Check failed: datum_channels > 0 (0 vs. 0) with non-image data

[Alex Ter-Sarkisov]

I use features as inputs in a ConvNN, dimensions 4096x19x32. Input layer looks like 

layer{

  name: "data"

  type: "Data"

  top: "data"

  #input_param { shape: {dim:20 dim: 4096 dim: 19 dim: 32 } }

  data_param {

     source: "train1lmdb"

     batch_size: 20

     backend: LMDB

  }

  include {

    phase: TRAIN

  }

}

The error I get is Check failed: datum_channels > 0 (0 vs. 0), so I guess caffe requies a Datum object; I created a blobproto object with pycaffe's array_to_blobproto().

It works if I change the layer type to "Input" and uncomment the input_param, but I know it's wrong and there's some error, as training/validation errors go to 0. 

What am I doing wrong? 

[Przemek D]

Is your database properly created? It looks like it was read okay, but the datum it contains has zero channels.