|
'KALDI'
|
The class ParseOptions handles the parsing of command-line options given to main() via argc and argv. First we give an example of how we invoke a typical Kaldi program from the command-line:
gmm-align --transition-scale=10.0 --beam=75 \
exp/mono/tree exp/mono/30.mdl data/L.fst \
'ark:add-deltas --print-args=false scp:data/train.scp ark:- |' \
ark:data/train.tra ark:exp/tri/0.ali
The command-line options, which only have a long form (there are no one-letter options), must appear before the positional arguments. In this case there are six positional arguments, starting from "exp/mono/tree"; notice that the one starting with "ark:add-deltas" is a single string with spaces in it; the single-quotes get interpreted by the shell; this argument gets invoked as a pipe.
We will illustrate how these options get handled at the C++ level by introducing some of the code from gmm-align.cc (we have modified it slightly to make it clearer):
int main(int argc, char *argv[]) { try { // try-catch block is standard and relates to handling of errors. using namespace kaldi; const char *usage = "Align features given [GMM-based] models.\n" "Usage: align-gmm [options] tree-in model-in lexicon-fst-in feature-rspecifier " "transcriptions-rspecifier alignments-wspecifier\n"; // Initialize the ParseOptions object with the usage string. ParseOptions po(usage); // Declare options and set default values. bool binary = false; BaseFloat beam = 200.0; // Below is a structure containing options; its initializer sets defaults. TrainingGraphCompilerOptions gopts; // Register the options with the ParseOptions object. po.Register("binary", &binary, "Write output in binary mode"); po.Register("beam", &beam, "Decoding beam"); gopts.Register(&po); // The command-line options get parsed here. po.Read(argc, argv); // Check that there are a valid number of positional arguments. if(po.NumArgs() != 6) { po.PrintUsage(); exit(1); } // The positional arguments get read here (they can only be obtained // from ParseOptions as strings). std::string tree_in_filename = po.GetArg(1); ... std::string alignment_wspecifier = po.GetArg(6); ... } catch(const std::exception& e) { std::cerr << e.what(); return -1; } }
The code above is mostly self-explanatory. In a normal Kaldi program, the sequence is as follows:
po.NumArgs() is in the valid range for your program.po.GetArg(1) and so on; for optional positional arguments that may be out of range, the convenience function po.GetOptArg(n) returns the n'th argument, or the empty string if n was out of range.Typically when writing a new command-line Kaldi program, it will be easiest to copy an existing one and modify it.
Certain command-line options are automatically registered by the ParseOptions object itself. These include the following:
--config This option loads command-line options from a config file. E.g. if we do --config=configs/my.conf, the file my.conf might contain:
--first-option=15 # This is the first option
--second-option=false # This is the second option
--print-args This boolean option controls whether the program prints the command-line arguments to the standard error (default is true); --print-args=false will turn this off.--help This boolean argument, if true will cause the program to print out a usage message (like other boolean arguments, just --help counts as true). You can usually get the usage message just by omitting all command-line arguments, because most programs require positional arguments.--verbose This controls the verbose level, so that messages logged with KALDI_VLOG will get printed out. More is higher (e.g. --verbose=2 is typical).