BIDI_ML_INFO_EXTRACTION/BiddingKG/dl/interface/modelFactory.py

'''
Created on 2019年5月16日

@author: User
'''

import os
import sys
sys.path.append(os.path.abspath("../.."))
from keras import models
from keras import layers
# from keras_contrib.layers import CRF
from keras.preprocessing.sequence import pad_sequences
from keras import optimizers,losses,metrics
from BiddingKG.dl.common.Utils import *
import tensorflow as tf


    
class Model_role_classify():
    def __init__(self,lazyLoad=getLazyLoad()):
        #self.model_role_file = os.path.abspath("../role/models/model_role.model.hdf5")
        self.model_role_file = os.path.dirname(__file__)+"/../role/log/new_biLSTM-ep012-loss0.028-val_loss0.040-f10.954.h5"
        self.model_role = None
        self.graph = tf.get_default_graph()
        if not lazyLoad:
            self.getModel()
        
        
    def getModel(self):
        if self.model_role is None:
            self.model_role = models.load_model(self.model_role_file,custom_objects={'precision':precision,'recall':recall,'f1_score':f1_score})
        return self.model_role
    
    def encode(self,tokens,begin_index,end_index,**kwargs):
        return embedding(spanWindow(tokens=tokens,begin_index=begin_index,end_index=end_index,size=10),shape=(2,10,128))
    
    def predict(self,x):
        x = np.transpose(np.array(x),(1,0,2,3))
        with self.graph.as_default():
            return self.getModel().predict([x[0],x[1]])
    
class Model_role_classify_word():
    def __init__(self,lazyLoad=getLazyLoad()):
        if USE_PAI_EAS:
            lazyLoad = True
        #self.model_role_file = os.path.abspath("../role/log/ep071-loss0.107-val_loss0.122-f10.956.h5")
        self.model_role_file = os.path.dirname(__file__)+"/../role/models/ep038-loss0.140-val_loss0.149-f10.947.h5"
        #self.model_role_file = os.path.abspath("../role/log/textcnn_ep017-loss0.088-val_loss0.125-f10.955.h5")
        self.model_role = None
        
        self.sess_role = tf.Session(graph=tf.Graph())
        if not lazyLoad:
            self.getModel()
        
        
    def getModel(self):
        if self.model_role is None:
          with self.sess_role.as_default() as sess:
            with self.sess_role.graph.as_default():
              meta_graph_def = tf.saved_model.loader.load(sess=self.sess_role, tags=["serve"], export_dir=os.path.dirname(__file__)+"/role_savedmodel")
              signature_key = tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY
              signature_def = meta_graph_def.signature_def
              
              input0 = self.sess_role.graph.get_tensor_by_name(signature_def[signature_key].inputs["input0"].name)
              input1 = self.sess_role.graph.get_tensor_by_name(signature_def[signature_key].inputs["input1"].name)
              input2 = self.sess_role.graph.get_tensor_by_name(signature_def[signature_key].inputs["input2"].name)
              output = self.sess_role.graph.get_tensor_by_name(signature_def[signature_key].outputs["outputs"].name)
              self.model_role = [[input0,input1,input2],output]
        return self.model_role
    '''
    def load_weights(self):
        model = self.getModel()
        model.load_weights(self.model_role_file)
    '''
    
    def encode(self,tokens,begin_index,end_index,entity_text,**kwargs):
        _span = spanWindow(tokens=tokens,begin_index=begin_index,end_index=end_index,size=12,center_include=True,word_flag=True,text=entity_text)
        # print(_span)
        _encode_span = encodeInput(_span, word_len=20, word_flag=True,userFool=False)
        # print(_encode_span)
        return _encode_span
    
    def predict(self,x):
        x = np.transpose(np.array(x),(1,0,2))
        model_role = self.getModel()
        assert len(x)==len(model_role[0])
        feed_dict = {}
        for _x,_t in zip(x,model_role[0]):
          feed_dict[_t] = _x
        list_result = limitRun(self.sess_role,[model_role[1]],feed_dict)[0]
        return list_result
        #return self.sess_role.run(model_role[1],feed_dict=feed_dict)
        
