0

我正在构建一个模型，其中 LSTM 层堆叠在 CNN 层上。此外，在这里我正在执行多通道 CNN，在每个通道之上，我正在应用 LSTM 层，然后将结果连接起来。

这里的问题是，在 CNN 上实现 TimeDistributed 层并将嵌入层（形状 - (None,300(seq.len),300(emb. dim.)) 传递给它向我展示了输入形状的值错误即expected dim = 3 , but passed dim =2. Input shape =(None,300-->(emb. dim.))。但是当我检查嵌入维度时，它们是正确的，即(None,300,300)。

sequence_length = 300
embed_dim = 300

class hyper():
  def __init__(self,embedding_dim,filter_sizes,num_filters,dropout_prob,hidden_dims,batch_size,num_epochs):
    # Model Hyperparameters
    self.embedding_dim = embedding_dim
    self.filter_sizes = filter_sizes
    self.num_filters = num_filters
    self.dropout_prob = dropout_prob
    self.hidden_dims = hidden_dims
    # Training parameters
    self.batch_size = batch_size
    self.num_epochs = num_epochs

class prep_hyper():
  def __init__(self,sequenceLength,max_words):
    # Prepossessing parameters
    self.sequenceLength = sequenceLength
    self.max_words = max_words
    
m_hyper=hyper(embedding_dim=embed_dim,filter_sizes=(3,4,5,6,8),num_filters=128,dropout_prob=(0.2,0.5),
              hidden_dims=128,batch_size=64,num_epochs= 50)

pr_hyper = prep_hyper(sequenceLength=sequence_length,max_words=vocab_size)

from keras.layers import Dense,TimeDistributed, Dropout, Input,LeakyReLU, GlobalMaxPooling1D, Convolution1D, Embedding,SpatialDropout1D,BatchNormalization
from keras.layers.merge import Concatenate
from keras.models import Model
from keras.utils.vis_utils import plot_model
from keras import regularizers,layers
from keras.layers import Bidirectional
from keras.callbacks import ReduceLROnPlateau,EarlyStopping

def build_model(pr_hyper,m_hyper):
    
    model_input = Input(shape=(pr_hyper.sequenceLength))
    embedding = Embedding(pr_hyper.max_words, m_hyper.embedding_dim,weights=[emb],trainable=False)(model_input)

    print(embedding.shape) #---> (None,300,300)

    ## CONVULATIONAL
    conv_kern_reg = regularizers.l2(0.0001)
    conv_bias_reg = regularizers.l2(0.0001)
    
    conv_blocks = []
    for sz in m_hyper.filter_sizes:
        conv = TimeDistributed(Convolution1D(filters=m_hyper.num_filters,
                             kernel_size=sz,
                             #padding="same",
                             activation="relu",
                             strides=1,
                             #kernel_regularizer=conv_kern_reg,
                             #bias_regularizer=conv_bias_reg
                            ))(embedding) #---> showing error that shape coming is (None,300)
        conv = TimeDistributed(GlobalMaxPooling1D())(conv)
        conv = TimeDistributed(Flatten())(conv)
        lstm = Bidirectional(layers.LSTM(256,activation='relu',return_sequences=True,))(conv)
        lstm = BatchNormalization()(lstm)
        lstm = Bidirectional(layers.LSTM(128,return_sequences=True,activation='relu'))(lstm)
        lstm = BatchNormalization()(lstm)
        lstm = Bidirectional(layers.LSTM(128,kernel_regularizer=regularizers.l2(0.01),activation='relu'))(lstm)
        conv_blocks.append(lstm)
        
    # merge
    x = Concatenate()(conv_blocks) if len(conv_blocks) > 1 else conv_blocks[0]
        
    concat = Dense(m_hyper.hidden_dims, activation="relu")(x)
    concat = BatchNormalization()(concat)
    concat = Dropout(m_hyper.dropout_prob[1])(concat)
    
    concat = Dense(16)(concat)
    concat = BatchNormalization()(concat)
    concat = LeakyReLU()(concat)
    
    model_output = Dense(2, activation="softmax")(concat)
    model = Model(model_input, model_output)
    
    model.compile(loss="binary_crossentropy", optimizer=keras.optimizers.Adam(learning_rate=0.0005), metrics=["accuracy"]) #categorical_crossentropy
    print(model.summary())
    
    plot_model(model, show_shapes=True, to_file='multichannelCNN&LSTM.png')
    return model

但是该模型在没有 TimeDistributed Layer 的情况下工作得非常好（显然当时没有 LSTM 层堆叠在它们之上）。

有什么问题在这里我无法理解。

4

0 回答 0