我的小小AI 天地

本教學將教你如何改寫網路

使它可以當成deploy層 讓你取出數據

以下教學皆翻譯來自官網:

https://github.com/BVLC/caffe/wiki/Using-a-Trained-Network:-Deploy

這裡我們以autoencoder 作為例子

首先介紹 autoencoder 整個網路訓練架構:

name: "MNISTAutoencoder"

***************************************************************訓練層
layer {
  name: "data"
  type: "Data"
  top: "data"
  include {
    phase: TRAIN
  }
  transform_param {
    scale: 0.0039215684
  }
  data_param {
    source: "examples/mnist/mnist_train_lmdb"
    batch_size: 100
    backend: LMDB
  }
}

***************************************************************測試層:test-on-train
layer {
  name: "data"
  type: "Data"
  top: "data"
  include {
    phase: TEST
    stage: "test-on-train"
  }
  transform_param {
    scale: 0.0039215684
  }
  data_param {
    source: "examples/mnist/mnist_train1_lmdb"
    batch_size: 100
    backend: LMDB
  }
}

***************************************************************測試層:test-on-test
layer {
  name: "data"
  type: "Data"
  top: "data"
  include {
    phase: TEST
    stage: "test-on-test"
  }
  transform_param {
    scale: 0.0039215684
  }
  data_param {
    source: "examples/mnist/mnist_test_lmdb"
    batch_size: 100
    backend: LMDB
  }
}

***************************************************************攤平資料
layer {
  name: "flatdata"
  type: "Flatten"
  bottom: "data"
  top: "flatdata"
}

**********************************************從資料層到第一層encoder           output:1000
layer {
  name: "encode1"
  type: "InnerProduct"
  bottom: "data"
  top: "encode1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  inner_product_param {
    num_output: 1000
    weight_filler {
      type: "gaussian"
      std: 1
      sparse: 15
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

***************************************************************從encoder1-->encoder1neuron
layer {
  name: "encode1neuron"
  type: "Sigmoid"
  bottom: "encode1"
  top: "encode1neuron"
}

**********************************************從encoder1neuron-->encoder2        output:500
layer {
  name: "encode2"
  type: "InnerProduct"
  bottom: "encode1neuron"
  top: "encode2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  inner_product_param {
    num_output: 500
    weight_filler {
      type: "gaussian"
      std: 1
      sparse: 15
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

***************************************************************從encoder2-->encoder2neuron
layer {
  name: "encode2neuron"
  type: "Sigmoid"
  bottom: "encode2"
  top: "encode2neuron"
}

**********************************************從encoder2neuron-->encoder3        output:250
layer {
  name: "encode3"
  type: "InnerProduct"
  bottom: "encode2neuron"
  top: "encode3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  inner_product_param {
    num_output: 250
    weight_filler {
      type: "gaussian"
      std: 1
      sparse: 15
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

***************************************************************從encoder3-->encoder3neuron
layer {
  name: "encode3neuron"
  type: "Sigmoid"
  bottom: "encode3"
  top: "encode3neuron"
}

**********************************************從encoder3neuron-->encoder4          output:30
layer {
  name: "encode4"
  type: "InnerProduct"
  bottom: "encode3neuron"
  top: "encode4"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  inner_product_param {
    num_output: 30
    weight_filler {
      type: "gaussian"
      std: 1
      sparse: 15
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

************************************************************************************reverse

***************************************************************從encode4-->decode4
layer {
  name: "decode4"
  type: "InnerProduct"
  bottom: "encode4"
  top: "decode4"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  inner_product_param {
    num_output: 250
    weight_filler {
      type: "gaussian"
      std: 1
      sparse: 15
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

***************************************************************從decode4-->decoder4neuron
layer {
  name: "decode4neuron"
  type: "Sigmoid"
  bottom: "decode4"
  top: "decode4neuron"
}