    
class Model_money_classify():
    def __init__(self,lazyLoad=getLazyLoad()):
        if USE_PAI_EAS:
            lazyLoad = True
        self.model_money_file = os.path.dirname(__file__)+"/../money/models/model_money_word.h5"
        self.model_money = None
        self.sess_money = tf.Session(graph=tf.Graph())
        if not lazyLoad:
            self.getModel()
        
        
    def getModel(self):
      if self.model_money is None:
        with self.sess_money.as_default() as sess:
          with sess.graph.as_default():
            meta_graph_def = tf.saved_model.loader.load(sess,tags=["serve"],export_dir=os.path.dirname(__file__)+"/money_savedmodel")
            # meta_graph_def = tf.saved_model.loader.load(sess,tags=["serve"],export_dir=os.path.dirname(__file__)+"/money_savedmodel_bilstmonly")
            signature_key = tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY
            signature_def = meta_graph_def.signature_def
            
            input0 = sess.graph.get_tensor_by_name(signature_def[signature_key].inputs["input0"].name)
            input1 = sess.graph.get_tensor_by_name(signature_def[signature_key].inputs["input1"].name)
            input2 = sess.graph.get_tensor_by_name(signature_def[signature_key].inputs["input2"].name)
            
            output = sess.graph.get_tensor_by_name(signature_def[signature_key].outputs["outputs"].name)
            
            self.model_money = [[input0,input1,input2],output]
      return self.model_money
      '''
      if self.model_money is None:
          self.model_money = models.load_model(self.model_money_file,custom_objects={'precision':precision,'recall':recall,'f1_score':f1_score})
      return self.model_money
      '''
    
    '''
    def load_weights(self):
        model = self.getModel()
        model.load_weights(self.model_money_file)
    '''
    
    def encode(self,tokens,begin_index,end_index,**kwargs):
        _span = spanWindow(tokens=tokens, begin_index=begin_index, end_index=end_index, size=10, center_include=True, word_flag=True)
        # print(_span)
        return encodeInput(_span, word_len=30, word_flag=True,userFool=False)
        return embedding_word(_span,shape=(3,100,60))
    
    def predict(self,x):
        # print("shape",np.shape(x))
        x = np.transpose(np.array(x),(1,0,2))
         
        model_money = self.getModel()
        assert len(x)==len(model_money[0])
        feed_dict = {}
        for _x,_t in zip(x,model_money[0]):
          feed_dict[_t] = _x
        list_result = limitRun(self.sess_money,[model_money[1]],feed_dict)[0]
        #return self.sess_money.run(model_money[1],feed_dict=feed_dict)
        return list_result
        '''
        with self.graph.as_default():
            return self.getModel().predict([x[0],x[1],x[2]])
        '''

from itertools import groupby
from BiddingKG.dl.relation_extraction.model import get_words_matrix
class Model_relation_extraction():
    def __init__(self,lazyLoad=getLazyLoad()):
        if USE_PAI_EAS:
            lazyLoad = True
        self.subject_model_file = os.path.dirname(__file__)+"/../relation_extraction/models2/subject_model"
        self.object_model_file = os.path.dirname(__file__)+"/../relation_extraction/models2/object_model"
        self.model_subject = None
        self.model_object = None
        self.sess_subject = tf.Session(graph=tf.Graph())
        self.sess_object = tf.Session(graph=tf.Graph())
        if not lazyLoad:
            self.getModel1()
            self.getModel2()
        self.entity_type_dict = {
            'org': '<company/org>',
            'company': '<company/org>',
            'location': '<location>',
            'phone': '<phone>',
            'person': '<contact_person>'
        }
        self.id2predicate = {
            0: "rel_person",  # 公司——联系人
            1: "rel_phone",  # 联系人——电话
            2: "rel_address"  # 公司——地址
        }
        self.words_size = 128

