iter=10000)

    # LogisticRegression を使い学習する
    def fit(self, X_train, y_train):
        X_train_cv = self.cv.fit_transform(X_train)
        self.lr.fit(X_train_cv, y_train)

    # LogisticRegression を使い予測する
    def predict(self, X_test):
        x = self.cv.transform(X_test)
        return self.lr.predict(x)

    # 予測し、分類毎に確率を得る
    def predict_proba(self, X_test):
        x = self.cv.transform(X_test)
        return self.lr.predict_proba(x)

    # 学習済みデータをロードする
    def load(self):
        self.cv = joblib.load('chapter08/cv73.learn')
        self.lr = joblib.load('chapter08/lr73.learn')

    # 学習済みデータを保存する
    def save(self):
        # 学習したデータを保存する
        joblib.dump(self.cv, 'chapter08/cv73.learn')
        joblib.dump(self.lr, 'chapter08/lr73.learn')

    # 学習に利用するデータを取り出す
    # y[] は、センチメント
    # X[] は、素性データ
    @staticmethod
    def getFeatureData(filename):
        X = []
        y = []
        sf = SentimentFeatures()
        for sentiment, features in sf.enumerate(filename, 'utf-8'):
            y.append(1.0 if sentiment[0] == '+' else 0.0)
            X.append(' '.join(features))
        return X, y

def main():
    sa = SentimentAnalyser()
    X, y = sa.getFeatureData('chapter08/sentiment.txt')
    # 5分割交差検定を行うためにデータを分割する
    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

    # 4/5のデータで学習する
    # X_train_cv = sa.trainToVector(X_train)
    sa.fit(X_train, y_train)

    # 1/5のデータで予測する
    y_test_pred = sa.predict(X_test)

    print('正解率 accuracy:', accuracy_score(y_test, y_test_pred))
    print('適合率 precision:', precision_score(y_test, y_test_pred))
    print('再現率 recall:', recall_score(y_test, y_test_pred))
    print('F1スコア f1_score:', f1_score(y_test, y_test_pred))

if __name__ == '__main__':
    main()

# 5分割交差検定を行うためにデータを分割する
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

正解率 accuracy: 0.768870135959
適合率 precision: 0.770398481973
再現率 recall: 0.763875823142
F1スコア f1_score: 0.767123287671

   Tweet

77. 正解率の計測
76の出力を受け取り，予測の正解率，正例に関する適合率，再現率，F1スコアを求めるプログラムを作成せよ．

import re
from nltk import stem
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import classification_report, accuracy_score, \
     f1_score, precision_score, recall_score
from sklearn.externals import joblib

class Stopwords:
    words = [ \
        'a', 'about', 'all', 'an', 'and', 'any', 'are', 'as', \
        'at', 'be', 'been', 'but', 'by', 'can', 'could', 'do', \
        'does', 'for', 'from', 'has', 'have', 'he', 'her', 'his', \
        'how', 'i', 'if', 'in', 'into', 'is', 'it', 'its', 'made', \
        'make', 'may', 'me', 'my', 'no', 'not', 'of', 'on', 'one', \
        'or', 'out', 'she', 'should', 'so', 'some', 'than', 'that', \
        'the', 'their', 'them', 'there', 'then', 'they', 'this', \
        'those', 'to', 'too', 'us', 'was', 'we', 'what', 'when',\
        'which', 'who', 'with', 'would', 'you', 'your', ''
        ]

    @staticmethod
    def exists(word):
        return word in  Stopwords.words

class SentimentFeatures:
    def __init__(self):
        self.stemmer = stem.PorterStemmer()
        self.validreg = re.compile(r'^[-=!@#$%^&*()_+|;";,.<>/?]+$')
        self.splitreg = re.compile(r'\s|,|\.|\(|\)|\'|/|\'|\[|\]|-')

    def isValid(self, word):
        if word == '' or len(word) <= 2:
            return False
        if self.validreg.match(word):
            return False
        return not Stopwords.exists(word)

    def getFromLine(self, line):
        array = self.splitreg.split(line)
        # こういう時はlambda キーワードいらないんですね。
        words = filter(self.isValid, array)
        xs = map(self.stemmer.stem, words)
        return xs

    def enumerate(self, filename, encoding):
        with open(filename, 'r', encoding=encoding) as fin:
            for line in fin:
                sentiment = line[:3]
                yield sentiment, self.getFromLine(line[3:])

class SentimentAnalyser:
    def __init__(self):
        self.cv = CountVectorizer(encoding='utf-8')
        self.lr = LogisticRegression(solver='sag', max_iter=10000)

