caffe中的sgd,與激活函數(activation function)

caffe中activation function的形式，直接決定了其訓練速度以及SGD的求解。

在caffe中，不同的activation function對應的sgd的方式是不同的，因此，在配置文件中指定activation layer的type，目前caffe中用的最多的是relu的activation function.

caffe中，目前實現的activation function有以下幾種：

absval, bnll, power, relu, sigmoid, tanh等幾種，分別有單獨的layer層。其數學公式分別為:

 

算了，這部分我不解釋了，直接看caffe的tutorial吧

ReLU / Rectified-Linear and Leaky-ReLU

• LayerType: RELU
• CPU implementation: ./src/caffe/layers/relu_layer.cpp
• CUDA GPU implementation: ./src/caffe/layers/relu_layer.cu
• Parameters (ReLUParameter relu_param)
  • Optional
    • negative_slope [default 0]: specifies whether to leak the negative part by multiplying it with the slope value rather than setting it to 0.

• Sample (as seen in ./examples/imagenet/imagenet_train_val.prototxt)

  layers {
    name: "relu1"
    type: RELU
    bottom: "conv1"
    top: "conv1"
  }
  

Given an input value x, The RELU layer computes the output as x if x > 0 and negative_slope * x if x <= 0. When the negative slope parameter is not set, it is equivalent to the standard ReLU function of taking max(x, 0). It also supports in-place computation, meaning that the bottom and the top blob could be the same to preserve memory consumption.

Sigmoid

• LayerType: SIGMOID
• CPU implementation: ./src/caffe/layers/sigmoid_layer.cpp
• CUDA GPU implementation: ./src/caffe/layers/sigmoid_layer.cu

• Sample (as seen in ./examples/imagenet/mnist_autoencoder.prototxt)

  layers {
    name: "encode1neuron"
    bottom: "encode1"
    top: "encode1neuron"
    type: SIGMOID
  }
  

The SIGMOID layer computes the output as sigmoid(x) for each input element x.

TanH / Hyperbolic Tangent

• LayerType: TANH
• CPU implementation: ./src/caffe/layers/tanh_layer.cpp
• CUDA GPU implementation: ./src/caffe/layers/tanh_layer.cu

• Sample

  layers {
    name: "layer"
    bottom: "in"
    top: "out"
    type: TANH
  }
  

The TANH layer computes the output as tanh(x) for each input element x.

Absolute Value

• LayerType: ABSVAL
• CPU implementation: ./src/caffe/layers/absval_layer.cpp
• CUDA GPU implementation: ./src/caffe/layers/absval_layer.cu

• Sample

  layers {
    name: "layer"
    bottom: "in"
    top: "out"
    type: ABSVAL
  }
  

The ABSVAL layer computes the output as abs(x) for each input element x.

Power

• LayerType: POWER
• CPU implementation: ./src/caffe/layers/power_layer.cpp
• CUDA GPU implementation: ./src/caffe/layers/power_layer.cu
• Parameters (PowerParameter power_param)
  • Optional
    • power [default 1]
    • scale [default 1]
    • shift [default 0]

• Sample

  layers {
    name: "layer"
    bottom: "in"
    top: "out"
    type: POWER
    power_param {
      power: 1
      scale: 1
      shift: 0
    }
  }
  

The POWER layer computes the output as (shift + scale * x) ^ power for each input element x.

BNLL

• LayerType: BNLL
• CPU implementation: ./src/caffe/layers/bnll_layer.cpp
• CUDA GPU implementation: ./src/caffe/layers/bnll_layer.cu

• Sample

  layers {
    name: "layer"
    bottom: "in"
    top: "out"
    type: BNLL
  }
  

The BNLL (binomial normal log likelihood) layer computes the output as log(1 + exp(x)) for each input element x.