    # subject_model
    def getModel1(self):
        if self.model_subject is None:
            with self.sess_subject.as_default() as sess:
                with sess.graph.as_default():
                    meta_graph_def = tf.saved_model.loader.load(sess,tags=["serve"],export_dir=self.subject_model_file)
                    signature_key = tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY
                    signature_def = meta_graph_def.signature_def
                    input0 = sess.graph.get_tensor_by_name(signature_def[signature_key].inputs["input0"].name)
                    input1 = sess.graph.get_tensor_by_name(signature_def[signature_key].inputs["input1"].name)
                    output = sess.graph.get_tensor_by_name(signature_def[signature_key].outputs["outputs"].name)
                    self.model_subject = [[input0,input1],output]
        return self.model_subject
    # object_model
    def getModel2(self):
        if self.model_object is None:
            with self.sess_object.as_default() as sess:
                  with sess.graph.as_default():
                        meta_graph_def = tf.saved_model.loader.load(sess,tags=["serve"],export_dir=self.object_model_file)
                        signature_key = tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY
                        signature_def = meta_graph_def.signature_def
                        input0 = sess.graph.get_tensor_by_name(signature_def[signature_key].inputs["input0"].name)
                        input1 = sess.graph.get_tensor_by_name(signature_def[signature_key].inputs["input1"].name)
                        input2 = sess.graph.get_tensor_by_name(signature_def[signature_key].inputs["input2"].name)
                        output = sess.graph.get_tensor_by_name(signature_def[signature_key].outputs["outputs"].name)
                        self.model_object = [[input0,input1,input2],output]
        return self.model_object

    def encode(self,entity_list,list_sentence):
        list_sentence = sorted(list_sentence, key=lambda x: x.sentence_index)
        entity_list = sorted(entity_list, key=lambda x: (x.sentence_index, x.begin_index))
        pre_data = []
        text_data = []
        last_sentence_index = -1
        for key, group in groupby(entity_list, key=lambda x: x.sentence_index):
            if key - last_sentence_index > 1:
                for i in range(last_sentence_index + 1, key):
                    pre_data.extend(list_sentence[i].tokens)
                    text_data.extend([0] * len(list_sentence[i].tokens))
            group = list(group)
            for i in range(len(group)):
                ent = group[i]
                _tokens = list_sentence[key].tokens
                if i == len(group) - 1:
                    if i == 0:
                        pre_data.extend(_tokens[:ent.begin_index])
                        text_data.extend([0] * len(_tokens[:ent.begin_index]))
                        pre_data.append(self.entity_type_dict[ent.entity_type])
                        text_data.append(ent)
                        pre_data.extend(_tokens[ent.end_index + 1:])
                        text_data.extend([0] * len(_tokens[ent.end_index + 1:]))
                        break
                    else:
                        pre_data.append(self.entity_type_dict[ent.entity_type])
                        text_data.append(ent)
                        pre_data.extend(_tokens[ent.end_index + 1:])
                        text_data.extend([0] * len(_tokens[ent.end_index + 1:]))
                        break
                if i == 0:
                    pre_data.extend(_tokens[:ent.begin_index])
                    text_data.extend([0] * len(_tokens[:ent.begin_index]))
                    pre_data.append(self.entity_type_dict[ent.entity_type])
                    text_data.append(ent)
                    pre_data.extend(_tokens[ent.end_index + 1:group[i + 1].begin_index])
                    text_data.extend([0] * len(_tokens[ent.end_index + 1:group[i + 1].begin_index]))
                else:
                    pre_data.append(self.entity_type_dict[ent.entity_type])
                    text_data.append(ent)
                    pre_data.extend(_tokens[ent.end_index + 1:group[i + 1].begin_index])
                    text_data.extend([0] * len(_tokens[ent.end_index + 1:group[i + 1].begin_index]))
            last_sentence_index = key
        return text_data, pre_data

