Excluding Layers: Train and Test Phase

#Excluding Layers

Caffe can be told to include or exclude a layer when running a network in a particular phase.

For examples, when training a model with caffe via caffe train ..., or caffe test ..., we are explicitly telling caffe to train in a particular phase, train and test respectively.

There are times when a network should behave slightly differently depending on whether is being trained or being tested, and thus by setting the appropriate include-phase we can achieve this behaviour without having to define many nearly identical network prototxts.

Example: MNist

The first two layers of the MNist example is presented. Note that the input data and batch size differ depending on the current phase.

name: "LeNet"
layer {
  name: "mnist"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TRAIN
  }
  transform_param {
    scale: 0.00390625
  }
  data_param {
    source: "examples/mnist/mnist_train_lmdb"
    batch_size: 64
    backend: LMDB
  }
}
layer {
  name: "mnist"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TEST
  }
  transform_param {
    scale: 0.00390625
  }
  data_param {
    source: "examples/mnist/mnist_test_lmdb"
    batch_size: 100
    backend: LMDB
  }
}
...