def train_one_utterance(labeled_data_event):
print("Beginning training on utterance %d (%s) (%d frames)" %
(utterance_num[0], labeled_data_event.name,
labeled_data_event.data.shape[0]))
sys.stdout.flush()
label_lattice = labels_to_lattice(labeled_data_event.labels)
# Build model lattice
model_lattice = build_model_lattice(label_lattice, model_dict,
epsilon_index)
# Build training graph
training_graph = TrainingGraph(model_lattice, hmm_mgr, dict())
# Train on data
if gmm_mgr.get_adaptation_state() != "INITIALIZING":
gmm_mgr.set_adaptation_state("INITIALIZING")
hmm_mgr.set_adaptation_state("INITIALIZING")
training_graph.begin_training()
gmm_mgr.set_adaptation_state("ACCUMULATING")
hmm_mgr.set_adaptation_state("ACCUMULATING")
training_graph.train_one_sequence(labeled_data_event.data)
training_graph.end_training()
print("Finished training on utterance %d (%s)" %
(utterance_num[0], labeled_data_event.name))
sys.stdout.flush()
utterance_num[0] += 1
def test4(num_passes, num_obs):
# Each of the 4 nodes contains a 4 (or 6)-node order-3 Hmm; the nodes are connected in a
# diamond pattern
ret = ""
dimension = 2
# Data generator setup and data generation
obs_gen = make_data_generator(dimension)
obs_list = [obs_gen.next() for i in xrange(num_obs)]
# GmmMgr setup
num_models = 10
models = make_standard_gmms(dimension, num_models)
gmm_mgr = GmmMgr(models)
# Hmm setup
# Make three Hmms with 4 (or 6) states and order 3 (self loop, forward 1, forward 2)
num_states = 4
seed(0)
hmm0 = make_forward_hmm(gmm_mgr, num_states, 3, exact=True)
hmm1 = make_forward_hmm(gmm_mgr, num_states + 2, 3, exact=True)
hmm2 = make_forward_hmm(gmm_mgr, num_states, 3, exact=True)
hmm_mgr = HmmMgr((hmm0, hmm1, hmm2))
# TrainingGraph setup
gb = GraphBuilder()
# Note that here we are using the same HMM in two different TG nodes
node_id0 = gb.new_node((0, 0))
node_id1 = gb.new_node((1, 1))
node_id2 = gb.new_node((2, 2))
node_id3 = gb.new_node((3, 0))
arc_id = gb.new_arc(node_id0, node_id1)
arc_id = gb.new_arc(node_id0, node_id2)
arc_id = gb.new_arc(node_id1, node_id3)
arc_id = gb.new_arc(node_id2, node_id3)
gr0 = FrozenGraph(gb)
spd = {}
spd[(0, 1)] = (0.4, 0.3, 0.8)
spd[(0, 2)] = (0.6, 0.7, 0.2)
tg0 = TrainingGraph(gr0, hmm_mgr, spd)
# Now adapt original TrainingGraph
for i in xrange(num_passes):
gmm_mgr.set_adaptation_state("INITIALIZING")
gmm_mgr.clear_all_accumulators()
tg0.begin_training()
gmm_mgr.set_adaptation_state("ACCUMULATING")
for obs in obs_list:
tg0.train_one_sequence(obs)
tg0.end_training()
gmm_mgr.set_adaptation_state("APPLYING")
gmm_mgr.apply_all_accumulators()
gmm_mgr.set_adaptation_state("NOT_ADAPTING")
ret = tg0.to_string(full=True)
return ret
def test4(num_passes, num_obs):
# Each of the 4 nodes contains a 4 (or 6)-node order-3 Hmm; the nodes are connected in a
# diamond pattern
ret = ""
dimension = 2
# Data generator setup and data generation
obs_gen = make_data_generator(dimension)
obs_list = [obs_gen.next() for i in xrange(num_obs)]
# GmmMgr setup
num_models = 10
models = make_standard_gmms(dimension, num_models)
gmm_mgr = GmmMgr(models)
# Hmm setup
# Make three Hmms with 4 (or 6) states and order 3 (self loop, forward 1, forward 2)
num_states = 4
seed(0)
hmm0 = make_forward_hmm(gmm_mgr, num_states, 3, exact=True)
hmm1 = make_forward_hmm(gmm_mgr, num_states + 2, 3, exact=True)
hmm2 = make_forward_hmm(gmm_mgr, num_states, 3, exact=True)
hmm_mgr = HmmMgr((hmm0, hmm1, hmm2))
# TrainingGraph setup
gb = GraphBuilder()
# Note that here we are using the same HMM in two different TG nodes
node_id0 = gb.new_node((0, 0))
node_id1 = gb.new_node((1, 1))
node_id2 = gb.new_node((2, 2))
node_id3 = gb.new_node((3, 0))
arc_id = gb.new_arc(node_id0, node_id1)
arc_id = gb.new_arc(node_id0, node_id2)
arc_id = gb.new_arc(node_id1, node_id3)
arc_id = gb.new_arc(node_id2, node_id3)
gr0 = FrozenGraph(gb)
spd = {}
spd[(0, 1)] = (0.4, 0.3, 0.8)
spd[(0, 2)] = (0.6, 0.7, 0.2)
tg0 = TrainingGraph(gr0, hmm_mgr, spd)
# Now adapt original TrainingGraph
for i in xrange(num_passes):
gmm_mgr.set_adaptation_state("INITIALIZING")
gmm_mgr.clear_all_accumulators()
tg0.begin_training()
gmm_mgr.set_adaptation_state("ACCUMULATING")
for obs in obs_list:
tg0.train_one_sequence(obs)
tg0.end_training()
gmm_mgr.set_adaptation_state("APPLYING")
gmm_mgr.apply_all_accumulators()
gmm_mgr.set_adaptation_state("NOT_ADAPTING")
ret = tg0.to_string(full=True)
return ret
def train_one_utterance(labeled_data_event):
print("Beginning training on utterance %d (%s) (%d frames)" % (utterance_num[0], labeled_data_event.name,
labeled_data_event.data.shape[0]))
sys.stdout.flush()
label_lattice = labels_to_lattice(labeled_data_event.labels)
# Build model lattice
model_lattice = build_model_lattice(label_lattice, model_dict, epsilon_index)
# Build training graph
training_graph = TrainingGraph(model_lattice, hmm_mgr, dict())
# Train on data
if gmm_mgr.get_adaptation_state() != "INITIALIZING":
gmm_mgr.set_adaptation_state("INITIALIZING")
hmm_mgr.set_adaptation_state("INITIALIZING")
training_graph.begin_training()
gmm_mgr.set_adaptation_state("ACCUMULATING")
hmm_mgr.set_adaptation_state("ACCUMULATING")
training_graph.train_one_sequence(labeled_data_event.data)
training_graph.end_training()
print("Finished training on utterance %d (%s)" % (utterance_num[0], labeled_data_event.name))
sys.stdout.flush()
utterance_num[0] += 1