    # LogisticRegression を使い学習する
    def fit(self, X_train, y_train):
        X_train_cv = self.cv.fit_transform(X_train)
        self.lr.fit(X_train_cv, y_train)

    # LogisticRegression を使い予測する
    def predict(self, X_test):
        x = self.cv.transform(X_test)
        return self.lr.predict(x)

    # 予測し、分類毎に確率を得る
    def predict_proba(self, X_test):
        x = self.cv.transform(X_test)
        return self.lr.predict_proba(x)

    # 学習済みデータをロードする
    def load(self):
        self.cv = joblib.load('chapter08/cv73.learn')
        self.lr = joblib.load('chapter08/lr73.learn')

    # 学習済みデータを保存する
    def save(self):
        # 学習したデータを保存する
        joblib.dump(self.cv, 'chapter08/cv73.learn')
        joblib.dump(self.lr, 'chapter08/lr73.learn')

    # 学習に利用するデータを取り出す
    # y[] は、センチメント
    # X[] は、素性データ
    @staticmethod
    def getFeatureData(filename):
        X = []
        y = []
        sf = SentimentFeatures()
        for sentiment, features in sf.enumerate(filename,'utf-8'):
            y.append(1.0 if sentiment[0] == '+' else 0.0)
            X.append(' '.join(features))
        return X, y

def main():
    sa = SentimentAnalyser()
    sa.load()
    X_test, y_test = sa.getFeatureData('chapter08/sentiment.txt')
    y_test_pred = sa.predict(X_test)
    print('正解率 accuracy:', accuracy_score(y_test, y_test_pred))
    print('適合率 precision:', precision_score(y_test, y_test_pred))
    print('再現率 recall:', recall_score(y_test, y_test_pred))
    print('F1スコア f1_score:', f1_score(y_test, y_test_pred))

if __name__ == '__main__':
    main()

正解率 accuracy: 0.94363158882
適合率 precision: 0.946563444109
再現率 recall: 0.940348902645
F1スコア f1_score: 0.943445939588

    print(classification_report(y_test, y_test_pred))

             precision    recall  f1-score   support

        0.0       0.94      0.95      0.94      5331
        1.0       0.95      0.94      0.94      5331

avg / total       0.94      0.94      0.94     10662

   Tweet

76. ラベル付け
学習データに対してロジスティック回帰モデルを適用し，正解のラベル，予測されたラベル，予測確率をタブ区切り形式で出力せよ．

import re
from nltk import stem
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.externals import joblib

class Stopwords:
    words = [ \
        'a', 'about', 'all', 'an', 'and', 'any', 'are', 'as', \
        'at', 'be', 'been', 'but', 'by', 'can', 'could', 'do', \
        'does', 'for', 'from', 'has', 'have', 'he', 'her', 'his', \
        'how', 'i', 'if', 'in', 'into', 'is', 'it', 'its', 'made', \
        'make', 'may', 'me', 'my', 'no', 'not', 'of', 'on', 'one', \
        'or', 'out', 'she', 'should', 'so', 'some', 'than', 'that', \
        'the', 'their', 'them', 'there', 'then', 'they', 'this', \
        'those', 'to', 'too', 'us', 'was', 'we', 'what', 'when',\
        'which', 'who', 'with', 'would', 'you', 'your', ''
        ]

