Skip to content

set a smaller beam for pruned lattice rescoring with kaldi-rnnlm #2325

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
danpovey merged 2 commits into from
Apr 13, 2018
Merged

set a smaller beam for pruned lattice rescoring with kaldi-rnnlm #2325

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
89c1e39
set a smaller beam for pruned lattice rescoring with kaldi-rnnlm
hainan-xv Mar 30, 2018
9f9b977
fix an issue in pruned lattice rescoring where small beams might not …
hainan-xv Apr 5, 2018
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 4 additions & 1 deletion scripts/rnnlm/lmrescore_pruned.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ max_ngram_order=4 # Approximate the lattice-rescoring by limiting the max-ngram-
# the same ngram history and this prevents the lattice from
# exploding exponentially. Details of the n-gram approximation
# method are described in section 2.3 of the paper
# http://www.danielpovey.com/files/2018_icassp_lattice_pruning.pdm
# http://www.danielpovey.com/files/2018_icassp_lattice_pruning.pdf
max_arcs= # limit the max arcs in lattice while rescoring. E.g., 20000

acwt=0.1
Expand All @@ -26,6 +26,8 @@ normalize=false # If true, we add a normalization step to the output of the RNNL
# as in our RNNLM setup, a properly trained network would automatically
# have its normalization term close to 1. The details of this
# could be found at http://www.danielpovey.com/files/2018_icassp_rnnlm.pdf
lattice_prune_beam=4 # Beam used in pruned lattice composition
Copy link
Copy Markdown
Contributor

@danpovey danpovey Mar 30, 2018

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@hainan-xv, in your experience, how much difference does this make to WER results? (4 vs. 6)

# This option affects speed and how large the composed lattice may be

# End configuration section.

Expand Down Expand Up @@ -97,6 +99,7 @@ cp $indir/num_jobs $outdir

$cmd JOB=1:$nj $outdir/log/rescorelm.JOB.log \
lattice-lmrescore-kaldi-rnnlm-pruned --lm-scale=$weight $special_symbol_opts \
--lattice-compose-beam=$lattice_prune_beam \
--acoustic-scale=$acwt --max-ngram-order=$max_ngram_order $normalize_opt $max_arcs_opt \
$carpa_option $oldlm $word_embedding "$rnnlm_dir/final.raw" \
"ark:gunzip -c $indir/lat.JOB.gz|" "ark,t:|gzip -c>$outdir/lat.JOB.gz" || exit 1;
Expand Down
15 changes: 13 additions & 2 deletions src/lat/compose-lattice-pruned.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -272,7 +272,14 @@ class PrunedCompactLatticeComposer {
// BaseFloat expected_cost_offset;
};


// This bool variable is initialized to false, and will be updated to true
// the first time a Final() function is called on the det_fst_. Then we will
// immediately call RecomputeRruningInfo() so that the output_best_cost_ is
// changed from +inf to a finite value, to be used in beam search. This is the
// only time the RecomputeRruningInfo() function is called manually; otherwise
// it always follows an automatic schedule based on the num-arcs of the output
// lattice.
bool output_reached_final_;
const ComposeLatticePrunedOptions &opts_;
const CompactLattice &clat_in_;
fst::DeterministicOnDemandFst<fst::StdArc> *det_fst_;
Expand Down Expand Up @@ -584,7 +591,7 @@ PrunedCompactLatticeComposer::PrunedCompactLatticeComposer(
const ComposeLatticePrunedOptions &opts,
const CompactLattice &clat_in,
fst::DeterministicOnDemandFst<fst::StdArc> *det_fst,
CompactLattice* composed_clat):
CompactLattice* composed_clat): output_reached_final_(false),
opts_(opts), clat_in_(clat_in), det_fst_(det_fst),
clat_out_(composed_clat),
num_arcs_out_(0),
Expand Down Expand Up @@ -697,6 +704,10 @@ void PrunedCompactLatticeComposer::ProcessQueueElement(
double final_cost = ConvertToCost(final_lat_weight);
if (final_cost < src_composed_state_info.backward_cost)
src_composed_state_info.backward_cost = final_cost;
if (!output_reached_final_) {
output_reached_final_ = true;
RecomputePruningInfo();
}
}
} else {
// It really was an arc. This code is very complicated, so we make it its
Expand Down