python - 有状态 RNN 在 Keras 功能模型中张量形状错误

Question

我定义了一个 Keras 功能模型，其中包含一个带有状态 LSTM 的 block ，如下所示:

import numpy as np
from tensorflow.python import keras


data = np.ones((1,2,3))

input_shape = data.shape  # batch size, step size, input size
output_units = 2

### input block ###
inputs = keras.layers.Input(batch_shape=input_shape, name="inputs")


### model block with stateful RNN ###
dummy_inputs_1 = keras.layers.Input(batch_shape=input_shape, name="dummy_inputs_1")
recurrent_1 = keras.layers.LSTM(units=input_shape[-1], batch_input_shape=input_shape,
                                return_sequences=True, stateful=True,
                                name="recurrent_1")(dummy_inputs_1)

dense_1 = keras.layers.Dense(output_units, batch_input_shape=(
    input_shape[0], input_shape[-1], input_shape[1]),
                             name="dense_1")
output_1 = keras.layers.TimeDistributed(dense_1, input_shape=input_shape, name="output_1")(recurrent_1)

model_1 = keras.models.Model(inputs=[dummy_inputs_1], outputs=[output_1], name="model_1")
model_1.compile(loss='mean_squared_error',
                optimizer='Nadam',
                metrics=['accuracy'])

model_1.predict(data) #works

### add model block to model ###
model_block = model_1(inputs)
model = keras.models.Model(inputs=[inputs], outputs=[model_block], name="model")
model.compile(loss='mean_squared_error',
                optimizer='Nadam',
                metrics=['accuracy'])

model_1.predict(data) #works

model.predict(data)  #fails

正如所写，第一个 predict() 调用(对包含有状态 LSTM 层的内部模型 block )工作正常，但第二个调用失败并出现以下错误:

 Traceback (most recent call last):
  File ".../functional_stateful.py", line 38, in <module>
    model_1.predict(data)
  File ".../local/lib/python2.7/site-packages/tensorflow/python/keras/engine/training.py", line 1478, in predict
    self, x, batch_size=batch_size, verbose=verbose, steps=steps)
  File ".../local/lib/python2.7/site-packages/tensorflow/python/keras/engine/training_arrays.py", line 363, in predict_loop
    batch_outs = f(ins_batch)
  File ".../local/lib/python2.7/site-packages/tensorflow/python/keras/backend.py", line 2897, in __call__
    fetched = self._callable_fn(*array_vals)
  File ".../local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 1454, in __call__
    self._session._session, self._handle, args, status, None)
  File ".../local/lib/python2.7/site-packages/tensorflow/python/framework/errors_impl.py", line 519, in __exit__
    c_api.TF_GetCode(self.status.status))
tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'inputs' with dtype float and shape [1,2,3]
     [[Node: inputs = Placeholder[dtype=DT_FLOAT, shape=[1,2,3], _device="/job:localhost/replica:0/task:0/device:CPU:0"]()]]

在 LSTM 定义中注释掉 stateful=True 后，整个过程运行良好。有谁知道发生了什么事吗？

编辑: 显然，仅调用另一层上的有状态模型 block 就足以导致该 block 的 predict() 失败(即此代码因相同的错误而失败):

import numpy as np
from tensorflow.python import keras

data = np.ones((1,2,3))

input_shape = data.shape  # batch size, step size, input size
output_units = 2

### input block ###
inputs = keras.layers.Input(batch_shape=input_shape, name="inputs")


### sample model block with stateful RNN ###
dummy_inputs_1 = keras.layers.Input(batch_shape=input_shape, name="dummy_inputs_1")
recurrent_1 = keras.layers.LSTM(units=input_shape[-1], batch_input_shape=input_shape,
                                return_sequences=True, stateful=True,
                                name="recurrent_1")(dummy_inputs_1)

model_1 = keras.models.Model(inputs=[dummy_inputs_1], outputs=[recurrent_1], name="model_1")
model_1.compile(loss='mean_squared_error',
                optimizer='Nadam',
                metrics=['accuracy'])

# ### add model block to model ###
model_block = model_1(inputs)

model_1.predict(data) #fails 

编辑 2: 但显然，在另一个 block 上调用有状态 block 之前添加对 Predict() 的调用可以让您在之后仍然使用它(即下面的代码运行良好):

import numpy as np
from tensorflow.python import keras

data = np.ones((1,2,3))

input_shape = data.shape  # batch size, step size, input size
output_units = 2

### input block ###
inputs = keras.layers.Input(batch_shape=input_shape, name="inputs")


### sample model block with stateful RNN ###
dummy_inputs_1 = keras.layers.Input(batch_shape=input_shape, name="dummy_inputs_1")
recurrent_1 = keras.layers.LSTM(units=input_shape[-1], batch_input_shape=input_shape,
                                return_sequences=True, stateful=True,
                                name="recurrent_1")(dummy_inputs_1)

model_1 = keras.models.Model(inputs=[dummy_inputs_1], outputs=[recurrent_1], name="model_1")
model_1.compile(loss='mean_squared_error',
                optimizer='Nadam',
                metrics=['accuracy'])

model_1.predict(data) #works

# ### add model block to model ###
model_block = model_1(inputs)

model_1.predict(data) #works

Answer 1

我怀疑 stateful=True RNN 与多个输入不兼容。
(在您的代码中，您有 dummy_inputs_1 和 inputs。keras 在许多消息中将其称为“多个入站节点”。实际上，您有两个并行分支，一个关于原始 dummy_inputs_1，另一个关于新 inputs)

这是为什么呢？ Stateful=True 层旨在接收“一个序列”(或一批中的许多“并行”序列)，该序列被分成时间步长组。

当它收到批处理 2 时，它会将其解释为批处理 1 关于序列时间步长的续集。

当你有两个输入张量时，RNN 应该如何解释什么继续什么？你会失去“连续序列”的一致性。该层只有“一个状态张量”，它无法用它来跟踪并行张量。

因此，如果您要使用具有多个输入的有状态 RNN，我建议您创建该层的副本。如果您希望它们共享相同的权重，这可能需要具有公共(public)权重张量的自定义层。

现在，如果您打算使用此 block 一次，您可能应该使用 model_1.input 和 model_1.output 而不是提供另一个输入张量。