service_mail/classifier.py

from sklearn.base import BaseEstimator, TransformerMixin
from sklearn.feature_extraction import DictVectorizer
from sklearn.feature_extraction.text import TfidfTransformer, CountVectorizer
from sklearn.preprocessing import LabelEncoder
from sklearn.pipeline import Pipeline, FeatureUnion
from sklearn.naive_bayes import MultinomialNB
from sklearn.neural_network import MLPClassifier
from sklearn.model_selection import train_test_split
import numpy as np
import yaml
from storage import MailThread,db_session

with open("data.yml", 'r') as stream:
    try:
        train=yaml.load(stream)
    except yaml.YAMLError as exc:
        print(exc)

data_types= { "answered": bool, "maintopic": str}

def store_training_data(i, d,key=b"answered"):
    global train
    if not data_types.has_key(key):
        raise ValueError("Key "+str(key)+" unknown")
    if not train.has_key(i):
        train[i]={}
    if not key is None and type(train[i]) is dict:
        if not type(d)  is data_types[key]:
            #            print str(type(d)) + " vs " + str(data_types[key])
            raise TypeError("Data - %s - for key "% d +str(key)+" must be " +str(data_types[key])+ " but it is "+ str(type(d)))
        train[i][key]=d
        
    
    with open("data.yml","w") as file:
        file.write(yaml.dump(train,default_flow_style=True))
        file.close()


# Lade Trainingsdaten fuer einen angegebenen key (Label/Eigenschaft) 
def get_training_threads(key="answered"):
    t_a=[]
    d_a=[]
    d_a2=[]
    for i in train:
        t=db_session.query(MailThread).filter(MailThread.firstmail==i).first()
        if not t is None:   # Thread muss in der Datenbank sein
            if train[i].has_key(key): # In den Trainingsdaten muss der relevante Key sein
                t_a.append(t)
                d_a.append(train[i][key])
                le=LabelEncoder()
                d_a2=le.fit_transform(d_a)
    return (t_a,d_a2,le)


def in_training(i, key="answered"):
    return train.has_key(i) and train[i].has_key(key)


def print_answers(l):
    cc=l.classes_
    c_id=l.transform(cc)
    for i,c in enumerate(cc):
        print str(i) + ":  " + str(c)
    return None


class ThreadDictExtractor(BaseEstimator, TransformerMixin):
    def fit(self, x, y=None):
        return self
    def transform(self, X,y=None):
        return [t.mail_flat_dict() for t in X]

class ThreadSubjectExtractor(BaseEstimator, TransformerMixin):
    def fit(self, x, y=None):
        return self
    def transform(self, X,y=None):
        return [t.subject() for t in X]

class ThreadTextExtractor(BaseEstimator, TransformerMixin):
    def fit(self, x, y=None):
        return self
    def transform(self, X,y=None):
        return [t.text() for t in X]





def build_pipe(p=b"pipe1"):

    if p == "pipe1":
        p=Pipeline([('tde', ThreadDictExtractor()),
                    ('dv',DictVectorizer()),
                    ('clf', MultinomialNB())
        ])
    elif p=="pipe2":
        p = Pipeline([
            ('union', FeatureUnion(transformer_list=[
                ('subject', Pipeline([('tse', ThreadSubjectExtractor()),
                                      ('cv',CountVectorizer()),
                                      ('tfidf', TfidfTransformer())
                ])),
                ('text',    Pipeline([('tte',ThreadTextExtractor()),
                                      ('cv',CountVectorizer()),
                                      ('tfidf', TfidfTransformer())
                ])),
                ('envelope', Pipeline([('tde', ThreadDictExtractor()),
                                       ('dv',DictVectorizer())
                ]))
            ], transformer_weights={
                'subject': 1,
                'text': 0.7,
                'envelope': 0.7
            } )),
            ('clf', MultinomialNB())
        ])
    elif p=="pipe2b":
        p = Pipeline([
            ('union', FeatureUnion(transformer_list=[
            ('subject', Pipeline([('tse', ThreadSubjectExtractor()),
                                ('cv',CountVectorizer()),
                                ('tfidf', TfidfTransformer())
        ])),
            ('text',    Pipeline([('tte',ThreadTextExtractor()),
                                  ('cv',CountVectorizer()),
                                ('tfidf', TfidfTransformer())
            ])),
            ('envelope', Pipeline([('tde', ThreadDictExtractor()),
                                   ('dv',DictVectorizer())
            ]))
            ], transformer_weights={
                'subject': 1,
                'text': 0.7,
                'envelope': 0.7
            } )),
            ('mlc', MLPClassifier())
        ])
    elif p=="pipe2c":
        p = Pipeline([
            ('union', FeatureUnion(transformer_list=[
            ('subject', Pipeline([('tse', ThreadSubjectExtractor()),
                                ('cv',CountVectorizer()),
                                ('tfidf', TfidfTransformer())
        ])),
            ('text',    Pipeline([('tte',ThreadTextExtractor()),
                                  ('cv',CountVectorizer()),
                                ('tfidf', TfidfTransformer())
            ])),
            ('envelope', Pipeline([('tde', ThreadDictExtractor()),
                                   ('dv',DictVectorizer())
            ]))
            ], transformer_weights={
                'subject': 1,
                'text': 1,
                'envelope': 0.4
            } )),
            ('mlc', MLPClassifier())
        ])
    else:
        raise ValueError("The pipe %s is not a valid pipe")
    return p

def get_pipe(p=b"pipe1",k=b"answered"):
    p=build_pipe(p)
    tt= get_training_threads(k)
    p.fit(tt[0],tt[1])
    return p,tt[2]

from sklearn.metrics import accuracy_score

def test_pipe(pp,k):
    tt= get_training_threads(k)
    X_train,X_test,y_train,y_test=train_test_split(tt[0],tt[1],test_size=0.2)
    if type(pp) is list:
        for p in pp:
            print "pipe: %s" % p
            p=build_pipe(p)
            p.fit(X_train,y_train)
            ypred=p.predict(X_test)
            print accuracy_score(y_test,ypred)

#pipe1=get_pipe("pipe1", "answered")
#pipe2=get_pipe("pipe2", "maintopic")
#pipe2b=get_pipe("pipe2b", "maintopic")