    @staticmethod
    def exists(word):
        return word in  Stopwords.words

class SentimentFeatures:
    def __init__(self):
        self.stemmer = stem.PorterStemmer()
        self.validreg = re.compile(r'^[-=!@#$%^&*()_+|;";,.<>/?]+$')
        self.splitreg = re.compile(r'\s|,|\.|\(|\)|\'|/|\'|\[|\]|-')

    def isValid(self, word):
        if word == '' or len(word) <= 2:
            return False
        if self.validreg.match(word):
            return False
        return not Stopwords.exists(word)

    def getFromLine(self, line):
        array = self.splitreg.split(line)
        # こういう時はlambda キーワードいらないんですね。
        words = filter(self.isValid, array)
        xs = map(self.stemmer.stem, words)
        return xs

    def enumerate(self, filename, encoding):
        with open(filename, 'r', encoding=encoding) as fin:
            for line in fin:
                sentiment = line[:3]
                yield sentiment, self.getFromLine(line[3:])

class SentimentAnalyser:
    def __init__(self):
        self.cv = CountVectorizer(encoding='utf-8')
        self.lr = LogisticRegression(solver='sag', max_iter=10000)

    # LogisticRegression を使い学習する
    def fit(self, X_train, y_train):
        X_train_cv = self.cv.fit_transform(X_train)
        self.lr.fit(X_train_cv, y_train)

    # LogisticRegression を使い予測する
    def predict(self, X_test):
        x = self.cv.transform(X_test)
        return self.lr.predict(x)

    # 予測し、分類毎に確率を得る
    def predict_proba(self, X_test):
        x = self.cv.transform(X_test)
        return self.lr.predict_proba(x)

    # 学習済みデータをロードする
    def load(self):
        self.cv = joblib.load('chapter08/cv73.learn')
        self.lr = joblib.load('chapter08/lr73.learn')

    # 学習済みデータを保存する
    def save(self):
        # 学習したデータを保存する
        joblib.dump(self.cv, 'chapter08/cv73.learn')
        joblib.dump(self.lr, 'chapter08/lr73.learn')

    # 学習に利用するデータを取り出す
    # y[] は、センチメント
    # X[] は、素性データ
    @staticmethod
    def getFeatureData(filename):
        X = []
        y = []
        sf = SentimentFeatures()
        for sentiment, features in sf.enumerate(filename, 'cp1252'):
            y.append(1.0 if sentiment[0] == '+' else 0.0)
            X.append(' '.join(features))
        return X, y

def getLabel(val):
    return '+1' if val > 0.5 else '-1'


def main():
    sa = SentimentAnalyser()
    sa.load()

    X_test, y_test = sa.getFeatureData('chapter08/sentiment.txt')
    y_test_pred = sa.predict(X_test)
    pr = sa.predict_proba(X_test)
    for right, pred, proba in zip(y_test, y_test_pred, pr):
        print('{}\t{}\t{}'.format(getLabel(right), getLabel(pred), \
            proba[0] if pred == 0 else proba[1]))

if __name__ == '__main__':
    main()

-1	-1	0.6097303429740366
-1	-1	0.7237875917615046
+1	+1	0.8452377813768515
+1	+1	0.6449560336941579
+1	+1	0.543156360826781
+1	+1	0.9113562170418062
+1	+1	0.6398913869466663
-1	-1	0.6568220959620661
+1	+1	0.7594696176352425
-1	-1	0.7002498960488786
+1	+1	0.927910364837036
+1	+1	0.9032093714572444
+1	+1	0.714278110906943
+1	+1	0.8908909959384128
+1	+1	0.588722939785566
+1	+1	0.7755451604607824
+1	+1	0.9985328530583656
+1	+1	0.747913188658218
-1	-1	0.6764951558532446
+1	+1	0.9859825558428327

