caffe 版本的 non local

曾经终败给现在 2022-05-10 23:52 140阅读 0赞

主要是根据不同模块的组合构建的：

假设输入的特征维数为1024通道的特征：

矩阵乘法和transpose 可以参考下面链接实现：

[https://github.com/simbaforrest/caffe/blob/master/src/caffe/layers/matrix\_multiplication\_layer.cpp][https_github.com_simbaforrest_caffe_blob_master_src_caffe_layers_matrix_multiplication_layer.cpp]

layer {
      name: "G"
      type: "Convolution"
      bottom: "input"
      top: "G"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 1024
        kernel_size: 1
        weight_filler {
          type: "gaussian"
          std: 0.01
        }
        bias_filler {
          type: "constant"
          value: 0.0
        }
      }
    }
    
    
    
    layer {
      name: "Theta"
      type: "Convolution"
      bottom: "input"
      top: "Theta"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 1024
        kernel_size: 1
        weight_filler {
          type: "gaussian"
          std: 0.01
        }
        bias_filler {
          type: "constant"
          value: 0.0
        }
      }
    }
    
    
    layer {
      name: "Phi"
      type: "Convolution"
      bottom: "input"
      top: "Phi"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 1024
        kernel_size: 1
        weight_filler {
          type: "gaussian"
          std: 0.01
        }
        bias_filler {
          type: "constant"
          value: 0.0
        }
      }
    }
    
    
    layer {
        name: "reshape_G"
        type: "Reshape"
        bottom: "G"
        top: "output_G"
        reshape_param {
          shape {
            dim: 0  # copy the dimension from below
            dim: 0
            dim: -1 # infer it from the other dimensions
          }
        }
    }
    
    
    layer {
        name: "reshape_Theta"
        type: "Reshape"
        bottom: "Theta"
        top: "output_Theta"
        reshape_param {
          shape {
            dim: 0  # copy the dimension from below
            dim: 0
            dim: -1 # infer it from the other dimensions
          }
        }
    }
    
    
    layer {
        name: "reshape_Phi"
        type: "Reshape"
        bottom: "Phi"
        top: "output_Phi"
        reshape_param {
          shape {
            dim: 0  # copy the dimension from below
            dim: 0
            dim: -1 # infer it from the other dimensions
          }
        }
    }
    
    layer {
        name: "Transpose_G"
        type: "Transpose"
        bottom: "output_G"
        top: "Transpose_G"
        transpose_param { dim: 0 dim: 2 dim: 1 }
    }
    
    layer {
        name: "Transpose_Theta"
        type: "Transpose"
        bottom: "output_Theta"
        top: "Transpose_Theta"
        transpose_param {dim: 0 dim: 2 dim: 1 }
    }
    
    layer {
        name: "MM_Theta_Phi"
        type: "MatrixMultiplication"
    	#type: "MATRIXMultiplication"
        bottom: "Transpose_Theta"
    	bottom: "output_Phi"
        top: "MM_Theta_Phi"
    }
    
    layer {
      name: "MM_Theta_Phi_soft"
      type: "Softmax"
      bottom: "MM_Theta_Phi"
      top: "MM_Theta_Phi_soft"
      SoftmaxParameter {
         axis:2
      }
    }
    
    
    layer {
        name: "MM_o"
        type: "MatrixMultiplication"
    	#type: "MATRIXMultiplication"
        bottom: "MM_Theta_Phi_soft"
    	bottom: "output_G"
        top: "MM_o"
    }
    
    
    layer {
        name: "Transpose_MM"
        type: "Transpose"
        bottom: "MM_o"
        top: "Transpose_MM"
        transpose_param {dim: 0 dim: 2 dim: 1 }
    }
    
    
    
    layer {
        name: "re_MM"
        type: "Reshape"
        bottom: "Transpose_MM"
        top: "re_MM"
        reshape_param {
          shape {
            dim: 0  # copy the dimension from below
            dim: 0
    		dim: 176 # infer it from the other dimensions
            dim: -1 # infer it from the other dimensions
          }
        }
    }
    
    
    layer {
      name: "conv_re"
      type: "Convolution"
      bottom: "re_MM"
      top: "conv_re"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 1024
        kernel_size: 1
        weight_filler {
          type: "gaussian"
          std: 0.01
        }
        bias_filler {
          type: "constant"
          value: 0.0
        }
      }
    }
    
