词格原先的整句打分是一份没校准过的启发式 unigram,质量很差。这里内嵌 libime 在用的 kenlm 查询子集,改由它的三元模型 sc.lm 来打路径分。 - engine/ngram 只编 kenlm 的 query / trie-build 路径。不定义 WITH_THREADS, parallel_read.cc 因此退化成串行实现,整个子集零 Boost 依赖。 KENLM_MAX_ORDER=3 必须与建模阶数一致,改它等于读不懂已发布的模型文件。 - ngram::LanguageModel 是唯一 include kenlm 头文件的地方,对外只给出一个纯 std 的头,_HAS_AUTO_PTR_ETC 和 kenlm 的 include 目录都保持 PRIVATE。 - shared_language_model 按资源路径共享一份 mmap。kenlm 的查询是只读的且把 上下文放在调用方的 State 里,所以并发会话不需要额外加锁。 - 模型缺失时 valid() 为假,词格退回原来的启发式打分:整句候选只是变差,不会 停止工作。测试覆盖了缺模型降级与按路径共享。 - 出货的 ARPA 不带 <s>/</s>,所以解码从 null_state() 起算而非 begin_state()。 kenlm 是上游副本,不重排版、不进 clang-format,其 LGPL-2.1+ 与 BSD 例外清单 已补进 THIRD_PARTY_NOTICES.txt。googlepinyinime 那条 Fallback 整句路径不变。
141 lines
3.7 KiB
C++
141 lines
3.7 KiB
C++
#ifndef UTIL_THREAD_POOL_H
|
|
#define UTIL_THREAD_POOL_H
|
|
|
|
#include "pcqueue.hh"
|
|
|
|
#include <boost/ptr_container/ptr_vector.hpp>
|
|
#include <boost/optional.hpp>
|
|
#include <boost/thread.hpp>
|
|
|
|
#include <iostream>
|
|
#include <cstdlib>
|
|
|
|
namespace util {
|
|
|
|
template <class HandlerT> class Worker : boost::noncopyable {
|
|
public:
|
|
typedef HandlerT Handler;
|
|
typedef typename Handler::Request Request;
|
|
|
|
template <class Construct> Worker(PCQueue<Request> &in, Construct &construct, const Request &poison)
|
|
: in_(in), handler_(construct), poison_(poison), thread_(boost::ref(*this)) {}
|
|
|
|
// Only call from thread.
|
|
void operator()() {
|
|
Request request;
|
|
while (1) {
|
|
in_.Consume(request);
|
|
if (request == poison_) return;
|
|
try {
|
|
(*handler_)(request);
|
|
}
|
|
catch(const std::exception &e) {
|
|
std::cerr << "Handler threw " << e.what() << std::endl;
|
|
abort();
|
|
}
|
|
catch(...) {
|
|
std::cerr << "Handler threw an exception, dropping request" << std::endl;
|
|
abort();
|
|
}
|
|
}
|
|
}
|
|
|
|
void Join() {
|
|
thread_.join();
|
|
}
|
|
|
|
private:
|
|
PCQueue<Request> &in_;
|
|
|
|
boost::optional<Handler> handler_;
|
|
|
|
const Request poison_;
|
|
|
|
boost::thread thread_;
|
|
};
|
|
|
|
template <class HandlerT> class ThreadPool : boost::noncopyable {
|
|
public:
|
|
typedef HandlerT Handler;
|
|
typedef typename Handler::Request Request;
|
|
|
|
template <class Construct> ThreadPool(std::size_t queue_length, std::size_t workers, Construct handler_construct, Request poison) : in_(queue_length), poison_(poison) {
|
|
for (size_t i = 0; i < workers; ++i) {
|
|
workers_.push_back(new Worker<Handler>(in_, handler_construct, poison));
|
|
}
|
|
}
|
|
|
|
~ThreadPool() {
|
|
for (std::size_t i = 0; i < workers_.size(); ++i) {
|
|
Produce(poison_);
|
|
}
|
|
for (typename boost::ptr_vector<Worker<Handler> >::iterator i = workers_.begin(); i != workers_.end(); ++i) {
|
|
i->Join();
|
|
}
|
|
}
|
|
|
|
void Produce(const Request &request) {
|
|
in_.Produce(request);
|
|
}
|
|
|
|
// For adding to the queue.
|
|
PCQueue<Request> &In() { return in_; }
|
|
|
|
private:
|
|
PCQueue<Request> in_;
|
|
|
|
boost::ptr_vector<Worker<Handler> > workers_;
|
|
|
|
Request poison_;
|
|
};
|
|
|
|
template <class Handler> class RecyclingHandler {
|
|
public:
|
|
typedef typename Handler::Request Request;
|
|
|
|
template <class Construct> RecyclingHandler(PCQueue<Request> &recycling, Construct &handler_construct)
|
|
: inner_(handler_construct), recycling_(recycling) {}
|
|
|
|
void operator()(Request &request) {
|
|
inner_(request);
|
|
recycling_.Produce(request);
|
|
}
|
|
|
|
private:
|
|
Handler inner_;
|
|
PCQueue<Request> &recycling_;
|
|
};
|
|
|
|
template <class HandlerT> class RecyclingThreadPool : boost::noncopyable {
|
|
public:
|
|
typedef HandlerT Handler;
|
|
typedef typename Handler::Request Request;
|
|
|
|
// Remember to call PopulateRecycling afterwards in most cases.
|
|
template <class Construct> RecyclingThreadPool(std::size_t queue, std::size_t workers, Construct handler_construct, Request poison)
|
|
: recycling_(queue), pool_(queue, workers, RecyclingHandler<Handler>(recycling_, handler_construct), poison) {}
|
|
|
|
// Initialization: put stuff into the recycling queue. This could also be
|
|
// done by calling Produce without Consume, but it's often easier to
|
|
// initialize with PopulateRecycling then do a Consume/Produce loop.
|
|
void PopulateRecycling(const Request &request) {
|
|
recycling_.Produce(request);
|
|
}
|
|
|
|
Request Consume() {
|
|
return recycling_.Consume();
|
|
}
|
|
|
|
void Produce(const Request &request) {
|
|
pool_.Produce(request);
|
|
}
|
|
|
|
private:
|
|
PCQueue<Request> recycling_;
|
|
ThreadPool<RecyclingHandler<Handler> > pool_;
|
|
};
|
|
|
|
} // namespace util
|
|
|
|
#endif // UTIL_THREAD_POOL_H
|