   Tweet

75. 素性の重み
73で学習したロジスティック回帰モデルの中で，重みの高い素性トップ10と，重みの低い素性トップ10を確認せよ．

import re
from nltk import stem
import numpy as np
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.externals import joblib

class Stopwords:
    words = ['a', 'about', 'all', 'an', 'and', 'any', 'are', 'as', \
            'at', 'be', 'been', 'but', 'by', 'can', 'could', 'do', \
            'does', 'for', 'from', 'has', 'have', 'he', 'her', 'his', \
            'how', 'i', 'if', 'in', 'into', 'is', 'it', 'its', 'made', \
            'make', 'may', 'me', 'my', 'no', 'not', 'of', 'on', 'one', \
            'or', 'out', 'she', 'should', 'so', 'some', 'than', 'that', \
            'the', 'their', 'them', 'there', 'then', 'they', 'this', \
            'those', 'to', 'too', 'us', 'was', 'we', 'what', 'when',\
            'which', 'who', 'with', 'would', 'you', 'your', ''
        ]

    @staticmethod
    def exists(word):
        return word in  Stopwords.words

class SentimentFeatures:
    def __init__(self):
        self.stemmer = stem.PorterStemmer()
        self.validreg = re.compile(r'^[-=!@#$%^&*()_+|;";,.<>/?]+$')
        self.splitreg = re.compile(r'\s|,|\.|\(|\)|\'|/|\'|\[|\]|-')

    def isValid(self, word):
        if word == '' or len(word) <= 2:
            return False
        if self.validreg.match(word):
            return False
        return not Stopwords.exists(word)

    def getFromLine(self, line):
        array = self.splitreg.split(line)
        # こういう時はlambda キーワードいらないんですね。
        words = filter(self.isValid, array)
        xs = map(self.stemmer.stem, words)
        return xs

    def enumerate(self, filename, encoding):
        with open(filename, 'r', encoding=encoding) as fin:
            for line in fin:
                sentiment = line[:3]
                yield sentiment, self.getFromLine(line[3:])

class SentimentAnalyser:
    def __init__(self):
        self.cv = CountVectorizer(encoding='utf-8')
        self.lr = LogisticRegression(solver='sag', max_iter=10000)

    # LogisticRegression を使い学習する
    def fit(self, X_train, y_train):
        X_train_cv = self.cv.fit_transform(X_train)
        self.lr.fit(X_train_cv, y_train)

    # LogisticRegression を使い予測する
    def predict(self, X_test):
        x = self.cv.transform(X_test)
        return self.lr.predict(x)

    # 予測し、分類毎に確率を得る
    def predict_proba(self, X_test):
        x = self.cv.transform(X_test)
        return self.lr.predict_proba(x)

    # 学習済みデータをロードする
    def load(self):
        self.cv = joblib.load('chapter08/cv73.learn')
        self.lr = joblib.load('chapter08/lr73.learn')

    # 学習済みデータを保存する
    def save(self):
        # 学習したデータを保存する
        joblib.dump(self.cv, 'chapter08/cv73.learn')
        joblib.dump(self.lr, 'chapter08/lr73.learn')

    # 学習に利用するデータを取り出す
    # y[] は、センチメント
    # X[] は、素性データ
    @staticmethod
    def getFeatureData(filename):
        X = []
        y = []
        sf = SentimentFeatures()
        for sentiment, features in sf.enumerate(filename, 'cp1252'):
            y.append(1.0 if sentiment[0] == '+' else 0.0)
            X.append(' '.join(features))
        return X, y

def main():
    sa = SentimentAnalyser()
    sa.load()

    # fit_transform/transformに渡した単語一覧(学習データ一覧)を得る (重複はなし)
    features = sa.cv.get_feature_names()
    # coef_ には学習した結果の重みが入る。
    # これをソートして、そのインデックスを得る argsortはソートした結果のインデックスが返る
    sorted_idx = np.argsort(sa.lr.coef_)[0]
    print('重みの高い素性トップ10')
    for i in sorted_idx[-1:-11:-1]:
        print(features[i])
    print()
    print('重みの低い素性トップ10')
    for i in sorted_idx[:10]:
        print(features[i])

if __name__ == '__main__':
    main()

重みの高い素性トップ10
engross
refresh
smarter
unexpect
remark
resist
grown
examin
refreshingli
confid

重みの低い素性トップ10
bore
dull
fail
mediocr
plod
routin
wasn
neither
badli
wast