CODEWITHSUNDEEP
neural-networkdeep-learningcaffeconv-neural-network
Lemm Ras
Lemm Ras

Reputation: 1032

Caffe, operations among batches

Since I have a classifier based on single patch scores, I would like to sum together the predictions a network produces for different images.

From https://github.com/BVLC/caffe/blob/master/src/caffe/proto/caffe.proto, Reduction does not support operation over axis different than the last one. Also a pooling operation would produce an average of its input but, obviously, without touching on the full batch.

I have implemented a python layer, but this is not fast enough for large scale experiments.

Is there a way to "sum" or, more generally, operate over the first axis with the tools already available?

Upvotes: 1

Views: 506

Answers (2)

Jonathan
Jonathan

Reputation: 5278

If your predictions have the dimensions: N x c (for mini-batch size of N and c channels), then you can splice this into c blobs with dimension N. You can feed these into a Reduction layer.

For example, you to write the following as a Jinja2 template:

layer {
  name: "pred-slice"
  type: "Slice"
  bottom: "pred"
{%- for num in range(10) %}
  top: "pred-{{ num }}-vector"
{%- endfor %}
  slice_param {
    slice_dim: 1
{%- for num in range(1, 10) %}
    slice_point: {{ num }}
{%- endfor %}
  }
  include {
    phase: TEST
  }
}

{%- for num in range(10) %}
layer {
  name: "pred-{{num}}"
  type: "Reduction"
  bottom: "pred-{{ num }}-vector"
  top: "pred-{{ num }}"
  include {
    phase: TEST
  }
  reduction_param {
    operation: MEAN
  }
}
{%- endfor %}

which expands to:

layer {
  name: "pred-slice"
  type: "Slice"
  bottom: "pred"
  top: "pred-0-vector"
  top: "pred-1-vector"
  top: "pred-2-vector"
  top: "pred-3-vector"
  top: "pred-4-vector"
  top: "pred-5-vector"
  top: "pred-6-vector"
  top: "pred-7-vector"
  top: "pred-8-vector"
  top: "pred-9-vector"
  slice_param {
    slice_dim: 1
    slice_point: 1
    slice_point: 2
    slice_point: 3
    slice_point: 4
    slice_point: 5
    slice_point: 6
    slice_point: 7
    slice_point: 8
    slice_point: 9
  }
  include {
    phase: TEST
  }
}
layer {
  name: "pred-0"
  type: "Reduction"
  bottom: "pred-0-vector"
  top: "pred-0"
  include {
    phase: TEST
  }
  reduction_param {
    operation: MEAN
  }
}
layer {
  name: "pred-1"
  type: "Reduction"
  bottom: "pred-1-vector"
  top: "pred-1"
  include {
    phase: TEST
  }
  reduction_param {
    operation: MEAN
  }
}
layer {
  name: "pred-2"
  type: "Reduction"
  bottom: "pred-2-vector"
  top: "pred-2"
  include {
    phase: TEST
  }
  reduction_param {
    operation: MEAN
  }
}
layer {
  name: "pred-3"
  type: "Reduction"
  bottom: "pred-3-vector"
  top: "pred-3"
  include {
    phase: TEST
  }
  reduction_param {
    operation: MEAN
  }
}
layer {
  name: "pred-4"
  type: "Reduction"
  bottom: "pred-4-vector"
  top: "pred-4"
  include {
    phase: TEST
  }
  reduction_param {
    operation: MEAN
  }
}
layer {
  name: "pred-5"
  type: "Reduction"
  bottom: "pred-5-vector"
  top: "pred-5"
  include {
    phase: TEST
  }
  reduction_param {
    operation: MEAN
  }
}
layer {
  name: "pred-6"
  type: "Reduction"
  bottom: "pred-6-vector"
  top: "pred-6"
  include {
    phase: TEST
  }
  reduction_param {
    operation: MEAN
  }
}
layer {
  name: "pred-7"
  type: "Reduction"
  bottom: "pred-7-vector"
  top: "pred-7"
  include {
    phase: TEST
  }
  reduction_param {
    operation: MEAN
  }
}
layer {
  name: "pred-8"
  type: "Reduction"
  bottom: "pred-8-vector"
  top: "pred-8"
  include {
    phase: TEST
  }
  reduction_param {
    operation: MEAN
  }
}
layer {
  name: "pred-9"
  type: "Reduction"
  bottom: "pred-9-vector"
  top: "pred-9"
  include {
    phase: TEST
  }
  reduction_param {
    operation: MEAN
  }
}

Upvotes: 1

Autonomous
Autonomous

Reputation: 9075

Yes. You can. If you have N x p x q x r blob of prediction, the first use Slice (SliceLayer), creating N blobs, each of shape 1 x p x q x r. Then use these N blobs as N bottoms for the eltwise (EltwiseLayer) layer to produce a single top.

Upvotes: 1

Related Questions