differences between Convolution1D/2D and MaxPooling1D/2D

[Kyle McDonald]

hi,

i'm interested in working with 1d data in keras (waveforms and frequency spectra) similar to studies i've done with lasagne[1].

when i use the convolution and pooling layers following the example of mnist_cnn, but in 1d instead of 2d, i get unexpected layer output sizes.[2]

with both, the input shape looks correct:

2D: Initial input shape: (None, 1, 28, 28)

1D: Initial input shape: (None, 1, 28)

but after the same convolution, they look very different:

2D: Convolution2D (convolution2d) (None, 32, 28, 28)

1D: Convolution1D (convolution1d) (None, 1, 32)

while the 2D convolution has resulted in (batch_size, nb_filters, nb_conv, nb_conv), the 1D convolution results in (batch_size, ?, nb_filters).

then after a max pooling operation it gets even stranger:

2D: MaxPooling2D (maxpooling2d)   (None, 32, 14, 14)

1D: MaxPooling1D (maxpooling1d)   (None, 0, 32)

since it seems like the 2d implementation is correct, should i stick with 2d convolutions for now and just use 1 for the row size? or would this be inefficient?

thanks!

kyle

[1] https://gist.github.com/kylemcdonald/662f93a4bcaf0c046b6a

[2] https://gist.github.com/kylemcdonald/ff2aaa33b78f9fb48867

[François Chollet]

Convolution1D is doing temporal convolution over the time dimension (which is dimension 1). If that's not what you need, then you should probably use Convolution2D instead. For a typical use case of Convolution1D with time series, you can check out our example "CNN LSTM" for IMDB.

--
You received this message because you are subscribed to the Google Groups "Keras-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to keras-users...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/keras-users/CAO9UE66ZccrU4XwqUaARTsTErwE2MiNBRTETWu9UuL0Eg_njnw%40mail.gmail.com.
For more options, visit https://groups.google.com/d/optout.

[Kyle McDonald]

I see. So Convolution2D operates on (batch size, channels, rows, columns) and Covolution1D operates on (batch size, length, channels). I was expecting Convolution1D to be (batch size, channels, length) keeping consistent with Convolution2D, but I can imagine most applications of Convolution1D have a large number of "channels" (e.g., one hot token vectors) so it's more efficient to make it the final dimension.

[François Chollet]

This is due to dimension ordering conventions in Keras, which are inherited from Theano. Note that you can also run Convolution2D on (samples, rows, columns, channels) if you specify dim_ordering="tf".