diff options
author | Chris Dyer <redpony@gmail.com> | 2014-10-13 00:42:37 -0400 |
---|---|---|
committer | Chris Dyer <redpony@gmail.com> | 2014-10-13 00:42:37 -0400 |
commit | b1ed81ef3216b212295afa76c5d20a56fb647204 (patch) | |
tree | 9633cdc1b8a341dfa58b0b7fec0e2cae44d28835 /klm/lm/neural | |
parent | 1b17f61d359be6e1c3cea29f8c100db3bcdd73a0 (diff) |
new kenlm
Diffstat (limited to 'klm/lm/neural')
-rw-r--r-- | klm/lm/neural/wordvecs.cc | 23 | ||||
-rw-r--r-- | klm/lm/neural/wordvecs.hh | 38 |
2 files changed, 61 insertions, 0 deletions
diff --git a/klm/lm/neural/wordvecs.cc b/klm/lm/neural/wordvecs.cc new file mode 100644 index 00000000..09bb4260 --- /dev/null +++ b/klm/lm/neural/wordvecs.cc @@ -0,0 +1,23 @@ +#include "lm/neural/wordvecs.hh" + +#include "util/file_piece.hh" + +namespace lm { namespace neural { + +WordVecs::WordVecs(util::FilePiece &f) { + const unsigned long lines = f.ReadULong(); + const std::size_t vocab_mem = ngram::ProbingVocabulary::Size(lines, 1.5); + vocab_backing_.reset(util::CallocOrThrow(vocab_mem)); + vocab_.SetupMemory(vocab_backing_.get(), vocab_mem); + const unsigned long width = f.ReadULong(); + vecs_.resize(width, lines); + for (unsigned long i = 0; i < lines; ++i) { + WordIndex column = vocab_.Insert(f.ReadDelimited()); + for (unsigned int row = 0; row < width; ++row) { + vecs_(row,column) = f.ReadFloat(); + } + } + vocab_.FinishedLoading(); +} + +}} // namespaces diff --git a/klm/lm/neural/wordvecs.hh b/klm/lm/neural/wordvecs.hh new file mode 100644 index 00000000..921a2b22 --- /dev/null +++ b/klm/lm/neural/wordvecs.hh @@ -0,0 +1,38 @@ +#ifndef LM_NEURAL_WORDVECS_H +#define LM_NEURAL_WORDVECS_H + +#include "util/scoped.hh" +#include "lm/vocab.hh" + +#include <Eigen/Dense> + +namespace util { class FilePiece; } + +namespace lm { +namespace neural { + +class WordVecs { + public: + // Columns of the matrix are word vectors. The column index is the word. + typedef Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::ColMajor> Storage; + + /* The file should begin with a line stating the number of word vectors and + * the length of the vectors. Then it's followed by lines containing a + * word followed by floating-point values. + */ + explicit WordVecs(util::FilePiece &in); + + const Storage &Vectors() const { return vecs_; } + + WordIndex Index(StringPiece str) const { return vocab_.Index(str); } + + private: + util::scoped_malloc vocab_backing_; + ngram::ProbingVocabulary vocab_; + + Storage vecs_; +}; + +}} // namespaces + +#endif // LM_NEURAL_WORDVECS_H |