张量连接输出到输入以提供新的 lstm 层

Question

我尝试重塑和连接一些输出以完成原始输入并在我的模型的下一阶段使用它。尺寸似乎匹配，但出现此错误：

Concatenate(axis=2)([tensor_input2, out_first_try])
*** ValueError: A `Concatenate` layer requires inputs with matching 
shapes except for the concat axis. Got inputs shapes: [(64, 10, 8), [(), 
(), ()]]

我也试试 :

tf.concat([tensor_input2, out_first_try], 2)

出现此错误：

tf.concat([tensor_input2, out_first_try], 2)
*** ValueError: Shape must be rank 3 but is rank 1 for '{{node 
tf.concat/concat}} = ConcatV2[N=2, T=DT_FLOAT, Tidx=DT_INT32] 
(Placeholder, tf.concat/concat/values_1, tf.concat/concat/axis)' with 
input shapes: [64,10,8], [3], [].

原因似乎是一样的，但我不知道如何处理，

    # tensor_input1 = [64,365,9]
    tensor_input1 = Input(batch_size=batch, shape=(X.shape[1], 
                    X.shape[2]), name='input1')
    # tensor_input2 = [64,10,8]
    tensor_input2 = Input(batch_size=batch, shape=(X2.shape[1], 
                    X2.shape[2]), name='input2')

    extractor = CuDNNLSTM(100, return_sequences=False, 
                 stateful=False, name='LSTM1')(tensor_input2)
    extractor = Dropout(rate = .2)(extractor)
   
    extractor = Dense(100, activation='softsign')(extractor)

    out_1 = Dense(10, activation='linear')(extractor2)

    # add a dimension to out_1 [64,10] to fit tensor_input2 
    out_first_try = tf.expand_dims(out_1, axis=2).shape.as_list()

    # concat in 3d dim the output to the original input
    # tensor_input2 =[64,10,8] 
    # out_first_try, after tf.expend [64,10,1]
    forcast_input  = Concatenate(axis=2)([tensor_input2, 
                     out_first_try])

    # forcast_input expected size [64,10,9]

    # finaly concat tensor_input1, new tensor_input2 side to side
    allin_input = Concatenate(axis=1)([tensor_input1, forcast_input])
    # allin_input  expected size [64,365+10,9]

    extractor2 = CuDNNLSTM(100, return_sequences=False, 
                 stateful=False, name='LSTM1')(allin_input )
    ...

Answer 1

将张量与列表连接将不起作用。所以，也许可以尝试这样的事情：

out_first_try = tf.expand_dims(out_1, axis=2)
forcast_input  = Concatenate(axis=2)([tensor_input2, out_first_try])

请注意，我已删除 shape.as_list()，因为顾名思义，它 returns 将张量的形状作为列表。你可以用这个例子来验证：

import tensorflow as tf

out_1 = tf.random.normal((5, 10))
out_first_try = tf.expand_dims(out_1, axis=2).shape.as_list()
tf.print(type(out_first_try))
#<class 'list'>

张量连接输出到输入以提供新的 lstm 层

tensor concat output to input to feed new lstm layer

python

concatenation

tensorflow

tensor

tensorflow2.0