How to reshape input data between data layer and convolution layer?

[Hugo G]

Hi,

What I wanted to do is to use convolution layer to train a network on non-image data(1 dimension PCM audio data). So that the dimension of the input LMDB data is i.e. 256(batch size)  x  1  x 1960 x  1. In order to use convolutional filters whose size length are 2 or larger, I'll have to reshape the input data dimension to have 2-dimensional image like format. So I defined the following right after data layer:

layer {

    name: "reshape"

    type: "Reshape"

    bottom: "data"

    top: "conv1"

    reshape_param {

      shape {

        dim: 0  # copy the dimension from below

        dim: 0  # copy the dimension from below

        dim: 7

        dim: 280 #  -1: infer it from the other dimensions

      }

    }

} 

The data layer looks like this:

name: "LeNet"

layer {

   name: "data"

   type: "Data"

   top: "data"

   top: "label"

   include {

   phase: TRAIN

   }

   data_param {

     source: "/path/to/lmdb_train"

     batch_size: 256

backend: LMDB

   }

 }

 

 layer {

   name: "data"

   type: "Data"

   top: "data"

   top: "label"

   include {

   phase: TEST

   }

   data_param {

     source: "/path/to/lmdb_test"

     batch_size: 100

backend: LMDB

   }

 }

However, Caffe complained: 

F0105 11:38:01.728590 32701 insert_splits.cpp:29] Unknown bottom blob 'data' (layer 'reshape', bottom index 0)

If I add one more output at data layer, i.e:

   top: "data"

   top: "label"

   top: "reshape"

It complained again that one layer cannot have more that 2 outputs:    "Check failed: MaxTopBlobs() >= top.size() (2 vs. 3) Data Layer produces at most 2 top blob(s) as output."

 

Does anyone know how to reshape 1 dimensional data to 2 dimensional between data layer and convolution layer so that LeNet is working? I can find zero resources regarding this issue.

Thanks

Hugo 

[Przemek D]

Will a 1x1960x1 blob reshaped to 1x7x280 still make sense? Will a convolution over it make sense?

I can't see why convolution over the original blob would be impossible... have you tried using a rectangular kernel? That is one where you don't specify a kernel_size but instead kernel_h and kernel_w - you could use a 1x7x1 kernel to slide over that volume with no problems, but I haven't done  that myself.

For a good documentation on layers I'll recommend you read this one by Jonathan Williford.