1D Convolutions support?

[ddofer]

Hi,
I'm trying out your (lovely) package on non image data, and would like to try using convolutions - I see that currently only 2D convolutions are supported, any chance of adding in 1D support/for non image/square data?

(i.e As my first layer I tried:
Convolution("Rectifier", channels=32, kernel_shape=(1,1))
)

'AssertionError: Input array is not in image shape, and could not assume a square.'

Thanks!

[alexjc]

1D convolutions work already, there's a test for it. The problem here is that you want to apply the convolution over a 2D image but you're giving the NN a 1D array, and it can't figure out what size you mean. (It tries to guess square, but your data is not square.)

Try reshaping your inputs?

EDIT: In case you want to convolve over a 1D image, you still need to specify either (N, 1) or (1, N) as an array size—just like the kernel you have is 2D.

[ddofer]

I am using a 1d array of features.. ;
So, The solution is to reshape the features (X) into a square? Ok, thanks.
On Jun 28, 2015 2:33 PM, "Alex J. Champandard" notifications@github.com
wrote:

1D convolutions work already, there's a test for it. The problem here is
that you want to apply the convolution over a 2D image but you're giving
the NN a 1D array, and it can't figure out what size you mean.

Try reshaping your inputs?

—
Reply to this email directly or view it on GitHub
#96 (comment)
.

[alexjc]

If you have a (1,1) kernel you can use any input shape as long as it's 2D!

[ddofer]

By "N", do you mean a string input, or an integer with a preset size?

[alexjc]

I mean an integer with preset size that matches your data.

[elenacuoco]

Hi,
I have the same request. Your package is really useful. I did not understand how to use conv1D in your package. I'm looking for something like Conv1D in Lasagne (http://lasagne.readthedocs.org/en/latest/modules/layers/conv.html),without reshaping the data...

[alexjc]

I will have to look at this and add a test or two. Thanks for the feedback!

[alexjc]

We use Lasagne now, I'd consider adding this if there's enough interest!