***************************************************************從decoder4neuron-->decoder3
layer {
  name: "decode3"
  type: "InnerProduct"
  bottom: "decode4neuron"
  top: "decode3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  inner_product_param {
    num_output: 500
    weight_filler {
      type: "gaussian"
      std: 1
      sparse: 15
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

***************************************************************從decode3-->decoder3neuron
layer {
  name: "decode3neuron"
  type: "Sigmoid"
  bottom: "decode3"
  top: "decode3neuron"
}

***************************************************************從decoder3neuron-->decoder2
layer {
  name: "decode2"
  type: "InnerProduct"
  bottom: "decode3neuron"
  top: "decode2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  inner_product_param {
    num_output: 1000
    weight_filler {
      type: "gaussian"
      std: 1
      sparse: 15
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

***************************************************************從decoder2-->decoder2neuron
layer {
  name: "decode2neuron"
  type: "Sigmoid"
  bottom: "decode2"
  top: "decode2neuron"
}

***************************************************************從decoder2neuron-->decoder1
layer {
  name: "decode1"
  type: "InnerProduct"
  bottom: "decode2neuron"
  top: "decode1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  inner_product_param {
    num_output: 784
    weight_filler {
      type: "gaussian"
      std: 1
      sparse: 15
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

***************************************************************
layer {
  name: "loss"
  type: "SigmoidCrossEntropyLoss"
  bottom: "decode1"
  bottom: "flatdata"
  top: "cross_entropy_loss"
  loss_weight: 1
}

***************************************************************從decoder1-->decoder1neuron
layer {
  name: "decode1neuron"
  type: "Sigmoid"
  bottom: "decode1"
  top: "decode1neuron"
}

***************************************************************
layer {
  name: "loss"
  type: "EuclideanLoss"
  bottom: "decode1neuron"
  bottom: "flatdata"
  top: "l2_error"
  loss_weight: 0
}

***************************************************************end

下面說明如何將上面的訓練網路改寫成deploy 層

第一步: 移除所有data 層

第二步: 移除所有與data 層 的相關層

第三步: 建造 input 層

由於我們只需要encode後的資訊

decode只是訓練時需要

所以完整網路如下:

name: "MNISTAutoencoder_half"

***************************************************************
layer {
  name: "data"
  type: "Input"
  top: "data"
  input_param { shape: { dim: 1 dim: 1 dim: 28 dim: 28 } }
}

***************************************************************

layer {
  name: "encode1"
  type: "InnerProduct"
  bottom: "data"
  top: "encode1"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  inner_product_param {
    num_output: 1000
    weight_filler {
      type: "gaussian"
      std: 1
      sparse: 15
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

***************************************************************
layer {
  name: "encode1neuron"
  type: "Sigmoid"
  bottom: "encode1"
  top: "encode1neuron"
}

***************************************************************
layer {
  name: "encode2"
  type: "InnerProduct"
  bottom: "encode1neuron"
  top: "encode2"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  inner_product_param {
    num_output: 500
    weight_filler {
      type: "gaussian"
      std: 1
      sparse: 15
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

***************************************************************
layer {
  name: "encode2neuron"
  type: "Sigmoid"
  bottom: "encode2"
  top: "encode2neuron"
}

***************************************************************
layer {
  name: "encode3"
  type: "InnerProduct"
  bottom: "encode2neuron"
  top: "encode3"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  inner_product_param {
    num_output: 250
    weight_filler {
      type: "gaussian"
      std: 1
      sparse: 15
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

***************************************************************
layer {
  name: "encode3neuron"
  type: "Sigmoid"
  bottom: "encode3"
  top: "encode3neuron"
}

***************************************************************
layer {
  name: "encode4"
  type: "InnerProduct"
  bottom: "encode3neuron"
  top: "encode4"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 1
    decay_mult: 0
  }
  inner_product_param {
    num_output: 30
    weight_filler {
      type: "gaussian"
      std: 1
      sparse: 15
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}

***************************************************************必須再另外加上這層
layer {
  name: "encode4neuron"
  type: "Sigmoid"
  bottom: "encode4"
  top: "encode4neuron"
}