深度学习笔记（1）：caffe 添加新层 attention LSTM layer和LSTM layer代码精读

总结一下最近的工作：LSTM layer 代码，caffe 加入新层 Attention LSTM layer

LSTM layer

关键代码如下，可以参考图1进行阅读，图一来自博客

namespace caffe {

template <typename Dtype>
void LSTMLayer<Dtype>::RecurrentInputBlobNames(vector<string>* names) const {
  names->resize(2);
  (*names)[0] = "h_0";
  (*names)[1] = "c_0";   //定义h_0,c_0 的输入
}                      

template <typename Dtype>
void LSTMLayer<Dtype>::RecurrentOutputBlobNames(vector<string>* names) const {
  names->resize(2);
  (*names)[0] = "h_" + this->int_to_str(this->T_);
  (*names)[1] = "c_T";  // 定义输出，不同时刻的h_t
}

template <typename Dtype>
void LSTMLayer<Dtype>::OutputBlobNames(vector<string>* names) const {
  names->resize(1);
  (*names)[0] = "h";  // 最终输出h
}

template <typename Dtype>
void LSTMLayer<Dtype>::FillUnrolledNet(NetParameter* net_param) const {
  const int num_output = this->layer_param_.recurrent_param().num_output();
  CHECK_GT(num_output, 0) << "num_output must be positive";
  const FillerParameter& weight_filler =
      this->layer_param_.recurrent_param().weight_filler();
  const FillerParameter& bias_filler =
      this->layer_param_.recurrent_param().bias_filler(); // 权重W和偏差Bias

  // Add generic LayerParameter's (without bottoms/tops) of layer types we'll
  // use to save redundant code.
  LayerParameter hidden_param;
  hidden_par