caffe在別人修改的層中，加入了個全連接層，並且設置爲共享參數，訓練出的模型中只有該層的名稱，沒有具體的tensor

問題描述：在caffe中，做的是強化學習image caption，想要實現在網絡中多加一個全連接層（一共三個net要加，訓練baseline的net，訓練強化學習的scstnet，以及兩者共用的decode_net）。在訓練baseline的時候加入了全連接層。網絡可以訓練出加入那層的參數，並且decode的時候也能產生結果。但是在訓練強化學習的scstnet訓練能夠訓練處模型，不會報錯。但是在decode的過程中發現添加的那層全連接層並沒有傳入值，導致了
的問題。問題顯示加載的模型參數傳入不進decode網絡中。對模型參數進行打印，顯示加入的那層只有名稱，並沒有具體的數值（是訓練沒有產生，還是沒有保存？）!其中fc_atten是我添加的全連接層的層名。可以看到打印結果只有名稱，沒有參數的shape。其他參數由於都是共享參數，所以第一個節點都產生了具體的tensor。
2，相關代碼。
在訓練baseline的net和decode結構基本相同
```
layer {
  name: "fc_atten_0"
  type: "InnerProduct"
  bottom: "lstm0_hidden0"
  top: "fc_atten_0"
  param {
    name: "fc_param_0"
  }
  inner_product_param {
    num_output: 1000
    bias_term: true
    weight_filler {
      type: "gaussian"
      std: 0.00999999977648
    }
  }
}
```
在訓練scst的部分，也是這樣加入的。但scst和net有點不一樣的是scst多了個beamsearch層，並且加載了net訓練的模型參數（兩者結構基本一樣）。由於我訓練的是語言模型，一共20個lstm的時間步。在作者的beamsearch層中，它將第一個step的全部代碼放到該層中，後面的20個時間步都是共享該層中的參數。
```
layer {
  name: "beam"
  type: "BeamSearch"
  bottom: "num_boxes"
  bottom: "spatial_features"
  bottom: "fc"
  bottom: "context"
  top: "caption"
  top: "log_prob"
  top: "log_prob_sequence"
  param {
    name: "embed_param"
  }
  param {
    name: "lstm0_param_0"
  }
  param {
    name: "lstm0_param_1"
  }
  param {
    name: "hidden_att_param_0"
  }
  param {
    name: "predict_att_param_0"
  }
  param {
    name: "lstm1_param_0"
  }
  param {
    name: "lstm1_param_1"
  }
  param {
    name: "fc_param_0"
  }
  param {
    name: "predict_param_0"
  }
  param {
    name: "predict_param_1"
  }
  beam_search_param {
    net_param {
      layer {
        name: "input"
        type: "Input"
        top: "num_boxes"
        top: "spatial_features"
        top: "fc"
        top: "context"
        top: "input"
        input_param {
          shape {
            dim: 12
            dim: 1
          }
          shape {
            dim: 12
            dim: 100
            dim: 2048
          }
          shape {
            dim: 12
            dim: 100
            dim: 512
          }
          shape {
            dim: 12
            dim: 2048
          }
          shape {
            dim: 12
            dim: 1
          }
        }
      }
      layer {
        name: "lstm0_hidden_prev"
        type: "DummyData"
        top: "lstm0_hidden_prev"
        dummy_data_param {
          shape {
            dim: 12
            dim: 1000
          }
        }
      }
      layer {
        name: "lstm0_mem_cell_prev"
        type: "DummyData"
        top: "lstm0_mem_cell_prev"
        dummy_data_param {
          shape {
            dim: 12
            dim: 1000
          }
        }
      }
      layer {
        name: "lstm1_hidden_prev"
        type: "DummyData"
        top: "lstm1_hidden_prev"
        dummy_data_param {
          shape {
            dim: 12
            dim: 1000
          }
        }
      }
      layer {
        name: "lstm1_mem_cell_prev"
        type: "DummyData"
        top: "lstm1_mem_cell_prev"
        dummy_data_param {
          shape {
            dim: 12
            dim: 1000
          }
        }
      }
      layer {
        name: "embedding"
        type: "Embed"
        bottom: "input"
        top: "embedding"
        param {
          name: "embed_param"
        }
        propagate_down: false
        embed_param {
          num_output: 1000
          input_dim: 10010
          bias_term: false
          weight_filler {
            type: "gaussian"
            std: 0.00999999977648
          }
        }
      }
      layer {
        name: "concat0_t0"
        type: "Concat"
        bottom: "embedding"
        bottom: "context"
        bottom: "lstm1_hidden_prev"
        bottom: "lstm0_hidden_prev"
        top: "concat0_t0"
      }
      layer {
        name: "lstm1"
        type: "LSTMNode"
        bottom: "concat0_t0"
        bottom: "lstm0_mem_cell_prev"
        top: "lstm0_hidden0"
        top: "lstm0_mem_cell0"
        param {
          name: "lstm0_param_0"
        }
        param {
          name: "lstm0_param_1"
        }
        propagate_down: true
        propagate_down: false
        lstm_param {
          num_cells: 1000
          input_weight_filler {
            type: "gaussian"
            std: 0.00999999977648
          }
          input_gate_weight_filler {
