当input_shape指定为3-d时，Keras SimpleRNN出错

Question

I'm doing trying to train from text on a SimpleRNN on Keras.

In Keras, i specified a very simple parameters for SimpleRNN as below:

model = Sequential()
model.add(SimpleRNN(output_dim=1, input_shape=(1,1,1))

I understand that input_shape should be (nb_samples, timesteps, input_dim), the same as my train_x.shape

so i was surprised that i received the following error.

Traceback (most recent call last):
  File "C:/Users/xxx/xxxx/xxx/xxx.py", line 262, in <module>
    model.add(SimpleRNN(output_dim=vocab_size, input_shape=train_x.shape))
  File "C:\Anaconda3\envs\py34\lib\site-packages\keras\models.py", line 275, in add
    layer.create_input_layer(batch_input_shape, input_dtype)
  File "C:\Anaconda3\envs\py34\lib\site-packages\keras\engine\topology.py", line 367, in create_input_layer
    self(x)
  File "C:\Anaconda3\envs\py34\lib\site-packages\keras\engine\topology.py", line 467, in __call__
    self.assert_input_compatibility(x)
  File "C:\Anaconda3\envs\py34\lib\site-packages\keras\engine\topology.py", line 408, in assert_input_compatibility
    str(K.ndim(x)))
Exception: Input 0 is incompatible with layer simplernn_1: expected ndim=3, found ndim=4

Not sure why keras "found ndim=4" when only 3 was specified!

for clarity, my

train_x.shape = (73, 84, 400)

and

vocab_size=400

. As long as input_shape is fed 3d and above, i realised an error will result.

Any help will be greatly appreciated!!! :))

Answer 1

You are not supposed to include n_samples in the input shape of the model. So you have to specify a tuple of size 2 for input shape of a your layer (or set the first element of shape to None ). Here Keras automatically adds None to your input shape resulting in ndim=4 . More info on this can be found here .

Also it appears that your input_dim=400 (assuming you use one-hot coding representation of words in vocabulary) and that your training data consists of 73 texts (pretty small) each having length of 84 . So you should probably set input_shape=(84,400).