Siamese with Sigmoid Cross Entropy Loss, conversely to Contrastive Divergence

[Alessandro Ferrari]

Hi, I am attempting to build a siamese network, that similarly to "Siamese Neural Networks for One-Shot Image Recognition" by Gregory Koch, 2015 (http://www.cs.utoronto.ca/~gkoch/files/msc-thesis.pdf), Chapter 3. page 8, instead of using a Contrastive Divergence in the siamese, uses a sigmoid cross-entropy objective.

The distance function is computed as
p = sigmoid{ SUM_j[ a_j * abs( feat_j - featp_j ) ]}

that is in the cross-entropy loss with the label sim, 1 if the pair is genuine, 0 if it is an impostoir.

I have defined the following way:

name: "mnist_siamese_train_test"
layer {
  name: "pair_data"
  type: "Data"
  top: "pair_data"
  top: "sim"
  include {
    phase: TRAIN
  }
  transform_param {
    scale: 0.00390625
  }
  data_param {
    source: "/home/imaging/new_caffe/examples/siamese/mnist_siamese_train_leveldb"
    batch_size: 64
  }
}
layer {
  name: "pair_data"
  type: "Data"
  top: "pair_data"
  top: "sim"
  include {
    phase: TEST
  }
  transform_param {
    scale: 0.00390625
  }
  data_param {
    source: "/home/imaging/new_caffe/examples/siamese/mnist_siamese_test_leveldb"
    batch_size: 100
  }
}
layer {
  name: "slice_pair"
  type: "Slice"
  bottom: "pair_data"
  top: "data"
  top: "data_p"
  slice_param {
    slice_dim: 1
    slice_point: 1
  }
}
layer {
  name: "conv1"
  type: "Convolution"
  bottom: "data"
  top: "conv1"
  param {
    name: "conv1_w"
    lr_mult: 1
  }
  param {
    name: "conv1_b"
    lr_mult: 2
  }
  convolution_param {
    num_output: 20
    kernel_size: 5
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "pool1"
  type: "Pooling"
  bottom: "conv1"
  top: "pool1"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv2"
  type: "Convolution"
  bottom: "pool1"
  top: "conv2"
  param {
    name: "conv2_w"
    lr_mult: 1
  }
  param {
    name: "conv2_b"
    lr_mult: 2
  }
  convolution_param {
    num_output: 50
    kernel_size: 5
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "pool2"
  type: "Pooling"
  bottom: "conv2"
  top: "pool2"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "ip1"
  type: "InnerProduct"
  bottom: "pool2"
  top: "ip1"
  param {
    name: "ip1_w"
    lr_mult: 1
  }
  param {
    name: "ip1_b"
    lr_mult: 2
  }
  inner_product_param {
    num_output: 500
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "relu1"
  type: "ReLU"
  bottom: "ip1"
  top: "ip1"
}
layer {
  name: "ip2"
  type: "InnerProduct"
  bottom: "ip1"
  top: "ip2"
  param {
    name: "ip2_w"
    lr_mult: 1
  }
  param {
    name: "ip2_b"
    lr_mult: 2
  }
  inner_product_param {
    num_output: 10
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "feat"
  type: "InnerProduct"
  bottom: "ip2"
  top: "feat"
  param {
    name: "feat_w"
    lr_mult: 1
  }
  param {
    name: "feat_b"
    lr_mult: 2
  }
  inner_product_param {
    num_output: 2
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "conv1_p"
  type: "Convolution"
  bottom: "data_p"
  top: "conv1_p"
  param {
    name: "conv1_w"
    lr_mult: 1
  }
  param {
    name: "conv1_b"
    lr_mult: 2
  }
  convolution_param {
    num_output: 20
    kernel_size: 5
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "pool1_p"
  type: "Pooling"
  bottom: "conv1_p"
  top: "pool1_p"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "conv2_p"
  type: "Convolution"
  bottom: "pool1_p"
  top: "conv2_p"
  param {
    name: "conv2_w"
    lr_mult: 1
  }
  param {
    name: "conv2_b"
    lr_mult: 2
  }
  convolution_param {
    num_output: 50
    kernel_size: 5
    stride: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "pool2_p"
  type: "Pooling"
  bottom: "conv2_p"
  top: "pool2_p"
  pooling_param {
    pool: MAX
    kernel_size: 2
    stride: 2
  }
}
layer {
  name: "ip1_p"
  type: "InnerProduct"
  bottom: "pool2_p"
  top: "ip1_p"
  param {
    name: "ip1_w"
    lr_mult: 1
  }
  param {
    name: "ip1_b"
    lr_mult: 2
  }
  inner_product_param {
    num_output: 500
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "relu1_p"
  type: "ReLU"
  bottom: "ip1_p"
  top: "ip1_p"
}
layer {
  name: "ip2_p"
  type: "InnerProduct"
  bottom: "ip1_p"
  top: "ip2_p"
  param {
    name: "ip2_w"
    lr_mult: 1
  }
  param {
    name: "ip2_b"
    lr_mult: 2
  }
  inner_product_param {
    num_output: 10
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "feat_p"
  type: "InnerProduct"
  bottom: "ip2_p"
  top: "feat_p"
  param {
    name: "feat_w"
    lr_mult: 1
  }
  param {
    name: "feat_b"
    lr_mult: 2
  }
  inner_product_param {
    num_output: 2
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
    }
  }
}
layer {
  name: "diff"
  type: "Eltwise"
  bottom: "feat"
  bottom: "feat_p"
  top: "diff"
  eltwise_param {
    operation: SUM
    coeff: 1
    coeff: -1
  }
}
layer {
  name: "absdiff"
  bottom: "diff"
  top: "absdiff"
  type: "AbsVal"
}
layer {
  name: "dist"
  type: "InnerProduct"
  bottom: "absdiff"
  top: "dist"
  param {
    name: "distfeat_w"
    lr_mult: 1
  }
  param {
    name: "distfeat_b"
    lr_mult: 0
  }
  inner_product_param {
    num_output: 1
    weight_filler {
      type: "xavier"
    }
    bias_filler {
      type: "constant"
      value: 0
    }
  }
}
layer {
  name: "loss"
  type: "SigmoidCrossEntropyLoss"
  bottom: "sim"
  bottom: "dist"
  top: "loss"
}

However, I get this error:

sigmoid_cross_entropy_loss_layer.cu:13] SigmoidCrossEntropyLoss Layer cannot backpropagate to label inputs.

Anybody has any hints on how to fix it?

Thank you

[mehdi Noroozi]

Hi,

I have the same problem. Would be grateful if you could let me know in the case that you have found the solution.

[zhiqiang Tao]

Hi Alessandro,

I also have this problem, and it seems shutting down the "propagate_down" of label bottom works for me.  

See https://groups.google.com/forum/#!searchin/caffe-users/backpropagate$20to$20label$20data/caffe-users/0_8t7ZI83tI/discussion 

for more details.

best