Question

我正在尝试在 Keras 中实现神经网络模型，但我遇到了维度问题。根据模型架构，我应该从最后一个（全连接）层获得 1 作为输出维度，但我获得了 2D 数据作为输出。

我正在尝试实现 paper 中的图 4。

我的实现如下：

import numpy as np
from tensorflow.keras import layers, Model
from tensorflow.keras.optimizers import Adam, RMSprop
from tensorflow.keras.utils import plot_model
from tensorflow.python.keras.regularizers import l2
from keras.datasets import imdb
from keras.preprocessing import sequence

def get_model(vocabulary_size, embedding_dim, input_length, summary=True):
  inputs = layers.Input(shape=(input_length))
  x = layers.Embedding(vocabulary_size, embedding_dim)(inputs)

  branch1 = layers.Conv1D(128, kernel_size=(3),padding='same',kernel_regularizer=l2(0.01), activation='relu')(x)
  branch1 = layers.MaxPool1D(pool_size=(2))(branch1)
  branch1 = layers.Dropout(0.5)(branch1)
  branch1 = layers.BatchNormalization()(branch1)
  branch1 = layers.LSTM(128, return_sequences=True)(branch1)

  branch2 = layers.Conv1D(128, kernel_size=(5),padding='same',kernel_regularizer=l2(0.01), activation='relu')(x)
  branch2 = layers.MaxPool1D(pool_size=(2))(branch2)
  branch2 = layers.Dropout(0.5)(branch2)
  branch2 = layers.BatchNormalization()(branch2)
  branch2 = layers.LSTM(128, return_sequences=True)(branch2)


  branch3 = layers.Conv1D(128, kernel_size=(7),padding='same',kernel_regularizer=l2(0.01), activation='relu')(x)
  branch3 = layers.MaxPool1D(pool_size=(2))(branch3)
  branch3 = layers.Dropout(0.5)(branch3)
  branch3 = layers.BatchNormalization()(branch3)
  branch3 = layers.LSTM(128, return_sequences=True)(branch3)

  branch4 = layers.Conv1D(128, kernel_size=(9),padding='same',kernel_regularizer=l2(0.01), activation='relu')(x)
  branch4 = layers.MaxPool1D(pool_size=(2))(branch4)
  branch4 = layers.Dropout(0.5)(branch4)
  branch4 = layers.BatchNormalization()(branch4)
  branch4 = layers.LSTM(128, return_sequences=True)(branch4)

  concat = layers.concatenate([branch1, branch2, branch3, branch4], name='Concatenate')

  outputs = layers.Dense(1, activation='sigmoid')(concat)


  model = Model(inputs, outputs)
  model.compile(loss='binary_crossentropy', optimizer=Adam(lr=1e-3), metrics='accuracy')
  
  print(model.summary())

  return model

EMBEDDING_DIM = 32
VOCABULARY_SIZE = 5000
seq_length = 500

my_model = get_model(VOCABULARY_SIZE, EMBEDDING_DIM, seq_length)

Answer 1

在最后一个 LSTM 层中需要 return_sequences = False，因此它只返回最后一个隐藏状态。这样它只返回一个向量。所以 4 个分支返回 4 个向量，它们连接成一个。

更多详情：How to use return_sequences option and TimeDistributed layer in Keras?

深度学习模型在 Keras 中的实现

1 个答案: