def train_gmm_set(model_list,
training_list,
lab_dir,
model_dir,
orig_model_file,
n_train_iterations=10,
min_var_macro=None,
update_transitions=False):
'''
Train GMM files for different classes.
Input:
model_list : file with names of the models to be trained (one per line)
the model names should match the labels in the annotations
training_list : list of training feature files (absolute paths)
orig_dir : directory where the initial model files lie (HTK mmf-formatted file is required)
model_dir : directory where the output models will be stored
n_train_iterations : number of Baum-Welch iterations
min_var_macro : file with the minimum variance macros (vFloors)
'''
if min_var_macro != None and os.path.exists(min_var_macro):
args = [
'-u', 'mvwt', '-S', training_list, '-L', lab_dir, '-H',
orig_model_file, '-H', min_var_macro, '-M', model_dir, '-m',
str(1), model_list
]
else:
args = [
'-u', 'mvwt', '-S', training_list, '-L', lab_dir, '-H',
orig_model_file, '-M', model_dir, '-m',
str(1), model_list
]
herest.herest(args)
model_file = os.path.join(model_dir, 'newMacros')
if not os.path.exists(model_file):
model_file = os.path.join(model_dir, 'hmmdef')
assert (os.path.exists(model_file))
#print "Trained for a single iteration"
if min_var_macro != None and os.path.exists(min_var_macro):
args = [
'-u', 'mvwt', '-S', training_list, '-L', lab_dir, '-H', model_file,
'-H', min_var_macro, '-M', model_dir, '-m',
str(1), model_list
]
else:
args = [
'-u', 'mvwt', '-S', training_list, '-L', lab_dir, '-H', model_file,
'-M', model_dir, '-m',
str(1), model_list
]
# Ideally a convergence criterion should be applied instead
for iter in range(1, n_train_iterations):
#print "Iteration {0}/{1}".format(iter, n_train_iterations)
#print args
herest.herest(args)
def train_hlda(model_list, n_nuisance_dims, training_list, lab_dir, model_dir, min_var_macro=None, n_states=1):
'''
Estimate HLDA transformation by following the guidelines in the HTK
tutorial (HTKBook, p.49 'Semi-Tied and HLDA transforms'). Found that
HLDA estimation in HTK v3.4-1 only works
'''
# Find the number of GMM components
model_file = os.path.join(model_dir,'newMacros')
assert(os.path.exists(model_file))
m_file = open(model_file,'r')
n_comps = 0
for ln in m_file:
m = re.match('<NUMMIXES>\s+(\d+)',ln)
if m:
n_comps = m.group(1)
break
if n_comps==0:
raise Exception
# Write global class file
class_filename = os.path.join(model_dir,'global')
class_file = open(class_filename,'w')
class_file.write(get_global_class_string(n_comps, n_states))
class_file.close()
# Write configuration file
config_filename = os.path.join(model_dir,'config.hlda')
hlda_conf_map = config_hlda(n_nuisance_dims, max_iterations)
write_htk_config(hlda_conf_map, config_filename)
# Call HERest to estimate the HLDA transform
args = ['-H',os.path.join(model_dir,'newMacros'),'-u','s',
'-C',config_filename,'-L',lab_dir,'-K',model_dir,
'-J',model_dir,'-S',training_list,'-M',model_dir,model_list]
if min_var_macro != None and os.path.exists(min_var_macro):
args.insert(0, '-H')
args.insert(1, min_var_macro)
herest(args)
# Reiterate after the HLDA estimation
min_var_macro = os.path.join(model_dir,'vFloors')
if min_var_macro != None and os.path.exists(min_var_macro):
args = ['-H', os.path.join(model_dir,'newMacros'),'-C', config_filename,
'-L', lab_dir, '-S', training_list, '-H', min_var_macro,
'-M', model_dir, model_list]
else:
args = ['-H', os.path.join(model_dir,'newMacros'),'-C', config_filename,
'-L', lab_dir, '-S', training_list, '-M', model_dir, model_list]
for _ in range(1, 3):
herest(args)
def train_gmm_set(
model_list,
training_list,
lab_dir,
model_dir,
orig_model_file,
n_train_iterations=10,
min_var_macro=None,
update_transitions=False,
):
"""
Train GMM files for different classes.
Input:
model_list : file with names of the models to be trained (one per line)
the model names should match the labels in the annotations
training_list : list of training feature files (absolute paths)
orig_dir : directory where the initial model files lie (HTK mmf-formatted file is required)
model_dir : directory where the output models will be stored
n_train_iterations : number of Baum-Welch iterations
min_var_macro : file with the minimum variance macros (vFloors)
"""
if min_var_macro != None and os.path.exists(min_var_macro):
args = [
"-u",
"mvwt",
"-S",
training_list,
"-L",
lab_dir,
"-H",
orig_model_file,
"-H",
min_var_macro,
"-M",
model_dir,
"-m",
str(1),
model_list,
]
else:
args = [
"-u",
"mvwt",
"-S",
training_list,
"-L",
lab_dir,
"-H",
orig_model_file,
"-M",
model_dir,
"-m",
str(1),
model_list,
]
herest.herest(args)
model_file = os.path.join(model_dir, "newMacros")
if not os.path.exists(model_file):
model_file = os.path.join(model_dir, "hmmdef")
assert os.path.exists(model_file)
# print "Trained for a single iteration"
if min_var_macro != None and os.path.exists(min_var_macro):
args = [
"-u",
"mvwt",
"-S",
training_list,
"-L",
lab_dir,
"-H",
model_file,
"-H",
min_var_macro,
"-M",
model_dir,
"-m",
str(1),
model_list,
]
else:
args = [
"-u",
"mvwt",
"-S",
training_list,
"-L",
lab_dir,
"-H",
model_file,
"-M",
model_dir,
"-m",
str(1),
model_list,
]
# Ideally a convergence criterion should be applied instead
for iter in range(1, n_train_iterations):
# print "Iteration {0}/{1}".format(iter, n_train_iterations)
# print args
herest.herest(args)