def compute_presoftmax_prior_scale(dir, alidir, num_jobs, run_opts,
presoftmax_prior_scale_power=-0.25):
# getting the raw pdf count
common_lib.execute_command(
"""{command} JOB=1:{num_jobs} {dir}/log/acc_pdf.JOB.log \
ali-to-post "ark:gunzip -c {alidir}/ali.JOB.gz|" ark:- \| \
post-to-tacc --per-pdf=true {alidir}/final.mdl ark:- \
{dir}/pdf_counts.JOB""".format(command=run_opts.command,
num_jobs=num_jobs,
dir=dir,
alidir=alidir))
common_lib.execute_command(
"""{command} {dir}/log/sum_pdf_counts.log \
vector-sum --binary=false {dir}/pdf_counts.* {dir}/pdf_counts \
""".format(command=run_opts.command, dir=dir))
for file in glob.glob('{0}/pdf_counts.*'.format(dir)):
os.remove(file)
pdf_counts = common_lib.read_kaldi_matrix('{0}/pdf_counts'.format(dir))[0]
scaled_counts = smooth_presoftmax_prior_scale_vector(
pdf_counts,
presoftmax_prior_scale_power=presoftmax_prior_scale_power,
smooth=0.01)
output_file = "{0}/presoftmax_prior_scale.vec".format(dir)
common_lib.write_kaldi_matrix(output_file, [scaled_counts])
common_lib.force_symlink("../presoftmax_prior_scale.vec",
"{0}/configs/presoftmax_prior_scale.vec".format(
dir))
def get_full_config(self):
ans = []
# note: each value of self.descriptors is (descriptor, dim,
# normalized-string, output-string).
# by 'descriptor_final_string' we mean a string that can appear in
# config-files, i.e. it contains the 'final' names of nodes.
descriptor_final_string = self.descriptors['input']['final-string']
input_dim = self.descriptors['input']['dim']
output_dim = self.output_dim()
transform_file = self.config['affine-transform-file']
idct_mat = common_lib.compute_idct_matrix(
input_dim, output_dim, self.config['cepstral-lifter'])
# append a zero column to the matrix, this is the bias of the fixed
# affine component
for n in range(0, output_dim):
idct_mat[n].append(0)
common_lib.write_kaldi_matrix(transform_file, idct_mat)
# write the 'real' component to final.config
line = 'component name={0} type=FixedAffineComponent matrix={1}'.format(
self.name, transform_file)
ans.append(('final', line))
# write a random version of the component, with the same dims, to ref.config
line = 'component name={0} type=FixedAffineComponent input-dim={1} output-dim={2}'.format(
self.name, input_dim, output_dim)
ans.append(('ref', line))
# the component-node gets written to final.config and ref.config.
line = 'component-node name={0} component={0} input={1}'.format(
self.name, descriptor_final_string)
ans.append(('final', line))
ans.append(('ref', line))
return ans
def compute_presoftmax_prior_scale(dir, alidir, num_jobs, run_opts,
presoftmax_prior_scale_power=-0.25):
# getting the raw pdf count
common_lib.run_job(
"""{command} JOB=1:{num_jobs} {dir}/log/acc_pdf.JOB.log \
ali-to-post "ark:gunzip -c {alidir}/ali.JOB.gz|" ark:- \| \
post-to-tacc --per-pdf=true {alidir}/final.mdl ark:- \
{dir}/pdf_counts.JOB""".format(command=run_opts.command,
num_jobs=num_jobs,
dir=dir,
alidir=alidir))
common_lib.run_job(
"""{command} {dir}/log/sum_pdf_counts.log \
vector-sum --binary=false {dir}/pdf_counts.* {dir}/pdf_counts \
""".format(command=run_opts.command, dir=dir))
for file in glob.glob('{0}/pdf_counts.*'.format(dir)):
os.remove(file)
pdf_counts = common_lib.read_kaldi_matrix('{0}/pdf_counts'.format(dir))[0]
scaled_counts = smooth_presoftmax_prior_scale_vector(
pdf_counts,
presoftmax_prior_scale_power=presoftmax_prior_scale_power,
smooth=0.01)
output_file = "{0}/presoftmax_prior_scale.vec".format(dir)
common_lib.write_kaldi_matrix(output_file, [scaled_counts])
common_lib.force_symlink("../presoftmax_prior_scale.vec",
"{0}/configs/presoftmax_prior_scale.vec".format(
dir))
def get_full_config(self):
ans = []
# note: each value of self.descriptors is (descriptor, dim,
# normalized-string, output-string).
# by 'descriptor_final_string' we mean a string that can appear in
# config-files, i.e. it contains the 'final' names of nodes.
descriptor_final_string = self.descriptors['input']['final-string']
input_dim = self.descriptors['input']['dim']
output_dim = self.output_dim()
transform_file = self.config['affine-transform-file']
idct_mat = common_lib.compute_idct_matrix(
input_dim, output_dim, self.config['cepstral-lifter'])
# append a zero column to the matrix, this is the bias of the fixed
# affine component
for n in range(0, output_dim):
idct_mat[n].append(0)
common_lib.write_kaldi_matrix(transform_file, idct_mat)
# write the 'real' component to final.config
line = 'component name={0} type=FixedAffineComponent matrix={1}'.format(
self.name, transform_file)
ans.append(('final', line))
# write a random version of the component, with the same dims, to ref.config
line = 'component name={0} type=FixedAffineComponent input-dim={1} output-dim={2}'.format(
self.name, input_dim, output_dim)
ans.append(('ref', line))
# the component-node gets written to final.config and ref.config.
line = 'component-node name={0} component={0} input={1}'.format(
self.name, descriptor_final_string)
ans.append(('final', line))
ans.append(('ref', line))
return ans
def compute_presoftmax_prior_scale_targets(dir, counts_path,
presoftmax_prior_scale_power=-0.25):
# total num of frames per target already prepared
target_counts = common_lib.read_kaldi_matrix(counts_path)[0]
scaled_counts = smooth_presoftmax_prior_scale_vector(
target_counts,
presoftmax_prior_scale_power=presoftmax_prior_scale_power,
smooth=0.01)
output_file = "{0}/presoftmax_prior_scale.vec".format(dir)
common_lib.write_kaldi_matrix(output_file, [scaled_counts])
common_lib.force_symlink("../presoftmax_prior_scale.vec",
"{0}/configs/presoftmax_prior_scale.vec".format(
dir))
