1

我正在尝试在 Keras 中实现这个https://arxiv.org/abs/1603.01354，但是卷积层有问题。我不明白如何设置像 nb_filter ecc 这样的超参数，实际上真正的问题是将最大池输出与嵌入词层输出相匹配。

MAX_CHARACTER_LENGTH 61 MAX_SEQUENCE_LENGTH 124

输入数据

 x_train = pad_sequences(dataset.token_indices['train'],  maxlen=MAX_SEQUENCE_LENGTH)
 x_char_train= np.array(dataset.character_indices['train'])
 x_char_train=x_char_train.reshape((x_char_train.shape[0],  x_char_train.shape[1]*x_char_train.shape[2]))
 y_train = pad_sequences(dataset.label_indices['train'], maxlen=MAX_SEQUENCE_LENGTH)
 x_val = pad_sequences(dataset.token_indices['valid'], maxlen=MAX_SEQUENCE_LENGTH)
 x_char_val= np.array(dataset.character_indices['valid'])
 x_char_val=x_char_val.reshape((x_char_val.shape[0],   x_char_val.shape[1]*x_char_val.shape[2]))
 y_val = pad_sequences(dataset.label_indices['valid'], maxlen=MAX_SEQUENCE_LENGTH)
 y_train=np.expand_dims(y_train,-1)
 y_val=np.expand_dims(y_val,-1)

该模型

EMBEDDING_WORD_DIM= 100
EMBEDDING_CHAR_DIM= 30
N_FILTERS=20
embedding_char= True

class Model(object):
    def __init__(self, embedding_weights, dictonary_size, MAX_SEQUENCE_LENGTH,MAX_CHARACTER_LENGTH, alfabeth_size, tags):
       word_input= Sequential()
       word_input.add(Embedding(dictonary_size + 1,
                                EMBEDDING_WORD_DIM, 
                                weights=[embedding_weights],
                                        input_length=MAX_SEQUENCE_LENGTH,trainable=False))
       if embedding_char:
           character_input=Sequential()
           character_input.add(Embedding(alfabeth_size+1,
                                      EMBEDDING_CHAR_DIM,
                                      input_length=MAX_SEQUENCE_LENGTH*MAX_CHARACTER_LENGTH))
           print(character_input.output_shape)
           character_input.add(Reshape((character_input.output_shape[2],character_input.output_shape[1])))
           print(character_input.output_shape)
           character_input.add(Convolution1D(nb_filter=N_FILTERS, filter_length=3, activation='tanh'))
           print(character_input.output_shape)
           character_input.add(MaxPooling1D(pool_length=N_FILTERS))
           print(character_input.output_shape)
           character_input.add(Reshape((MAX_SEQUENCE_LENGTH,N_FILTERS)))

           self.model=Sequential()
           self.model.add(Merge([word_input,character_input], mode='concat'))
       else:
           self.model=word_input

   self.model.add(Dropout(0.5))
   self.model.add(Bidirectional(LSTM(MAX_SEQUENCE_LENGTH, return_sequences=True)))
   self.model.add(Dropout(0.5))
   self.model.add(TimeDistributed(Dense(tags)))
   crf=ChainCRF()
   self.model.add(crf)
   self.model.compile(loss=crf.loss, optimizer='rmsprop', metrics=['accuracy'])

形状：初始（14041,7564），嵌入后（None,7564,30），reshape1后（None,30,7564），conv1d后（None,28,20），maxpool1d后（None,1,20）

4

0 回答 0