CODEWITHSUNDEEP
caffe
konze
konze

Reputation: 893

Train bvlc_reference_caffenet with input data scaled to [0,1]

I followed the Caffe ImageNet Tutorial and could successfully train the bvlc_reference_caffenet. Now I would like to scale the input data from [0,255] to [0,1] (since I later have to run this CNN on hardware with limited fixed/floating point accuracy) like in the Caffe LeNet MNIST Tutorial where it is done by adding a scale parameter to the data layer:

layer {
  name: "mnist"
  type: "Data"
  transform_param {
    scale: 0.00390625
  }
  data_param {
    source: "mnist_train_lmdb"
    backend: LMDB
    batch_size: 64
  }
  top: "data"
  top: "label"
}

Due to this I added this scale parameter also to the bvlc_reference_caffenet and also divide the mean parameter which will be subtracted from each channel by 255:

layer {
  name: "data"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TRAIN
  }
  transform_param {
     crop_size: 227
     scale: 0.00390625
     mean_value: 0.40784313
     mean_value: 0.45882352
     mean_value: 0.48235294
     mirror: true
  }
  data_param {
    source: "examples/imagenet/ilsvrc12_train_lmdb"
    batch_size: 32
    backend: LMDB
  }
}
layer {
  name: "data"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TEST
  }
  transform_param {
    crop_size: 227
    scale: 0.00390625
    mean_value: 0.40784313
    mean_value: 0.45882352
    mean_value: 0.48235294
    mirror: false
  }
  data_param {
    source: "examples/imagenet/ilsvrc12_val_lmdb"
    batch_size: 32
    backend: LMDB
  }
}

When I train the net now the accuracy will always be lower than chance. What other parameters do I have to adopt to train the net with [0,1] images instead of [0,255] images?

Upvotes: 2

Views: 87

Answers (1)

rkellerm
rkellerm

Reputation: 5522

Caffe first subtracts the mean and then scales the result. You should retain your original 0...255 mean values.
Reference:

if (has_mean_file) {
          transformed_data[top_index] =
            (datum_element - mean[data_index]) * scale;
}

(data_transformer.cpp)

Upvotes: 3

Related Questions