SimpleRNN Error about keras HOT 4 CLOSED

In a Keras model, the only requirements on the input are those of the initial layer.

What does your input look like? I suspect you might be working with a 2D input. The imdb_lstm.py script in the examples folder provides a demo of how to build RNNs. You can replace the LSTM layer in that example with a SimpleRNN layer and it will work.

from keras.

This is extent of the code - no inputs defined. The compilation fails.

I haven't delved deep enough into the implementation yet, but I'm guessing Theano finding that the input will be a fmatrix, so the dimshuffle((1,0,2)) is invalid. I'm guessing this will require a reshape before-hand, depending on how the batches are being handled. Something along the line of

X = X.reshape((batch_size, X.shape[0]/batch_size, X.shape[1])).dimshuffle((1,0,2))

or

X = X.reshape((X.shape[0]/batch_size, batch_size, X.shape[1]))

I can't recall if Theano handles the second snippet intelligently, resulting in equivalent outputs from both of the above snippets.

EDIT:
Ah, a Reshape layer, perhaps?

from keras.

You were right, there was in fact an issue were recurrent layers had their default input shape set to matrix whereas they were expecting 3D inputs. This was not immediately apparent because the tests and example script did not implement any model where the recurrent layer came first.

The issue is fixed. Thanks for pointing this out!

from keras.

Awesome, thanks for the update.

from keras.