Shape of input blob - siamese network

[Makiko]

Hello, 

can you help me with input data to siamese network?

I would like to train siamese network. Data layer in my train.prototxt is (type: "Input") - http://caffe.berkeleyvision.org/tutorial/layers/input.html. My image data are in numpy array of shape (N x 2 x C x H x W), where N is number of (pairs) images, 2 is the pair, C is channel (in my case 1 B&W), H is height and W is width. I have defined in my train.prototxt input_param

 

input_param {

    shape: {

      dim: 50

      dim: 1

      dim: 255

      dim: 255

      }

  }

what is shape of blob what the net expects.

My last layer is contrastiveloss.

My question is how should I feed the net with my data?

I guess that I should reshape my array somehow...

I planned to use slice layer after data layer.

Could you provide me a short example how to do that?

Thank you in advance.

Makiko W.

[Przemek D]

If your data is 5-dimensional, so should be your input. In theory you would have to add another dim param equal to 2, but I never tried that in practice.
What I successfully used instead, was concatenating images side by side and then using the Slice layer to retrieve the individual ones. In your case the input would look like:

input_param {

  shape: {

    dim: 50

    dim: 1

    dim: 255

    dim: 510

  }

}

Then you'd simply feed two images as if they were one, and Slice them with param:
slice_dim: 3
slice_point: 255