Keras GRUCell缺少1个必需的位置参数：'states'。

Question

我尝试使用Keras构建一个三层的RNN模型。以下是其中一部分代码：

    model = Sequential()
    model.add(Embedding(input_dim = 91, output_dim = 128, input_length =max_length))
    model.add(GRUCell(units = self.neurons, dropout = self.dropval,  bias_initializer = bias))
    model.add(GRUCell(units = self.neurons, dropout = self.dropval,  bias_initializer = bias))
    model.add(GRUCell(units = self.neurons, dropout = self.dropval,  bias_initializer = bias))
    model.add(TimeDistributed(Dense(target.shape[2])))

然后我遇到了这个错误：

call() missing 1 required positional argument: 'states'

错误详情如下：

~/anaconda3/envs/hw3/lib/python3.5/site-packages/keras/models.py in add(self, layer)
487                           output_shapes=[self.outputs[0]._keras_shape])
488         else:
--> 489             output_tensor = layer(self.outputs[0])
490             if isinstance(output_tensor, list):
491                 raise TypeError('All layers in a Sequential model '

 ~/anaconda3/envs/hw3/lib/python3.5/site-packages/keras/engine/topology.py in __call__(self, inputs, **kwargs)
601 
602             # Actually call the layer, collecting output(s), mask(s), and shape(s).
--> 603             output = self.call(inputs, **kwargs)
604             output_mask = self.compute_mask(inputs, previous_mask)
605 

- beepretty

Answer 1

1. Don't use Cell classes (i.e. GRUCell or LSTMCell) in Keras directly. They are computation cells which are wrapped by the corresponding layers. Instead use the Layer classes (i.e. GRU or LSTM):

   model.add(GRU(units = self.neurons, dropout = self.dropval,  bias_initializer = bias))
   model.add(GRU(units = self.neurons, dropout = self.dropval,  bias_initializer = bias))
   model.add(GRU(units = self.neurons, dropout = self.dropval,  bias_initializer = bias))
   

   The LSTM and GRU use their corresponding cells to perform computations over the all timesteps. Read this SO answer to learn more about their difference.

2. When you are stacking multiple RNN layers on top of each other you need to set their return_sequences argument to True in order to produce the output of each timestep, which in turn is used by the next RNN layer. Note that you may or may not do this on the last RNN layer (it depends on your architecture and the problem you are trying to solve):

   model.add(GRU(units = self.neurons, dropout = self.dropval,  bias_initializer = bias, return_sequences=True))
   model.add(GRU(units = self.neurons, dropout = self.dropval,  bias_initializer = bias, return_sequences=True))
   model.add(GRU(units = self.neurons, dropout = self.dropval,  bias_initializer = bias))
   

- today

---

这完美地解决了我的问题，谢谢你提醒我使用"return_sequences"，我确实遇到过这个问题。非常感谢！ - beepretty

很好的解释。这也解决了我的问题（鼓掌）！ - UNagaswamy