    def check_data(self, words):
        # 检查数据是否包含可预测的subject和object
        # 没有需要预测的链接属性，直接return
        company_relation = 0
        person_relation = 0
        if '<company/org>' in words:
            company_relation += 1
        if '<contact_person>' in words:
            person_relation += 1
            if company_relation:
                company_relation += 1
        # 暂时不考虑地址location实体
        # if '<location>' in words and company_relation:
        #     company_relation += 1
        if '<phone>' in words and company_relation:
            person_relation += 1
        if company_relation < 2 and person_relation < 2:
            return False
        return True

    def predict(self,text_in, words, rate=0.5):
        # 没有需要预测的链接属性，直接return
        # if self.check_data(words):
        #     return []
        # 使用模型预测
        triple_list = []
        # print("tokens:",words)
        # _t2 = [self.words2id.get(c, 1) for c in words]
        _t2 = np.zeros((len(words), self.words_size))
        for i in range(len(words)):
            _t2[i] = np.array(get_words_matrix(words[i]))
        _t2 = np.array([_t2])
        _t3 = [1 for _ in words]
        _t3 = np.array([_t3])
        # _k1 = self.model_subject.predict([_t2, _t3])
        _k1 = limitRun(self.sess_subject,[self.model_subject[1]],feed_dict={self.model_subject[0][0]:_t2,
                                                                            self.model_subject[0][1]:_t3})[0]
        _k1 = _k1[0, :, 0]
        _k1 = np.where(_k1 > rate)[0]
        # print('k1',_k1)
        _subjects = []
        for i in _k1:
            _subject = text_in[i]
            _subjects.append((_subject, i, i))
        if _subjects:
            _t2 = np.repeat(_t2, len(_subjects), 0)
            _t3 = np.repeat(_t3, len(_subjects), 0)
            _k1, _ = np.array([_s[1:] for _s in _subjects]).T.reshape((2, -1, 1))
            # _o1 = self.model_object.predict([_t2, _t3, _k1])
            _o1 = limitRun(self.sess_object, [self.model_object[1]], feed_dict={self.model_object[0][0]: _t2,
                                                                                  self.model_object[0][1]: _t3,
                                                                                  self.model_object[0][2]: _k1})[0]
            for i, _subject in enumerate(_subjects):
                _oo1 = np.where(_o1[i] > 0.5)
                # print('_oo1', _oo1)
                for _ooo1, _c1 in zip(*_oo1):
                    _object = text_in[_ooo1]
                    _predicate = self.id2predicate[_c1]
                    triple_list.append((_subject[0], _predicate, _object))
            # print([(t[0].entity_text,t[1],t[2].entity_text) for t in triple_list])
            return triple_list
        else:
            return []

    
class Model_person_classify():
    def __init__(self,lazyLoad=getLazyLoad()):
        if USE_PAI_EAS:
            lazyLoad = True
        self.model_person_file = os.path.dirname(__file__)+"/../person/models/model_person.model.hdf5"
        self.model_person = None
        self.sess_person = tf.Session(graph=tf.Graph())
        if not lazyLoad:
            self.getModel()
        
        
    def getModel(self):
      if self.model_person is None:
        with self.sess_person.as_default() as sess:
          with sess.graph.as_default():
            # meta_graph_def = tf.saved_model.loader.load(sess,tags=["serve"],export_dir=os.path.dirname(__file__)+"/person_savedmodel_new")
            meta_graph_def = tf.saved_model.loader.load(sess,tags=["serve"],export_dir=os.path.dirname(__file__)+"/person_savedmodel_new_znj")
            signature_key = tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY
            signature_def = meta_graph_def.signature_def
            
            input0 = sess.graph.get_tensor_by_name(signature_def[signature_key].inputs["input0"].name)
            input1 = sess.graph.get_tensor_by_name(signature_def[signature_key].inputs["input1"].name)
            
            output = sess.graph.get_tensor_by_name(signature_def[signature_key].outputs["outputs"].name)
            
