python - 断言错误: Could not compute output Tensor when using multi_gpu_model() in Keras

Question

我有 2 个 Keras 子模型( model_1 、 model_2 )，从中我形成了完整的 model使用keras.models.Model()通过将它们逻辑地堆叠在“系列”中。我的意思是model_2接受 model_1 的输出加上一个额外的输入张量和 model_2 的输出是我的完整 model 的输出。完整model创建成功，我也可以使用compile/train/predict 。

但是，我想并行化 model 的训练通过在 2 个 GPU 上运行它，因此我使用 multi_gpu_model()失败并出现错误:

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

我尝试使用 multi_gpu_model(model_1, gpus=2) 单独并行化两个子模型和multi_gpu_model(model_2, gpus=2) ，但都成功了。该问题仅出现在完整模型中。

我正在使用Tensorflow 1.12.0和Keras 2.2.4。演示该问题的代码片段(至少在我的机器上)是:

from keras.layers import Input, Dense,TimeDistributed, BatchNormalization
from keras.layers import CuDNNLSTM as LSTM
from keras.models import Model
from keras.utils import multi_gpu_model


dec_layers = 2
codelayer_dim = 11
bn_momentum = 0.9
lstm_dim = 128
td_dense_dim = 0
output_dims = 22
dec_input_shape = [33, 44]


# MODEL 1
latent_input = Input(shape=(codelayer_dim,), name="Latent_Input")

# Initialize list of state tensors for the decoder
decoder_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_1
model_1 = Model(latent_input, decoder_state_list)


# MODEL 2
inputs = []

decoder_inputs = Input(shape=dec_input_shape, name="Decoder_Inputs")
inputs.append(decoder_inputs)

xo = decoder_inputs

for 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 probabilities
outputs = Dense(output_dims, activation='softmax', name="Dense_Decoder")(xo)

# Define model_2
model_2 = Model(inputs=inputs, outputs=[outputs])


# FULL MODEL
latent_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 model
x = model_1(latent_input)
# Insert decoder_inputs as the first input of the 2nd model
x.insert(0, decoder_inputs)
# Propagate tensors through 2nd model
x = model_2(x)

# Define full model
model = Model(inputs=[latent_input, decoder_inputs], outputs=[x])

# Parallelize the model
parallel_model = multi_gpu_model(model, gpus=2)
parallel_model.summary()

非常感谢您的帮助/提示。

Answer 1

我找到了问题的解决方案，但我不确定如何证明其合理性。

该问题是由 x.insert(0, detector_inputs) 引起的，我将其替换为 x = [decoder_inputs] + x。两者似乎都会产生相同的张量列表，但是 multi_gpu_model 在第一种情况下会提示。