hang: nnetbin/sat-nnet-train-frmshuff.cc注解1

Nnet读取

Nnet nnet;
nnet.Read(model_filename);
nnet.SetTrainOptions(trn_opts);

1. 声明nnet

• std::vector<Component *> components_；用于存储每一个component
• std::vector<CuMatrix<BaseFloat>> propagate_buf_,backpropagate_buf_;
  两个用于前向和后向的buffer（显存）
• NnetTrainOptions opts_; 用于存储learningrate，momentum，L1，L2

2. 利用model_filename读取模型和SetTrainOptions设置learningrate等参数

• 读模型的时候需要前一个component的输出维度==后一个component的输入维度
• propagate_buf_.resize(NumComponents()+1);backpropagate_buf_.resize(NumComponents()+1);
• opts_.learn_rate=0.0

构建映射表

std::map<std::string, int32> map_utt2spk;
if (utt2spk != "") {
std::ifstream fi_map(utt2spk.c_str());
  if (!fi_map) {
    KALDI_ERR << "failed to open file " << utt2spk;
  }
  std::string key;
  int32 value;
  while (fi_map >> key >> value) {
    map_utt2spk[key] = value;
  }
}

1. 这里的key为utterance-id (注意区分utterance-id和recording-id)
2. 这里的value为0-speakernum-1的数值，对应speaker_code矩阵的行号