Caffe中的Min-Max规范化层

时间:2016-12-26 07:54:11

标签: c++ neural-network deep-learning normalization caffe

我是caffe的新手,我正在尝试使用Min-Max Normalization将卷积输出归一化到0到1之间。

Out = X - Xmin /(Xmax - Xmin)

我检查过很多层(电源,缩放,批量标准化,MVN),但没有人在图层中给出最小 - 最大标准化输出。谁能帮助我?

*************我的原型文件*****************

name: "normalizationCheck"
layer {
  name: "data"
  type: "Input"
  top: "data"
  input_param { shape: { dim: 1 dim: 1 dim: 512 dim: 512 } }
}

layer {
  name: "normalize1"
  type: "Power"
  bottom: "data"
  top: "normalize1"
  power_param { 
    shift: 0
    scale: 0.00392156862
    power: 1
   }
}

layer {
    bottom: "normalize1"
    top: "Output"
    name: "conv1"
    type: "Convolution"
    convolution_param {
        num_output: 1
        kernel_size: 1
        pad: 0
        stride: 1
        bias_term: false
        weight_filler {
        type: "constant"
        value: 1
        }
    }
}

卷积层输出不是标准化形式我想要层格式的最小 - 最大标准化输出。手动我可以使用代码,但我需要在图层中。 谢谢

1 个答案:

答案 0 :(得分:4)

您可以在these guidelines之后编写自己的c ++图层,您将看到如何实现"仅转发"该页面中的图层。

或者,您可以在python中实现该层,并通过'"Python"' layer在caffe中执行它:

首先,在python中实现您的图层,将其存储在'/path/to/my_min_max_layer.py'

import caffe
import numpy as np

class min_max_forward_layer(caffe.Layer):
  def setup(self, bottom, top):
    # make sure only one input and one output
    assert len(bottom)==1 and len(top)==1, "min_max_layer expects a single input and a single output"

  def reshape(self, bottom, top):
    # reshape output to be identical to input
    top[0].reshape(*bottom[0].data.shape)

  def forward(self, bottom, top):
    # YOUR IMPLEMENTATION HERE!!
    in_ = np.array(bottom[0].data)
    x_min = in_.min()
    x_max = in_.max()
    top[0].data[...] = (in_-x_min)/(x_max-x_min)

  def backward(self, top, propagate_down, bottom):
    # backward pass is not implemented!
    pass

在python中实现图层后,您只需将其添加到网络中(确保'/path/to'中有$PYTHONPATH):

layer {
  name: "my_min_max_forward_layer"
  type: "Python"
  bottom: "name_your_input_here"
  top: "name_your_output_here"
  python_param {
    module: "my_min_max_layer"  # name of python file to be imported
    layer: "min_max_forward_layer" # name of layer class
  }
}