python - 如何在keras中连接不同的张量形状

Question

我正在尝试在keras中实现注意力机制，我的context_vector形状是shape=(?, 1024)
我的解码器嵌入形状是 shape=(?, 38, 1024)
context_vector 和解码器_embedding 都是张量，我如何连接它们......？

def B_Attention_layer(state_h,state_c,encoder_outputs):

  d0 = tf.keras.layers.Dense(1024,name='dense_layer_1')
  d1 = tf.keras.layers.Dense(1024,name='dense_layer_2')
  d2 = tf.keras.layers.Dense(1024,name='dense_layer_3')
  #below are the hidden states of LSTM 
  # my encoder output shape is shape=(?, 38, 1024)
  #my each hidden state shape is i.e.., state_c shape=(?, 1024) ,state_h shape=(?, 1024)
  hidden_with_time_axis_1 = tf.keras.backend.expand_dims(state_h, 1)
  hidden_with_time_axis_2 = tf.keras.backend.expand_dims(state_c, 1)
  score = d0(tf.keras.activations.tanh(encoder_outputs) + d1(hidden_with_time_axis_1) +  d2(hidden_with_time_axis_2))
  attention_weights = tf.keras.activations.softmax(score, axis=1)
  context_vector = attention_weights * encoder_outputs
  context_vector = tf.keras.backend.sum(context_vector, axis=1)
  input_to_decoder = tf.keras.layers.Concatenate(axis=-1)([context_vector,decoder_embedding])

  return input_to_decoder , attention_weights

当我尝试这个时，我遇到了如下的串联错误

ValueError: A `Concatenate` layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 1, 1024), (None, 38, 1024)]

Answer 1

在您的情况下，串联轴为 1，因为您有形状 (None, 1, 1024) 和 (None, 38, 1024)

示例:

I1 = tf.keras.Input(shape=(1, 1024))
# <tf.Tensor 'input_6:0' shape=(?, 1, 1024) dtype=float32>
I2 = tf.keras.Input(shape=(38, 1024))
# <tf.Tensor 'input_7:0' shape=(?, 38, 1024) dtype=float32>

concated = tf.keras.layers.Concatenate(axis=1)([I1,I2])
# output: <tf.Tensor 'concatenate_4/concat:0' shape=(?, 39, 1024) dtype=float32>

此外，由于连接轴为 1，因此错误消息指出其余尺寸必须相同。