def main(args, logger):
logger.debug(f'reading configuration file: {args.conf}')
with open(args.conf, 'r') as f:
conf = yaml.load(f)
groups_conf = {group_conf['group_name']: group_conf for group_conf in conf}
logger.debug(f'load the acoustic units name and group')
with open(args.units, 'r') as f:
grouped_unitnames = defaultdict(list)
for line in f:
name, group = line.strip().split()
grouped_unitnames[group].append(name)
if not args.dataset:
logger.debug('no dataset provided assuming zero mean and ' \
'identity covariance matrix')
mean, var = torch.zeros(args.dimension).float(), \
torch.ones(args.dimension).float()
else:
logger.debug(f'using "{args.dataset}" dataset for ' \
'initialization')
with open(args.dataset, 'rb') as f:
dataset = pickle.load(f)
mean, var = dataset.mean, dataset.var
start_pdf_id = 0
pdfs = []
units = {}
for group in grouped_unitnames:
logger.debug(f'creating HMM for group "{group}"')
tot_emitting_states = 0
for name in grouped_unitnames[group]:
logger.debug(f'creating HMM for unit "{name}"')
group_conf = groups_conf[group]
graph, start_pdf_id = create_unit_graph(group_conf['topology'],
start_pdf_id)
units[name] = graph
tot_emitting_states += count_emitting_state(graph)
pdfs.append(create_pdfs(mean, var, tot_emitting_states, group_conf))
emissions = beer.JointModelSet(pdfs)
logger.debug('saving the HMMs on disk...')
with open(args.out, 'wb') as f:
pickle.dump((units, emissions), f)
logger.info(f'created {len(units)} HMMs for a total of {len(emissions)}' \
f' emitting states')
logger.info(f'expected features dimension: {len(mean)}')
def main(args, logger):
logger.debug(f'reading configuration file: {args.conf}')
with open(args.conf, 'r') as f:
conf = yaml.load(f)
if not args.dataset:
logger.debug('no dataset provided assuming zero mean and ' \
'identity covariance matrix')
mean, var = torch.zeros(args.dim).float(), \
torch.ones(args.dim).float()
else:
logger.debug(f'using "{args.dataset}" dataset for ' \
'initialization')
with open(args.dataset, 'rb') as f:
dataset = pickle.load(f)
mean, var = dataset.mean, dataset.var
start_pdf_id = 0
pdfs = []
units = {}
for unit_group in conf:
prefix = unit_group['group_name']
logger.debug(f'creating HMM for group "{prefix}"')
tot_emitting_states = 0
for i in range(1, unit_group['n_units'] + 1):
unit_name = prefix + str(i)
graph, start_pdf_id = create_unit_graph(unit_group['topology'],
start_pdf_id)
units[unit_name] = graph
tot_emitting_states += count_emitting_state(graph)
pdfs.append(create_pdfs(mean, var, tot_emitting_states, unit_group))
emissions = beer.JointModelSet(pdfs)
logger.debug('saving the HMMs on disk...')
with open(args.out, 'wb') as f:
pickle.dump((units, emissions), f)
logger.info(f'created {len(units)} HMMs for a total of {len(emissions)}' \
f' emitting states')
logger.info(f'expected features dimension: {len(mean)}')
def main():
parser = argparse.ArgumentParser()
group = parser.add_mutually_exclusive_group(required=True)
group.add_argument('--stats', help='Feature statistics file for hmm model')
group.add_argument('--dim',
type=int,
help='Dimension of feature, used for vae-hmm model')
parser.add_argument('conf', help='Configuration file')
parser.add_argument('phones', help='list of phones')
parser.add_argument('hmm_graphs', help='hmm graph for each unit')
parser.add_argument('emissions', help='outout emissions')
args = parser.parse_args()
# Load the HMM configuration.
with open(args.conf, 'r') as fid:
conf = yaml.load(fid)
# Load the phones.
phones = []
with open(args.phones, 'r') as fid:
for line in fid:
tokens = line.split()
phones.append(tokens[0])
phones = phones
# Get the data statistics.
if args.stats:
stats = np.load(args.stats)
mean = torch.from_numpy(stats['mean']).float()
var = torch.from_numpy(stats['var']).float()
else:
dim = args.dim
mean = torch.zeros(dim).float()
var = torch.ones(dim).float()
unit_count = 0
pdf_id = 0
units = {}
emissions = []
for group in conf:
# Create the unit graphs.
for i in range(group['n_units']):
unit_id = unit_count
unit_count += 1
nstates = group['n_state_per_unit']
arcs = parse_topology(group['topology'])
unit, pdf_id = create_unit_graph(nstates, arcs, pdf_id)
units[phones[unit_id]] = unit
tot_states = group['n_units'] * (group['n_state_per_unit'] - 2)
modelset = create_emissions(group, mean, var)
modelset = beer.MixtureSet.create(tot_states, modelset)
emissions.append(modelset)
# Merge the pdfs into a single set.
emissions = beer.JointModelSet(emissions)
with open(args.hmm_graphs, 'wb') as fid:
pickle.dump(units, fid)
with open(args.emissions, 'wb') as fid:
pickle.dump(emissions, fid)