            self.model_person = [[input0,input1],output]
      return self.model_person
            
      
      '''
        if self.model_person is None:
            self.model_person = models.load_model(self.model_person_file,custom_objects={'precision':precision,'recall':recall,'f1_score':f1_score})
        return self.model_person
      '''
    '''
    def load_weights(self):
        model = self.getModel()
        model.load_weights(self.model_person_file)
    '''
    
    def encode(self,tokens,begin_index,end_index,**kwargs):
        # return embedding(spanWindow(tokens=tokens,begin_index=begin_index,end_index=end_index,size=10),shape=(2,10,128))
        return embedding(spanWindow(tokens=tokens,begin_index=begin_index,end_index=end_index,size=20),shape=(2,20,128))

    def predict(self,x):
        x = np.transpose(np.array(x),(1,0,2,3))
        
        model_person = self.getModel()
        assert len(x)==len(model_person[0])
        feed_dict = {}
        for _x,_t in zip(x,model_person[0]):
          feed_dict[_t] = _x

        list_result = limitRun(self.sess_person,[model_person[1]],feed_dict)[0]
        return list_result
        #return self.sess_person.run(model_person[1],feed_dict=feed_dict)
        '''
        with self.graph.as_default():
            return self.getModel().predict([x[0],x[1]])
        '''
    
class Model_form_line():
    def __init__(self,lazyLoad=getLazyLoad()):
        self.model_file = os.path.dirname(__file__)+"/../form/model/model_form.model - 副本.hdf5"
        self.model_form = None
        self.graph = tf.get_default_graph()
        if not lazyLoad:
            self.getModel()
        
        
    def getModel(self):
        if self.model_form is None:
            self.model_form = models.load_model(self.model_file,custom_objects={"precision":precision,"recall":recall,"f1_score":f1_score})
        return self.model_form
    
    def encode(self,data,shape=(100,60),expand=False,**kwargs):
        embedding = np.zeros(shape)
        word_model = getModel_word()
        for i in range(len(data)):
            if i>=shape[0]:
                break
            if data[i] in word_model.vocab:
                embedding[i] = word_model[data[i]]
        if expand:
            embedding = np.expand_dims(embedding,0)
        return embedding
    
    def predict(self,x):
        with self.graph.as_default():
            return self.getModel().predict(x)
    
class Model_form_item():
    def __init__(self,lazyLoad=getLazyLoad()):
        self.model_file = os.path.dirname(__file__)+"/../form/log/ep039-loss0.038-val_loss0.064-f10.9783.h5"
        self.model_form = None
        self.sess_form = tf.Session(graph=tf.Graph())
        if not lazyLoad:
            self.getModel()


    def getModel(self):
      if self.model_form is None:
        with self.sess_form.as_default() as sess:
          with sess.graph.as_default():
            meta_graph_def = tf.saved_model.loader.load(sess,tags=["serve"],export_dir="%s/form_savedmodel"%(os.path.dirname(__file__)))
            signature_key = tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY
            signature_def = meta_graph_def.signature_def

            inputs = sess.graph.get_tensor_by_name(signature_def[signature_key].inputs["inputs"].name)
            output = sess.graph.get_tensor_by_name(signature_def[signature_key].outputs["outputs"].name)

            self.model_form = [[inputs],output]
      return self.model_form
      '''
        if self.model_form is None:
            with self.graph.as_defalt():
                self.model_form = models.load_model(self.model_file,custom_objects={"precision":precision,"recall":recall,"f1_score":f1_score})
        return self.model_form
      '''

    def encode(self,data,**kwargs):
        return encodeInput([data], word_len=50, word_flag=True,userFool=False)[0]
        return encodeInput_form(data)

    def predict(self,x):

      model_form = self.getModel()

      list_result = limitRun(self.sess_form,[model_form[1]],feed_dict={model_form[0][0]:x})[0]
      return list_result
      #return self.sess_form.run(model_form[1],feed_dict={model_form[0][0]:x})



