diff options
Diffstat (limited to 'klm/lm/build_binary.cc')
| -rw-r--r-- | klm/lm/build_binary.cc | 54 |
1 files changed, 35 insertions, 19 deletions
diff --git a/klm/lm/build_binary.cc b/klm/lm/build_binary.cc index 4552c419..b7aee4de 100644 --- a/klm/lm/build_binary.cc +++ b/klm/lm/build_binary.cc @@ -15,12 +15,12 @@ namespace ngram { namespace { void Usage(const char *name) { - std::cerr << "Usage: " << name << " [-u log10_unknown_probability] [-s] [-n] [-p probing_multiplier] [-t trie_temporary] [-m trie_building_megabytes] [-q bits] [-b bits] [type] input.arpa output.mmap\n\n" -"-u sets the default log10 probability for <unk> if the ARPA file does not have\n" -"one.\n" + std::cerr << "Usage: " << name << " [-u log10_unknown_probability] [-s] [-i] [-p probing_multiplier] [-t trie_temporary] [-m trie_building_megabytes] [-q bits] [-b bits] [-c bits] [type] input.arpa [output.mmap]\n\n" +"-u sets the log10 probability for <unk> if the ARPA file does not have one.\n" +" Default is -100. The ARPA file will always take precedence.\n" "-s allows models to be built even if they do not have <s> and </s>.\n" -"-i allows buggy models from IRSTLM by mapping positive log probability to 0.\n" -"type is either probing or trie:\n\n" +"-i allows buggy models from IRSTLM by mapping positive log probability to 0.\n\n" +"type is either probing or trie. Default is probing.\n\n" "probing uses a probing hash table. It is the fastest but uses the most memory.\n" "-p sets the space multiplier and must be >1.0. The default is 1.5.\n\n" "trie is a straightforward trie with bit-level packing. It uses the least\n" @@ -29,10 +29,11 @@ void Usage(const char *name) { "-t is the temporary directory prefix. Default is the output file name.\n" "-m limits memory use for sorting. Measured in MB. Default is 1024MB.\n" "-q turns quantization on and sets the number of bits (e.g. -q 8).\n" -"-b sets backoff quantization bits. Requires -q and defaults to that value.\n\n" -"See http://kheafield.com/code/kenlm/benchmark/ for data structure benchmarks.\n" -"Passing only an input file will print memory usage of each data structure.\n" -"If the ARPA file does not have <unk>, -u sets <unk>'s probability; default 0.0.\n"; +"-b sets backoff quantization bits. Requires -q and defaults to that value.\n" +"-a compresses pointers using an array of offsets. The parameter is the\n" +" maximum number of bits encoded by the array. Memory is minimized subject\n" +" to the maximum, so pick 255 to minimize memory.\n\n" +"Get a memory estimate by passing an ARPA file without an output file name.\n"; exit(1); } @@ -63,12 +64,14 @@ void ShowSizes(const char *file, const lm::ngram::Config &config) { std::vector<uint64_t> counts; util::FilePiece f(file); lm::ReadARPACounts(f, counts); - std::size_t sizes[3]; + std::size_t sizes[5]; sizes[0] = ProbingModel::Size(counts, config); sizes[1] = TrieModel::Size(counts, config); sizes[2] = QuantTrieModel::Size(counts, config); - std::size_t max_length = *std::max_element(sizes, sizes + 3); - std::size_t min_length = *std::max_element(sizes, sizes + 3); + sizes[3] = ArrayTrieModel::Size(counts, config); + sizes[4] = QuantArrayTrieModel::Size(counts, config); + std::size_t max_length = *std::max_element(sizes, sizes + sizeof(sizes) / sizeof(size_t)); + std::size_t min_length = *std::min_element(sizes, sizes + sizeof(sizes) / sizeof(size_t)); std::size_t divide; char prefix; if (min_length < (1 << 10) * 10) { @@ -91,7 +94,9 @@ void ShowSizes(const char *file, const lm::ngram::Config &config) { std::cout << prefix << "B\n" "probing " << std::setw(length) << (sizes[0] / divide) << " assuming -p " << config.probing_multiplier << "\n" "trie " << std::setw(length) << (sizes[1] / divide) << " without quantization\n" - "trie " << std::setw(length) << (sizes[2] / divide) << " assuming -q " << (unsigned)config.prob_bits << " -b " << (unsigned)config.backoff_bits << " quantization \n"; + "trie " << std::setw(length) << (sizes[2] / divide) << " assuming -q " << (unsigned)config.prob_bits << " -b " << (unsigned)config.backoff_bits << " quantization \n" + "trie " << std::setw(length) << (sizes[3] / divide) << " assuming -a " << (unsigned)config.pointer_bhiksha_bits << " array pointer compression\n" + "trie " << std::setw(length) << (sizes[4] / divide) << " assuming -a " << (unsigned)config.pointer_bhiksha_bits << " -q " << (unsigned)config.prob_bits << " -b " << (unsigned)config.backoff_bits<< " array pointer compression and quantization\n"; } void ProbingQuantizationUnsupported() { @@ -106,11 +111,11 @@ void ProbingQuantizationUnsupported() { int main(int argc, char *argv[]) { using namespace lm::ngram; - bool quantize = false, set_backoff_bits = false; try { + bool quantize = false, set_backoff_bits = false, bhiksha = false; lm::ngram::Config config; int opt; - while ((opt = getopt(argc, argv, "siu:p:t:m:q:b:")) != -1) { + while ((opt = getopt(argc, argv, "siu:p:t:m:q:b:a:")) != -1) { switch(opt) { case 'q': config.prob_bits = ParseBitCount(optarg); @@ -121,6 +126,9 @@ int main(int argc, char *argv[]) { config.backoff_bits = ParseBitCount(optarg); set_backoff_bits = true; break; + case 'a': + config.pointer_bhiksha_bits = ParseBitCount(optarg); + bhiksha = true; case 'u': config.unknown_missing_logprob = ParseFloat(optarg); break; @@ -162,9 +170,17 @@ int main(int argc, char *argv[]) { ProbingModel(from_file, config); } else if (!strcmp(model_type, "trie")) { if (quantize) { - QuantTrieModel(from_file, config); + if (bhiksha) { + QuantArrayTrieModel(from_file, config); + } else { + QuantTrieModel(from_file, config); + } } else { - TrieModel(from_file, config); + if (bhiksha) { + ArrayTrieModel(from_file, config); + } else { + TrieModel(from_file, config); + } } } else { Usage(argv[0]); @@ -173,9 +189,9 @@ int main(int argc, char *argv[]) { Usage(argv[0]); } } - catch (std::exception &e) { + catch (const std::exception &e) { std::cerr << e.what() << std::endl; - abort(); + return 1; } return 0; } |