Keras padding replaces all my values by 0

Question

In order to adapt a sequence of data I want to use in a univariate LSTM and that have a "multi-step time", The simplest way is to do some padding. My initial dataset looks like this, it's a numpyarray:

X
[0.295046, 0.325147, 0.361293]
[0.249307,0.444077]
[0.570017,0.525082,0.475404,0.390616]

What I've tried so far is:

from keras.preprocessing.sequence import pad_sequences
padded_x = pad_sequences(X)
print(padded_x)

Instead of adding zeros to "complete" the dataset:

X
[0.0, 0.295046, 0.325147, 0.361293]
[0.0, 0.0, 0.249307,0.444077]
[0.570017,0.525082,0.475404,0.390616]

It just replaces all the values by 0. I don't know what I am missing...

Thanks in advance :)

score 4 · Accepted Answer · answered Sep 17 '19 at 14:47

4

By default, pad_sequences's dtype argument is set to 'int32', thus resulting in the rounding of your float values to 0. You thus need to use pad_sequences(x, dtype='float32') (or 'float64', as you see fit).

answered Sep 17 '19 at 14:47

pandrey

354
1
5

3

Maybe use `dtype=x.dtype` for generality? – Ollin Boer Bohan Sep 17 '19 at 14:48
2

@Nour You are welcome! And if I were you I would also take the clever advice from @OllinBoerBohan of using `x.dtype` :-) – pandrey Sep 17 '19 at 15:23

Keras padding replaces all my values by 0

1 Answers1