            type: "gaussian"
            std: 0.00999999977648
          }
          forget_gate_weight_filler {
            type: "gaussian"
            std: 0.00999999977648
          }
          output_gate_weight_filler {
            type: "gaussian"
            std: 0.00999999977648
          }
          input_bias_filler {
            type: "constant"
            value: 0.0
          }
          input_gate_bias_filler {
            type: "constant"
            value: 0.0
          }
          forget_gate_bias_filler {
            type: "constant"
            value: 1.0
          }
          output_gate_bias_filler {
            type: "constant"
            value: 0.0
          }
        }
      }
      layer {
        name: "hidden_att_0"
        type: "InnerProduct"
        bottom: "lstm0_hidden0"
        top: "hidden_att_0"
        param {
          name: "hidden_att_param_0"
        }
        inner_product_param {
          num_output: 512
          bias_term: false
          weight_filler {
            type: "gaussian"
            std: 0.00999999977648
          }
        }
      }
      layer {
        name: "tile_hidden_att_0"
        type: "Tile"
        bottom: "hidden_att_0"
        top: "tile_hidden_att_0"
        tile_param {
          axis: 1
          tiles: 100
        }
      }
      layer {
        name: "tile_hidden_reshape_0"
        type: "Reshape"
        bottom: "tile_hidden_att_0"
        top: "tile_hidden_reshape_0"
        reshape_param {
          shape {
            dim: 0
            dim: -1
            dim: 512
          }
        }
      }
      layer {
        name: "sum_hidden_att_0"
        type: "Eltwise"
        bottom: "fc"
        bottom: "tile_hidden_reshape_0"
        top: "sum_hidden_att_0"
        eltwise_param {
          operation: SUM
        }
      }
      layer {
        name: "hidden_tanh_0"
        type: "TanH"
        bottom: "sum_hidden_att_0"
        top: "sum_hidden_att_0"
      }
      layer {
        name: "predict_att_0"
        type: "InnerProduct"
        bottom: "sum_hidden_att_0"
        top: "predict_att_0"
        param {
          name: "predict_att_param_0"
        }
        inner_product_param {
          num_output: 1
          bias_term: false
          weight_filler {
            type: "gaussian"
            std: 0.00999999977648
          }
          axis: 2
        }
      }
      layer {
        name: "reshape_predict_att_0"
        type: "Reshape"
        bottom: "predict_att_0"
        top: "reshape_predict_att_0"
        reshape_param {
          shape {
            dim: 0
            dim: -1
          }
        }
      }
      layer {
        name: "att_weight_0"
        type: "Softmax"
        bottom: "reshape_predict_att_0"
        bottom: "num_boxes"
        top: "att_weight_0"
        softmax_param {
          engine: CAFFE
          axis: 1
        }
      }
      layer {
        name: "att_product_0"
        type: "Scale"
        bottom: "spatial_features"
        bottom: "att_weight_0"
        top: "att_product_0"
        scale_param {
          axis: 0
        }
      }
      layer {
        name: "permute_att_0"
        type: "Permute"
        bottom: "att_product_0"
        top: "permute_att_0"
        permute_param {
          order: 0
          order: 2
          order: 1
        }
      }
      layer {
        name: "fc8_0"
        type: "Reduction"
        bottom: "permute_att_0"
        top: "fc8_0"
        reduction_param {
          axis: 2
        }
      }
      layer {
        name: "concat1_t0"
        type: "Concat"
        bottom: "lstm0_hidden0"
        bottom: "fc8_0"
        bottom: "lstm1_hidden_prev"
        top: "concat1_t0"
      }
      layer {
        name: "lstm2"
        type: "LSTMNode"
        bottom: "concat1_t0"
        bottom: "lstm1_mem_cell_prev"
        top: "lstm1_hidden0"
        top: "lstm1_mem_cell0"
        param {
          name: "lstm1_param_0"
        }
        param {
          name: "lstm1_param_1"
        }
        propagate_down: true
        propagate_down: false
        lstm_param {
          num_cells: 1000
          input_weight_filler {
            type: "gaussian"
            std: 0.00999999977648
          }
          input_gate_weight_filler {
            type: "gaussian"
