def increase_n_components(n_mixes,
model_name,
model_file,
training_list,
lab_dir,
target_model_dir,
min_var_macro_file=None,
n_states=1,
do_kmeans=False):
'''
Increase the number of GMM components
'''
hhed_cfg_file = os.path.join(target_model_dir, 'hhed.cfg')
hhed_cfg = open(hhed_cfg_file, 'w')
hhed_cfg.write('KEEPDISTINCT=TRUE')
hhed_cfg.close()
hhed_cmd_file = os.path.join(target_model_dir, 'hhed.cmd')
hhed_cmd = open(hhed_cmd_file, 'w')
if n_states > 1:
hhed_cmd.write('MU {0} {{{1}.state[2-{2}].mix}}\n'.format(
str(n_mixes), model_name, str(n_states + 1)))
else:
hhed_cmd.write('MU {0} {{{1}.state[2].mix}}\n'.format(
str(n_mixes), model_name))
hhed_cmd.close()
model_list_file = os.path.join(target_model_dir, 'model.list')
model_list = open(model_list_file, 'w')
model_list.write(model_name)
model_list.close()
model_dir = os.path.split(model_file)[0]
hhed_args = [
'-C', hhed_cfg_file, '-d', model_dir, '-M', target_model_dir,
hhed_cmd_file, model_list_file
]
hhed(hhed_args)
if do_kmeans == True:
initialize_gmm.initialize_gmm_kmeans(model_name, training_list,
lab_dir, target_model_dir,
target_model_dir)
train_gmm(model_file=os.path.join(target_model_dir, model_name),
training_list=training_list,
lab_dir=lab_dir,
out_model_dir=target_model_dir,
n_train_iterations=30,
min_var_macro=min_var_macro_file)
def increase_n_components(
n_mixes,
model_name,
model_file,
training_list,
lab_dir,
target_model_dir,
min_var_macro_file=None,
n_states=1,
do_kmeans=False,
):
"""
Increase the number of GMM components
"""
hhed_cfg_file = os.path.join(target_model_dir, "hhed.cfg")
hhed_cfg = open(hhed_cfg_file, "w")
hhed_cfg.write("KEEPDISTINCT=TRUE")
hhed_cfg.close()
hhed_cmd_file = os.path.join(target_model_dir, "hhed.cmd")
hhed_cmd = open(hhed_cmd_file, "w")
if n_states > 1:
hhed_cmd.write("MU {0} {{{1}.state[2-{2}].mix}}\n".format(str(n_mixes), model_name, str(n_states + 1)))
else:
hhed_cmd.write("MU {0} {{{1}.state[2].mix}}\n".format(str(n_mixes), model_name))
hhed_cmd.close()
model_list_file = os.path.join(target_model_dir, "model.list")
model_list = open(model_list_file, "w")
model_list.write(model_name)
model_list.close()
model_dir = os.path.split(model_file)[0]
hhed_args = ["-C", hhed_cfg_file, "-d", model_dir, "-M", target_model_dir, hhed_cmd_file, model_list_file]
hhed(hhed_args)
if do_kmeans == True:
initialize_gmm.initialize_gmm_kmeans(model_name, training_list, lab_dir, target_model_dir, target_model_dir)
train_gmm(
model_file=os.path.join(target_model_dir, model_name),
training_list=training_list,
lab_dir=lab_dir,
out_model_dir=target_model_dir,
n_train_iterations=30,
min_var_macro=min_var_macro_file,
)
def train_M_sized_gmm_classifier_incrementally(M, n_dims, feature_type, training_list, model_list,
lab_dir, model_dir, apply_hlda=False, hlda_nuisance_dims=0,
n_train_iterations=10):
'''
Train a GMM-based classifier with a known number of components
The assumption is that training files are HTK-formatted feature files and that
there exist HTK-formatted transcription files as well named in the same way
as the feature files with just a .lab suffix. The lab_dir should correspond to
a flat structure.
'''
# Find the dimensionality of the feature vector
assert(os.path.exists(training_list))
t_list = open(training_list,'r')
fea_file = t_list.readline().rstrip('\r\n')
t_list.close()
sampSize = read_htk_header(fea_file)[2]
orig_n_dims = sampSize/4
assert(orig_n_dims<max_dimensions)
# Find the model names
models = []
m_list = open(model_list,'r')
for line in m_list:
line = line.rstrip('\r\n')
models.append(line)
# Directory structure generation
if not os.path.exists(model_dir):
os.makedirs(model_dir)
hmm0_dir = os.path.join(model_dir,'hmm0')
if not os.path.exists(hmm0_dir):
os.makedirs(hmm0_dir)
hmm1_dir = os.path.join(model_dir,'hmm1')
if not os.path.exists(hmm1_dir):
os.makedirs(hmm1_dir)
n_models = len(models)
final_model_files = []
for count in range(n_models):
if len(M) > 1:
n_comps = M[count]
else:
n_comps = M[0]
initial_n_comps = 1
if n_comps>1:
initial_n_comps = 2
model_name = models[count]
# Define the GMM as a trivial HMM with a single state in HTK format
create_gmm(orig_n_dims, initial_n_comps, feature_type, model_name, covar_type, hmm0_dir)
initialize_gmm.initialize_gmm_kmeans(model_name, training_list, lab_dir, hmm0_dir, hmm1_dir)
if count == 0:
# Estimate variance floors as percentage of the global variances (only once)
initialize_gmm.estimate_minimum_variances(training_list, os.path.join(hmm0_dir, models[count]), hmm0_dir)
min_var_macro_file = os.path.join(hmm0_dir,'vFloors')
train_gmm.prepare_vfloor_macros_file(model_file=os.path.join(hmm1_dir, model_name),
vfloors_file=min_var_macro_file)
# Increase the number of mixtures
n_mixes = 4;
orig_model_dir = hmm1_dir
target_model_dir = os.path.join(model_dir,'hmm4');
while n_mixes < n_comps+1:
if not os.path.exists(target_model_dir):
os.makedirs(target_model_dir)
model_file = os.path.join(orig_model_dir, model_name)
train_gmm.increase_n_components(n_mixes, model_name, model_file, training_list, lab_dir,
target_model_dir, min_var_macro_file)
n_mixes *= 2
orig_model_dir = target_model_dir
target_model_dir = os.path.join(model_dir,'hmm{0}'.format(str(n_mixes)))
final_model_files.append(os.path.join(orig_model_dir,model_name))
# Gather all HMM definitions into a single mmf file
model_file = os.path.join(model_dir, "newMacros")
mmf = open(model_file,'w')
for md in final_model_files:
md_file = open(md,'r')
mmf.writelines(md_file.readlines())
mmf.write('\n')
md_file.close()
mmf.close()
train_gmm.train_gmm_set(model_list, training_list, lab_dir, model_dir, model_file, n_train_iterations, min_var_macro_file)
model_file = os.path.join(model_dir, "newMacros")
shutil.copyfile(model_file, model_file+'_no_hlda')
if apply_hlda:
if hlda_nuisance_dims>0:
train_hlda(model_list, hlda_nuisance_dims, training_list, lab_dir, model_dir, min_var_macro_file)
