我有两个Keras子模型(model_1, model_2)，我通过逻辑上将它们“串联”起来，使用keras.models.Model()构建了我的完整model。我的意思是，model_2接受model_1的输出以及一个额外的输入张量，而model_2的输出就是我的完整model的输出。完整的model被成功创建，并且我也可以使用compile/train/predict。

然而，我希望通过在2个GPU上运行来并行化model的训练，因此我使用了multi_gpu_model()，但这导致了错误：

AssertionError: Could not compute output Tensor("model_2/Dense_Decoder/truediv:0", shape=(?, 33, 22), dtype=float32)

我尝试过分别使用multi_gpu_model(model_1, gpus=2)和multi_gpu_model(model_2, gpus=2)来并行化这两个子模型，两者都成功了。问题仅出现在完整模型上。

我使用的是Tensorflow 1.12.0和Keras 2.2.4。一个展示问题的代码片段（至少在我的机器上）是：

from keras.layers import Input, Dense,TimeDistributed, BatchNormalizationfrom keras.layers import CuDNNLSTM as LSTMfrom keras.models import Modelfrom keras.utils import multi_gpu_modeldec_layers = 2codelayer_dim = 11bn_momentum = 0.9lstm_dim = 128td_dense_dim = 0output_dims = 22dec_input_shape = [33, 44]# MODEL 1latent_input = Input(shape=(codelayer_dim,), name="Latent_Input")# Initialize list of state tensors for the decoderdecoder_state_list = []for dec_layer in range(dec_layers):    # The tensors for the initial states of the decoder    name = "Dense_h_" + str(dec_layer)    h_decoder = Dense(lstm_dim, activation="relu", name=name)(latent_input)    name = "BN_h_" + str(dec_layer)    decoder_state_list.append(BatchNormalization(momentum=bn_momentum, name=name)(h_decoder))    name = "Dense_c_" + str(dec_layer)    c_decoder = Dense(lstm_dim, activation="relu", name=name)(latent_input)    name = "BN_c_" + str(dec_layer)    decoder_state_list.append(BatchNormalization(momentum=bn_momentum, name=name)(c_decoder))# Define model_1model_1 = Model(latent_input, decoder_state_list)# MODEL 2inputs = []decoder_inputs = Input(shape=dec_input_shape, name="Decoder_Inputs")inputs.append(decoder_inputs)xo = decoder_inputsfor dec_layer in range(dec_layers):    name = "Decoder_State_h_" + str(dec_layer)    state_h = Input(shape=[lstm_dim], name=name)    inputs.append(state_h)    name = "Decoder_State_c_" + str(dec_layer)    state_c = Input(shape=[lstm_dim], name=name)    inputs.append(state_c)    # RNN layer    decoder_lstm = LSTM(lstm_dim,                   return_sequences=True,                   name="Decoder_LSTM_" + str(dec_layer))    xo = decoder_lstm(xo, initial_state=[state_h, state_c])    xo = BatchNormalization(momentum=bn_momentum, name="BN_Decoder_" + str(dec_layer))(xo)    if td_dense_dim > 0: # Squeeze LSTM interconnections using Dense layers        xo = TimeDistributed(Dense(td_dense_dim), name="Time_Distributed_" + str(dec_layer))(xo)# Final Dense layer to return probabilitiesoutputs = Dense(output_dims, activation='softmax', name="Dense_Decoder")(xo)# Define model_2model_2 = Model(inputs=inputs, outputs=[outputs])# FULL MODELlatent_input = Input(shape=(codelayer_dim,), name="Latent_Input")decoder_inputs = Input(shape=dec_input_shape, name="Decoder_Inputs")# Stack the two models# Propagate tensors through 1st modelx = model_1(latent_input)# Insert decoder_inputs as the first input of the 2nd modelx.insert(0, decoder_inputs)# Propagate tensors through 2nd modelx = model_2(x)# Define full modelmodel = Model(inputs=[latent_input, decoder_inputs], outputs=[x])# Parallelize the modelparallel_model = multi_gpu_model(model, gpus=2)parallel_model.summary()

非常感谢任何帮助/建议。

回答：

我找到了问题的解决方案，但我还不确定如何解释这个解决方案。

问题是由x.insert(0, decoder_inputs)引起的，我将其替换为x = [decoder_inputs] + x。这两种方法似乎都得到了相同的张量列表，但是multi_gpu_model在第一种情况下会报错。