Faster-rnnlm代码分析1

https://github.com/yandex/faster-rnnlm

   

Gdb ./rnnlm

r -rnnlm model-good.faster -train thread.title.good.train.txt -valid thread.title.good.valid.txt -hidden 5- -direct-order 3 -direct 200 -bptt 4 -bptt-block 10 -threads 1

[root@cq01-forum-rstree01.cq01.baidu.com faster-rnnlm]# more thread.title.good.train.txt

唉        稳        凉菜        干货        批发        稳        左成        个        月        都

咦        丢        图        跑

毕竟        新人

我        想去旅行

昨天        玩        个        满        深渊        人马                 才        踩        了        55

这        状态        还        不如        温网

新型        投资项目

晒        早饭        就        酱

渣土        哥        真是        太        放肆        了

推荐        就是        有        这样        的

白素贞        水        漫        文水        城

我知道        那些夏天        就像        你        一样        回        不

渑池        至        洛阳        最早        的        车        几        点        哪里        坐        到        洛阳        几点

宏观        方面        大        的        流动性        格局        虽无        明显        变化        但        眼下        地方        政府        债务        限

电工        行业        竞争        大        锦力        电器        有        优势

兄弟        啊                 影技        1        班        q        群        是        多少

你们        家乡        话        叫        什么

深深        的        孤独感        与        挫败        感        感觉        个人

一起去        旅游        吧

谁知道        四会        那里        有        修        打火机        的

[root@cq01-forum-rstree01.cq01.baidu.com faster-rnnlm]# pwd

/home/users/chenghuige/other/faster-rnnlm.debug/faster-rnnlm

1. 统计词频建立vocabulary

void Vocabulary::BuildFromCorpus(const std::string& fpath, bool show_progress)

首先添加一个 </s>

AddWord(kEOSTag); 只是编号0

   

然后逐个添加每行

每行处理的时候按照IsSpace切分

inline bool IsSpace(char c) {

return c == ' ' || c == ' ' || c == ' ' || c == ' ';

   

然后其实就是对每个词 类似 Identifer.h那样顺序编号，没出现的词 叫做oov 编号 -1

   

除了编号之外 同时统计频次

最后按照频次排序 从大到小 同时更新编号 也就是频次最大的 这里 </s> 编号为0

(gdb) p words_

$20 = std::vector of length 176788, capacity 262144 = {{freq = 900000, word = 0x6ae1c0 "</s>"}, {

freq = 258246, word = 0x6aef20 "265304"}, {freq = 126910, word = 0x6aeff0 "301313"}, {

freq = 101904, word = 0x6aedc0 "316322"}, {freq = 67328, word = 0x6aeee0 "323320"}, {

freq = 62290, word = 0x6aec10 "270366"}, {freq = 60866, word = 0x6afb20 "322273"}, {

   

[root@cq01-forum-rstree01.cq01.baidu.com faster-rnnlm]# wc -l thread.title.good.train.txt

900000 thread.title.good.train.txt

   

gdb) p cfg

$2 = {layer_size = 5, layer_count = 1, maxent_hash_size = 199947228, maxent_order = 3, use_nce = false, nce_lnz = 9, reverse_sentence = false, layer_type = "sigmoid"}

   

1. 构建网格结构

main_nnet = new NNet(vocab, cfg, use_cuda, use_cuda_memory_efficient);

构造函数调用Init 在这里

   

embeddings.resize(vocab.size(), cfg.layer_size);

//(word_num, hidden_size) 二维数组

   

rec_layer = CreateLayer(cfg.layer_type, cfg.layer_size, cfg.layer_count);

//隐层 建立一个layer 默认layer_type是sigmoid

   

maxent_layer.Init(cfg.maxent_hash_size);

//最大熵 @TODO

   

softmax_layer = HSTree::CreateHuffmanTree(vocab, cfg.layer_size);

//输出层 softmax 采用huffman树