machine-learning - 使用尺度参数训练 Caffe CNN

Question

我改编了 bvlc_reference_caffenet 中的 train_val.prototxt 以在 Caffe 中实现 VGG-16 克隆，并能够使用 batch_size: 6 和 base_lr 使用 GTX 1050 对其进行训练: 0.0648(~ 0.01 * sqrt(256/6)~ 0.01 * sqrt(42))。但是，我想将输入数据从 [0;255] 缩放到 [0;1]，因为该 CNN 的目标平台的精度有限。为了缩放数据，我引入了 scale: 0.00390625 参数(这取自 Caffe LeNet 示例，该示例在目标平台上运行良好)。但使用 scale 参数时，准确度不会增加(> 40000 次迭代)，并且训练过程中损失也不会改变。

如何使用 scale 参数训练此 CNN？

train_val.prototxt

name: "ES VGG"
layer {
  name: "data"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TRAIN
  }
  transform_param {
    scale: 0.00390625
    mirror: true
    crop_size: 224
    mean_file: "/local/datasets/imagenet/ilsvrc12/imagenet_mean.binaryproto"
  }
  data_param {
    source: "/local/datasets/imagenet/ilsvrc12_train_lmdb"
    batch_size: 6
    backend: LMDB
  }
}
layer {
  name: "data"
  type: "Data"
  top: "data"
  top: "label"
  include {
    phase: TEST
  }
  transform_param {
    scale: 0.00390625
    mirror: false
    crop_size: 224
    mean_file: "/local/datasets/imagenet/ilsvrc12/imagenet_mean.binaryproto"
  }
  data_param {
    source: "/local/datasets/imagenet/ilsvrc12_val_lmdb"
    batch_size: 6
    backend: LMDB
  }
}
layer {
  name: "conv1_1"
  type: "Convolution"
  bottom: "data"
  top: "conv1_1"
  convolution_param {
    num_output: 64
    kernel_size: 3
    pad: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu1_1"
  type: "ReLU"
  bottom: "conv1_1"
  top: "conv1_1"
}
layer {
  name: "conv1_2"
  type: "Convolution"
  bottom: "conv1_1"
  top: "conv1_2"
  convolution_param {
    num_output: 64
    kernel_size: 3
    pad: 1
    weight_filler {
    type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu1_2"
  type: "ReLU"
  bottom: "conv1_2"
  top: "conv1_2"
}
layer {
  name: "pool1"
  type: "Pooling"
  bottom: "conv1_2"
  top: "pool1"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv2_1"
  type: "Convolution"
  bottom: "pool1"
  top: "conv2_1"
  convolution_param {
    num_output: 128
    kernel_size: 3
    pad: 1
    weight_filler {
    type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu2_1"
  type: "ReLU"
  bottom: "conv2_1"
  top: "conv2_1"
}
layer {
  name: "conv2_2"
  type: "Convolution"
  bottom: "conv2_1"
  top: "conv2_2"
  convolution_param {
    num_output: 128
    kernel_size: 3
    pad: 1
    weight_filler {
    type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu2_2"
  type: "ReLU"
  bottom: "conv2_2"
  top: "conv2_2"
}
layer {
  name: "pool2"
  type: "Pooling"
  bottom: "conv2_2"
  top: "pool2"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv3_1"
  type: "Convolution"
  bottom: "pool2"
  top: "conv3_1"
  convolution_param {
    num_output: 256
    kernel_size: 3
    pad: 1
    weight_filler {
    type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu3_1"
  type: "ReLU"
  bottom: "conv3_1"
  top: "conv3_1"
}
layer {
  name: "conv3_2"
  type: "Convolution"
  bottom: "conv3_1"
  top: "conv3_2"
  convolution_param {
    num_output: 256
    kernel_size: 3
    pad: 1
    weight_filler {
    type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu3_2"
  type: "ReLU"
  bottom: "conv3_2"
  top: "conv3_2"
}
layer {
  name: "conv3_3"
  type: "Convolution"
  bottom: "conv3_2"
  top: "conv3_3"
  convolution_param {
    num_output: 256
    kernel_size: 3
    pad: 1
    weight_filler {
    type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu3_3"
  type: "ReLU"
  bottom: "conv3_3"
  top: "conv3_3"
}
layer {
  name: "pool3"
  type: "Pooling"
  bottom: "conv3_3"
  top: "pool3"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv4_1"
  type: "Convolution"
  bottom: "pool3"
  top: "conv4_1"
  convolution_param {
    num_output: 512
    kernel_size: 3
    pad: 1
    weight_filler {
    type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu4_1"
  type: "ReLU"
  bottom: "conv4_1"
  top: "conv4_1"
}
layer {
  name: "conv4_2"
