Multiple sequences about keras HOT 3 CLOSED

That is precisely how recurrent layers work, they take 3D inputs of shape (nb_samples, max_sample_length, input_dim). nb_samples is the batch size.

Can you post your model and explain what you were trying to do?

from keras.

Hey, so what I have is a list of strings of text of variable length. I turn them into a list of word sequences with the tokenizer. Ideally I'll turn it into a 1-of-N encoding but just for testing puposes now. I want to try to feed it one of these sequences and predict the next word.

toki.fit_on_texts(texts)
textseq = toki.texts_to_sequences(texts)
padtext = sequence.pad_sequences(textseq)
padtext = padtext.reshape(padtext.shape[0],padtext.shape[1],1) #4 1d sequences, each of length 400k
vocabsize = len(toki.word_index)

model = Sequential()
model.add(LSTM(vocabsize,128))
model.add(Dense(128, vocabsize, init='lecun_uniform'))
model.add(Activation('sigmoid'))

loss = RMSprop(lr=0.01, rho=0.9, epsilon=1e-6)
model.compile(loss='categorical_crossentropy', optimizer=loss, class_mode="categorical")

doing

model.fit(padtext[:,:-1,:],padtext[:,1:,:])

gives me the error

('Bad input argument to theano function with name "build/bdist.linux-x86_64/egg/keras/models.py:64"  at index 1(0-based)', 'Wrong number of dimensions: expected 2, got 3 with shape (4, 405985, 1).')

telling me LSTMs don't actually want 3d arrays. Ideas?

from keras.

It wasn't the LSTM, it was some other level in the network. Reshaping and flattening where required fixes most issues. Thanks :)

from keras.