Caffe

Question

我已经在我的数据集上微调了 imagenet 预训练模型，以下是在 train_val.prototxt 中所做的相关更改（另外我还没有进行过采样，只是准备 hdf5 时的中心裁剪）

name: "MyCaffeNet"
layer {
  type: "HDF5Data"
  name: "data"
  top: "X"  
  top: "Meta" 
  top: "Labels"
  hdf5_data_param {
    source: "/path/to/hdf5_train.txt"
    batch_size: 50
  }
  include { phase: TRAIN }
}
layer {
  type: "HDF5Data"
  name: "data"
  top: "X"    
  top: "Meta" 
  top: "Labels"
  hdf5_data_param {
    source: "/path/to/hdf5_test.txt"
    batch_size: 50
 }
  include { phase: TEST }
}
layer {
  name: "conv1"
  type: "Convolution"
  bottom: "X"

其余相同直到

layer {
  name: "concat"
  bottom: "fc7"
  bottom: "Meta"
  top: "combined"
  type: "Concat"
  concat_param {
    concat_dim: 1
  }
}
layer {
  name: "my-fc8"
  type: "InnerProduct"
  bottom: "combined"
  top: "my-fc8"
  # lr_mult is set to higher than for other layers, because this layer is starting from random while the others are already trained
  param {
    lr_mult: 10
    decay_mult: 1
  }
  param {
    lr_mult: 20
    decay_mult: 0
  }
  inner_product_param {
    num_output: 4098
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
      value: 0
    }
 }
}
layer {
  name: "my-fc9"
  type: "InnerProduct"
  bottom: "my-fc8"
  top: "my-fc9"
  # lr_mult is set to higher than for other layers, because this layer is starting from random while the others are already trained
  param {
    lr_mult: 10
    decay_mult: 1
  }
  param {
    lr_mult: 20
    decay_mult: 0
  }
  inner_product_param {
    num_output: 1
    weight_filler {
      type: "gaussian"
      std: 0.01
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "loss"
  type: "EuclideanLoss"
  bottom: "my-fc9"
  bottom: "Labels"
}
layer {
  name: "loss"
  type: "EuclideanLoss"
  bottom: "my-fc9"
  bottom: "Labels"
  top: "loss"
  include {
    phase: TEST
  }
}

问题是我不知道如何修改我的 deploy.prototxt 和我的特征提取代码（如下所示）来测试这个模型，现在使用额外的元信息作为输入特征

for i=1:n
    im = fgetl(file_list);
    im = imread(im);
    input_data = {prepare_image(im)};
    scores = caffe('forward', input_data);
    scores_original=scores;
    scores = scores{1};
    scores = squeeze(scores);
end

此处 prepare_image 将 rgb 转换为 bgr、置换和居中裁剪，基本上是预处理。

所以，归结为我应该如何修改'caffe('forward', input_data)';并制作一个 deploy.prototxt 以便在测试 caffe 时也提供 Meta (n*2) 功能。感谢您的耐心等待和帮助！

我觉得deploy.prototxt开头应该是这样的

name: "MyCaffeNet"
input: "X"
input_dim: 1
input_dim: 3
input_dim: 227
input_dim: 227
input: "Meta"
input_dim: 1
input_dim: 1
input_dim: 1
input_dim: 2 # two additional 'features' per image

Answer 1

变量input_data可以是一个4维斑点向量。函数 prepare_data 给出一个 4D 图像数据块，类似地，我们可以将元数据或任何其他类型的输入充分整形为 4D 块，然后将其传递给前向函数。

input_data = {prepare_data(im),prepare_meta()};

scores = caffe('forward', input_data');

Caffe - 从 train_val.prototxt 创建 deploy.prototxt

Caffe - Create deploy.prototxt from train_val.prototxt

matlab

input

hdf5

conv-neural-network