            std: 0.00999999977648
          }
          forget_gate_weight_filler {
            type: "gaussian"
            std: 0.00999999977648
          }
          output_gate_weight_filler {
            type: "gaussian"
            std: 0.00999999977648
          }
          input_bias_filler {
            type: "constant"
            value: 0.0
          }
          input_gate_bias_filler {
            type: "constant"
            value: 0.0
          }
          forget_gate_bias_filler {
            type: "constant"
            value: 1.0
          }
          output_gate_bias_filler {
            type: "constant"
            value: 0.0
          }
        }
      }
      layer {
        name: "fc_atten_0"
        type: "InnerProduct"
        bottom: "lstm0_hidden0"
        top: "fc_atten_0"
        param {
          name: "fc_param_0"
        }
        inner_product_param {
          num_output: 1000
          bias_term: true
          weight_filler {
            type: "gaussian"
            std: 0.00999999977648
          }
        }
      }
      layer {
        name: "concat_lstm0_lstm1_0"
        type: "Concat"
        bottom: "lstm1_hidden0"
        bottom: "fc_atten_0"
        top: "concat_lstm0_lstm1_0"
        }
      layer {
        name: "predict"
        type: "InnerProduct"
        bottom: "concat_lstm0_lstm1_0"
        top: "predict"
        param {
          name: "predict_param_0"
          lr_mult: 1.0
          decay_mult: 1.0
        }
        param {
          name: "predict_param_1"
          lr_mult: 2.0
          decay_mult: 0.0
        }
        inner_product_param {
          num_output: 10010
          weight_filler {
            type: "gaussian"
            std: 0.00999999977648
          }
          bias_filler {
            type: "constant"
            value: 0.0
          }
          axis: 1
        }
      }
      layer {
        name: "probs_0"
        type: "Softmax"
        bottom: "predict"
        top: "probs_0"
        softmax_param {
          axis: 1
        }
      }
      layer {
        name: "logp_0"
        type: "Log"
        bottom: "probs_0"
        top: "logp_0"
      }
    }
    sequence_length: 20
    beam_size: 5
    end_of_sequence: 0
    recurrent_connection {
      src: "lstm0_hidden0"
      dest: "lstm0_hidden_prev"
    }
    recurrent_connection {
      src: "lstm0_mem_cell0"
      dest: "lstm0_mem_cell_prev"
    }
    recurrent_connection {
      src: "lstm1_hidden0"
      dest: "lstm1_hidden_prev"
    }
    recurrent_connection {
      src: "lstm1_mem_cell0"
      dest: "lstm1_mem_cell_prev"
    }
    beam_search_connection {
      src: "logp_0"
      dest: "input"
    }
    allowed_multiple: 2
    allowed_multiple: 5
    allowed_multiple: 4
    allowed_multiple: 15
    allowed_multiple: 3
    allowed_multiple: 6
    allowed_multiple: 8
    allowed_multiple: 7
    allowed_multiple: 9
    allowed_multiple: 13
    allowed_multiple: 277
    allowed_multiple: 11
    allowed_multiple: 30
    allowed_multiple: 16
    allowed_multiple: 19
    allowed_multiple: 27
    allowed_multiple: 25
    allowed_multiple: 119
    allowed_multiple: 48
  }
}
layer {
  name: "silence_bs"
  type: "Silence"
  bottom: "log_prob"
  bottom: "log_prob_sequence"
}
```
其中fc_atten是我仿造該網絡加入的層。並且相同設置了共享參數，並且把參數名放到了前面。以下兩部分就是我在別人代碼上作的修改
```
param {
    name: "fc_param_0"
  }
```
```
 layer {
        name: "fc_atten_0"
        type: "InnerProduct"
        bottom: "lstm0_hidden0"
        top: "fc_atten_0"
        param {
          name: "fc_param_0"
        }
        inner_product_param {
          num_output: 1000
          bias_term: true
          weight_filler {
            type: "gaussian"
            std: 0.00999999977648
          }
        }
      }
```
其他都沒有改變。現在就是訓練出來不會報錯，但是模型中有該層fc_atten的名稱，沒有對應的tensor。
3，試過的方法，把寫的beam_search.cpp 和hpp都看過了，以及caffe.proto也看過修改過試着得到解決辦法。但是可能是作者寫的太nb自己水平不夠。
這裏貼上源代碼的連接 https://github.com/peteanderson80/caffe/tree/631806541c68658248ffdbbbde659f478fac4113
我考慮的是，1，作者是否把beam_search 傳入參數寫死，只允許傳入9個 我新加的傳不進去（但是目前併爲發現）2，既然在net和decode中可以得到結果，那麼問題可能就出現在作者寫的beamsearch.cpp中：https://github.com/peteanderson80/caffe/blob/631806541c68658248ffdbbbde659f478fac4113/src/caffe/layers/beam_search_layer.cpp。
4，提問：爲什麼在scst中訓練不出來參數，怎樣才能讓模型跑出來，並且decode出來結果
_______