我就廢話不多說了，大家還是直接看程式碼吧~

import keras
from sklearn.model_selection import train_test_split
import tensorflow as tf
from keras.callbacks import ModelCheckpoint,Callback
# import keras.backend as K
from keras.layers import *
from keras.models import Model
from keras.optimizers import SGD,RMSprop,Adagrad,Adam
from keras.models import *
from keras.metrics import *
from keras import backend as K
from keras.regularizers import *
from keras.metrics import categorical_accuracy
# from keras.regularizers import activity_l1 #通過L1正則項，使得輸出更加稀疏
from keras_contrib.layers import CRF

from visual_callbacks import AccLossPlotter
plotter = AccLossPlotter(graphs=['acc','loss'],save_graph=True,save_graph_path=sys.path[0])

# from crf import CRFLayer,create_custom_objects

class LossHistory(Callback):
  def on_train_begin(self,logs={}):
    self.losses = []

  def on_batch_end(self,batch,logs={}):
    self.losses.append(logs.get('loss'))
# def on_epoch_end(self,epoch,logs=None):

word_input = Input(shape=(max_len,),dtype='int32',name='word_input')
word_emb = Embedding(len(char_value_dict)+2,output_dim=64,input_length=max_len,dropout=0.2,name='word_emb')(word_input)
bilstm = Bidirectional(LSTM(32,dropout_W=0.1,dropout_U=0.1,return_sequences=True))(word_emb)
bilstm_d = Dropout(0.1)(bilstm)
half_window_size = 2
paddinglayer = ZeroPadding1D(padding=half_window_size)(word_emb)
conv = Conv1D(nb_filter=50,filter_length=(2 * half_window_size + 1),border_mode='valid')(paddinglayer)
conv_d = Dropout(0.1)(conv)
dense_conv = TimeDistributed(Dense(50))(conv_d)
rnn_cnn_merge = merge([bilstm_d,dense_conv],mode='concat',concat_axis=2)
dense = TimeDistributed(Dense(class_label_count))(rnn_cnn_merge)
crf = CRF(class_label_count,sparse_target=False)
crf_output = crf(dense)
model = Model(input=[word_input],output=[crf_output])
model.compile(loss=crf.loss_function,optimizer='adam',metrics=[crf.accuracy])
model.summary()

# serialize model to JSON
model_json = model.to_json()
with open("model.json","w") as json_file:
  json_file.write(model_json)

#編譯模型
# model.compile(loss='categorical_crossentropy',optimizer=adam,metrics=['acc',])

# 用於儲存驗證集誤差最小的引數，當驗證集誤差減少時，立馬儲存下來
checkpointer = ModelCheckpoint(filepath="bilstm_1102_k205_tf130.w",verbose=0,save_best_only=True,save_weights_only=True) #save_weights_only=True
history = LossHistory()

history = model.fit(x_train,y_train,batch_size=32,epochs=500,#validation_data = ([x_test,seq_lens_test],y_test),callbacks=[checkpointer,history,plotter],verbose=1,validation_split=0.1,)

補充知識：keras訓練模型使用自定義CTC損失函式，過載模型時報錯解決辦法

使用keras訓練模型，用到了ctc損失函式，需要自定義損失函式如下：

self.ctc_model.compile(loss={'ctc': lambda y_true,output: output},optimizer=opt)

其中loss為自定義函式，使用字典{‘ctc': lambda y_true,output: output}

訓練完模型後需要過載模型，如下：

from keras.models import load_model

model=load_model('final_ctc_model.h5')

報錯：

Unknown loss function : <lambda>

由於是自定義的損失函式需要加引數custom_objects，這裡需要定義字典{'': lambda y_true,output: output}，正確程式碼如下：

model=load_model('final_ctc_model.h5',custom_objects={'<lambda>': lambda y_true,output: output})

可能是因為要將自己定義的loss函式加入到keras函式裡

在這之前試了很多次，如果用lambda y_true,output: output定義loss

函式字典名只能是'<lambda>'，不能是別的字元

如果自定義一個函式如loss_func作為loss函式如：

self.ctc_model.compile(loss=loss_func,optimizer=opt)

可以在過載時使用

am=load_model('final_ctc_model.h5',custom_objects={'loss_func': loss_func})

此時注意字典名和函式名要相同

以上這篇使用keras實現BiLSTM+CNN+CRF文字標記NER就是小編分享給大家的全部內容了，希望能給大家一個參考，也希望大家多多支援我們。