      '''
      with self.graph.as_default():
          return self.getModel().predict(x)
      '''

class Model_form_context():
    def __init__(self,lazyLoad=getLazyLoad()):
        self.model_form = None
        self.sess_form = tf.Session(graph=tf.Graph())
        if not lazyLoad:
            self.getModel()


    def getModel(self):
        if self.model_form is None:
            with self.sess_form.as_default() as sess:
                with sess.graph.as_default():
                    meta_graph_def = tf.saved_model.loader.load(sess,tags=["serve"],export_dir="%s/form_context_savedmodel"%(os.path.dirname(__file__)))
                    signature_key = tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY
                    signature_def = meta_graph_def.signature_def

                    inputs = sess.graph.get_tensor_by_name(signature_def[signature_key].inputs["inputs"].name)
                    output = sess.graph.get_tensor_by_name(signature_def[signature_key].outputs["outputs"].name)

                    self.model_form = [[inputs],output]
        return self.model_form
        '''
          if self.model_form is None:
              with self.graph.as_defalt():
                  self.model_form = models.load_model(self.model_file,custom_objects={"precision":precision,"recall":recall,"f1_score":f1_score})
          return self.model_form
        '''

    def encode_table(self,inner_table,size=30):
        def encode_item(_table,i,j):
            _x = [_table[j-1][i-1],_table[j-1][i],_table[j-1][i+1],
                  _table[j][i-1],_table[j][i],_table[j][i+1],
                  _table[j+1][i-1],_table[j+1][i],_table[j+1][i+1]]
            e_x = [encodeInput_form(_temp[0],MAX_LEN=30) for _temp in _x]
            _label = _table[j][i][1]
            # print(_x)
            # print(_x[4],_label)
            return e_x,_label,_x

        def copytable(inner_table):
            table = []
            for line in inner_table:
                list_line = []
                for item in line:
                    list_line.append([item[0][:size],item[1]])
                table.append(list_line)
            return table

        table = copytable(inner_table)
        padding = ["#"*30,0]
        width = len(table[0])
        height = len(table)
        table.insert(0,[padding for i in range(width)])
        table.append([padding for i in range(width)])
        for item in table:
            item.insert(0,padding.copy())
            item.append(padding.copy())
        data_x = []
        data_y = []
        data_text = []
        data_position = []
        for _i in range(1,width+1):
            for _j in range(1,height+1):
                _x,_y,_text = encode_item(table,_i,_j)
                data_x.append(_x)
                _label = [0,0]
                _label[_y] = 1
                data_y.append(_label)
                data_text.append(_text)
                data_position.append([_i-1,_j-1])
                # input = table[_j][_i][0]
                # item_y = [0,0]
                # item_y[table[_j][_i][1]] = 1
                # data_x.append(encodeInput([input], word_len=50, word_flag=True,userFool=False)[0])
                # data_y.append(item_y)
        return data_x,data_y,data_text,data_position

    def encode(self,inner_table,**kwargs):
        data_x,_,_,data_position = self.encode_table(inner_table)
        return data_x,data_position

    def predict(self,x):

        model_form = self.getModel()

        list_result = limitRun(self.sess_form,[model_form[1]],feed_dict={model_form[0][0]:x})[0]
        return list_result

# class Model_form_item():
#     def __init__(self,lazyLoad=False):
#         self.model_file = os.path.dirname(__file__)+"/ep039-loss0.038-val_loss0.064-f10.9783.h5"
#         self.model_form = None
#
#         if not lazyLoad:
#             self.getModel()
#         self.graph = tf.get_default_graph()
#
#     def getModel(self):
#         if self.model_form is None:
#             self.model_form = models.load_model(self.model_file,custom_objects={"precision":precision,"recall":recall,"f1_score":f1_score})
#         return self.model_form
#
#     def encode(self,data,**kwargs):
#
#         return encodeInput_form(data)
#
#     def predict(self,x):
#         with self.graph.as_default():
#             return self.getModel().predict(x)