Implementing a Convolutional Rec Net

[geevi]

Hi,
I was trying to implement the following:

1.) A convolutional net convert input to a vector of 500 real numbers z[1] .. z[500]
2.) A reccurent net which is given the same input z for 5 time steps and outputs a sequence of 5 labels.

First used the repeater, but it runs the conv net 5 times on the same input. I want the conv net to run once to get z, and then the recurrent net is given the same input z for 5 time steps.

Any suggestions?

thanks

[nicholas-leonard]

Hi Girish,

Try something like :

nn.Sequential()
:add(convnet)
:add(nn.Repeater(rnn))

Regards,

Nicholas Léonard
917-741-7570

On Mon, Mar 14, 2016 at 6:14 AM, Girish Varma notifications@github.com
wrote:

Hi,
I was trying to implement the following:

1.) A convolutional net convert input to a vector of 500 real numbers z[1]
.. z[500]
2.) A reccurent net which is given the same input z for 5 time steps and
outputs a sequence of 5 labels.

First used the repeater, but it runs the conv net 5 times on the same
input. I want the conv net to run once to get z, and then the recurrent net
is given the same input z for 5 time steps.

Any suggestions?

thanks

—
Reply to this email directly or view it on GitHub
#159.

[geevi]

Hi Nicholas,
Wow. That was simple. I am just getting used to torch.
Thanks a lot.

Is there a way to make the rnn run on the same input, for time steps which depends on the target?
I was trying to build an ocr, where the input is an image of word and the target is a unicode string.
So the target has varying lengths. The hope was that the convnet would output a low dimensional representation of the image. Then the rnn would run for time steps equal to the length of the target string, on the same low dim. rep. (output of convnet).

I will be grateful for any suggestions.

[nicholas-leonard]

@geevi The best way to do that would be to pad your targets so that they all have the same length. You could for example pad the last elements of a target with a special end-of-sequence token.