    
    
    layer {
      name: "G"
      type: "Convolution"
      bottom: "input"
      top: "G"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 1024
        kernel_size: 1
        weight_filler {
          type: "gaussian"
          std: 0.01
        }
        bias_filler {
          type: "constant"
          value: 0.0
        }
      }
    }
    
    
    
    layer {
      name: "Theta"
      type: "Convolution"
      bottom: "input"
      top: "Theta"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 1024
        kernel_size: 1
        weight_filler {
          type: "gaussian"
          std: 0.01
        }
        bias_filler {
          type: "constant"
          value: 0.0
        }
      }
    }
    
    
    layer {
      name: "Phi"
      type: "Convolution"
      bottom: "input"
      top: "Phi"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 1024
        kernel_size: 1
        weight_filler {
          type: "gaussian"
          std: 0.01
        }
        bias_filler {
          type: "constant"
          value: 0.0
        }
      }
    }
    
    
    layer {
        name: "reshape_G"
        type: "Reshape"
        bottom: "G"
        top: "output_G"
        reshape_param {
          shape {
            dim: 0  # copy the dimension from below
            dim: 0
            dim: -1 # infer it from the other dimensions
          }
        }
    }
    
    
    layer {
        name: "reshape_Theta"
        type: "Reshape"
        bottom: "Theta"
        top: "output_Theta"
        reshape_param {
          shape {
            dim: 0  # copy the dimension from below
            dim: 0
            dim: -1 # infer it from the other dimensions
          }
        }
    }
    
    
    layer {
        name: "reshape_Phi"
        type: "Reshape"
        bottom: "Phi"
        top: "output_Phi"
        reshape_param {
          shape {
            dim: 0  # copy the dimension from below
            dim: 0
            dim: -1 # infer it from the other dimensions
          }
        }
    }
    
    layer {
        name: "Transpose_G"
        type: "Transpose"
        bottom: "output_G"
        top: "Transpose_G"
        transpose_param { dim: 0 dim: 2 dim: 1 }
    }
    
    layer {
        name: "Transpose_Theta"
        type: "Transpose"
        bottom: "output_Theta"
        top: "Transpose_Theta"
        transpose_param {dim: 0 dim: 2 dim: 1 }
    }
    
    layer {
        name: "MM_Theta_Phi"
        type: "MatrixMultiplication"
    	#type: "MATRIXMultiplication"
        bottom: "Transpose_Theta"
    	bottom: "output_Phi"
        top: "MM_Theta_Phi"
    }
    
    layer {
      name: "MM_Theta_Phi_soft"
      type: "Softmax"
      bottom: "MM_Theta_Phi"
      top: "MM_Theta_Phi_soft"
      SoftmaxParameter {
         axis:2
      }
    }
    
    
    layer {
        name: "MM_o"
        type: "MatrixMultiplication"
    	#type: "MATRIXMultiplication"
        bottom: "MM_Theta_Phi_soft"
    	bottom: "output_G"
        top: "MM_o"
    }
    
    
    layer {
        name: "Transpose_MM"
        type: "Transpose"
        bottom: "MM_o"
        top: "Transpose_MM"
        transpose_param {dim: 0 dim: 2 dim: 1 }
    }
    
    
    
    layer {
        name: "re_MM"
        type: "Reshape"
        bottom: "Transpose_MM"
        top: "re_MM"
        reshape_param {
          shape {
            dim: 0  # copy the dimension from below
            dim: 0
    		dim: 176 # infer it from the other dimensions
            dim: -1 # infer it from the other dimensions
          }
        }
    }
    
    
    layer {
      name: "conv_re"
      type: "Convolution"
      bottom: "re_MM"
      top: "conv_re"
      param {
        lr_mult: 1
        decay_mult: 1
      }
      param {
        lr_mult: 2
        decay_mult: 0
      }
      convolution_param {
        num_output: 1024
        kernel_size: 1
        weight_filler {
          type: "gaussian"
          std: 0.01
        }
        bias_filler {
          type: "constant"
          value: 0.0
        }
      }
    }
    
    layer{
      name: "ele"
      type: "Eltwise"
      bottom: "input"
      bottom: "conv_re"
      top: "ele_out"
    }

[https_github.com_simbaforrest_caffe_blob_master_src_caffe_layers_matrix_multiplication_layer.cpp]: https://github.com/simbaforrest/caffe/blob/master/src/caffe/layers/matrix_multiplication_layer.cpp