  type: "Convolution"
  bottom: "conv4_1"
  top: "conv4_2"
  convolution_param {
    num_output: 512
    kernel_size: 3
    pad: 1
    weight_filler {
    type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu4_2"
  type: "ReLU"
  bottom: "conv4_2"
  top: "conv4_2"
}
layer {
  name: "conv4_3"
  type: "Convolution"
  bottom: "conv4_2"
  top: "conv4_3"
  convolution_param {
    num_output: 512
    kernel_size: 3
    pad: 1
    weight_filler {
    type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu4_3"
  type: "ReLU"
  bottom: "conv4_3"
  top: "conv4_3"
}
layer {
  name: "pool4"
  type: "Pooling"
  bottom: "conv4_3"
  top: "pool4"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv5_1"
  type: "Convolution"
  bottom: "pool4"
  top: "conv5_1"
  convolution_param {
    num_output: 512
    kernel_size: 3
    pad: 1
    weight_filler {
    type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu5_1"
  type: "ReLU"
  bottom: "conv5_1"
  top: "conv5_1"
}
layer {
  name: "conv5_2"
  type: "Convolution"
  bottom: "conv5_1"
  top: "conv5_2"
  convolution_param {
    num_output: 512
    kernel_size: 3
    pad: 1
    weight_filler {
    type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "relu5_2"
  type: "ReLU"
  bottom: "conv5_2"
  top: "conv5_2"
}
layer {
  name: "conv5_3"
  type: "Convolution"
  bottom: "conv5_2"
  top: "conv5_3"
  convolution_param {
    num_output: 512
    kernel_size: 3
    pad: 1
    weight_filler {
    type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "pool5"
  type: "Pooling"
  bottom: "conv5_3"
  top: "pool5"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "fc6"
  type: "InnerProduct"
  bottom: "pool5"
  top: "fc6"
  inner_product_param {
    num_output: 4096
    weight_filler {
    type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0.01
    }
  }
}
layer {
  name: "relu6"
  type: "ReLU"
  bottom: "fc6"
  top: "fc6"
}
layer {
  name: "drop6"
  type: "Dropout"
  bottom: "fc6"
  top: "fc6"
  dropout_param {
    dropout_ratio: 0.5
  }
}
layer {
  name: "fc7"
  type: "InnerProduct"
  bottom: "fc6"
  top: "fc7"
  inner_product_param {
    num_output: 4096
    weight_filler {
    type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0.01
    }
  }
}
layer {
  name: "relu7"
  type: "ReLU"
  bottom: "fc7"
  top: "fc7"
}
layer {
  name: "drop7"
  type: "Dropout"
  bottom: "fc7"
  top: "fc7"
  dropout_param {
    dropout_ratio: 0.5
  }
}
layer {
  name: "fc8"
  type: "InnerProduct"
  bottom: "fc7"
  top: "fc8"
  inner_product_param {
    num_output: 1000
    weight_filler {
    type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0.01
    }
  }
}
layer {
  name: "loss"
  type: "SoftmaxWithLoss"
  bottom: "fc8"
  bottom: "label"
  top: "loss"
}
layer {
  name: "accuracytop1"
  type: "Accuracy"
  bottom: "fc8"
  bottom: "label"
  top: "accuracytop1"
  accuracy_param {
    top_k: 1
  }
  include {
    phase: TEST
  }
}
layer {
  name: "accuracytop5"
  type: "Accuracy"
  bottom: "fc8"
  bottom: "label"
  top: "accuracytop5"
  accuracy_param {
    top_k: 5
  }
  include {
    phase: TEST
  }
}

求解器.prototxt

net: "models/es_vgg/train_val.prototxt"
test_iter: 1000
test_interval: 1000
base_lr: 0.0648
lr_policy: "step"
gamma: 0.1
stepsize: 100000
display: 20
max_iter: 18900000
momentum: 0.9
weight_decay: 0.0005
snapshot: 10000
snapshot_prefix: "models/es_vgg/es_vgg_train"
solver_mode: GPU

Answer 1

如果您将输入除以 255，则需要将第一个卷积层 "conv1_1" 的权重乘以 255 以补偿此更改。
看net surgery看看如何做到这一点。

例如(在Python中):

import caffe
net = caffe.Net('models/es_vgg/train_val.prototxt', caffe.TEST)  # no .caffemodel weights supplied - weights are randomly init
# scale kernels of first conv layer by 255
net.params['conv1_1'][0].data[...] = 255. * net.params['conv1_1'][0].data
# save the scaled weights
net.save('models/es_vgg/init_scaled.caffemodel')

现在您需要使用'models/es_vgg/init_scaled.caffemodel'开始训练。