def test2(num_obs):
# Each of the 2 nodes contains a 4-node order-2 Hmm; the nodes are connected in single chain
dimension = 2
obs_gen = make_data_generator(dimension)
obs_list = [obs_gen.next() for i in xrange(num_obs)]
# GmmMgr setup
num_models = 20
models = make_standard_gmms(dimension, num_models)
gmm_mgr1 = GmmMgr(models[0:10])
gmm_mgr2 = GmmMgr(models[10:20])
# Hmm setup
# Make two Hmms with 4 states and order 2 (self loop, forward 1)
num_states = 4
seed(0)
hmm0 = make_forward_hmm(gmm_mgr1, num_states, 2, exact=True)
hmm1 = make_forward_hmm(gmm_mgr1, num_states, 2, exact=True)
hmm_mgr = HmmMgr((hmm0, hmm1))
# TrainingGraph setup
gb = GraphBuilder()
node_id0 = gb.new_node((0, 0))
node_id1 = gb.new_node((1, 1))
arc_id = gb.new_arc(node_id0, node_id1)
gr0 = FrozenGraph(gb)
tg0 = TrainingGraph(gr0, hmm_mgr, dict())
valid, ret = validate_training_graph(tg0, gmm_mgr1, hmm_mgr, obs_list, 1,
gmm_mgr2)
return ret
def test2(num_obs):
# Each of the 2 nodes contains a 4-node order-2 Hmm; the nodes are connected in single chain
dimension = 2
obs_gen = make_data_generator(dimension)
obs_list = [obs_gen.next() for i in xrange(num_obs)]
# GmmMgr setup
num_models = 20
models = make_standard_gmms(dimension, num_models)
gmm_mgr1 = GmmMgr(models[0:10])
gmm_mgr2 = GmmMgr(models[10:20])
# Hmm setup
# Make two Hmms with 4 states and order 2 (self loop, forward 1)
num_states = 4
seed(0)
hmm0 = make_forward_hmm(gmm_mgr1, num_states, 2, exact=True)
hmm1 = make_forward_hmm(gmm_mgr1, num_states, 2, exact=True)
hmm_mgr = HmmMgr((hmm0, hmm1))
# TrainingGraph setup
gb = GraphBuilder()
node_id0 = gb.new_node((0, 0))
node_id1 = gb.new_node((1, 1))
arc_id = gb.new_arc(node_id0, node_id1)
gr0 = FrozenGraph(gb)
tg0 = TrainingGraph(gr0, hmm_mgr, dict())
valid, ret = validate_training_graph(tg0, gmm_mgr1, hmm_mgr, obs_list, 1, gmm_mgr2)
return ret
def test1(num_obs):
# 1 node contains a 4-node order-2 Hmm
dimension = 2
obs_gen = make_data_generator(dimension)
obs_list = [obs_gen.next() for i in xrange(num_obs)]
# GmmMgr setup
num_models = 20
models = make_standard_gmms(dimension, num_models)
gmm_mgr1 = GmmMgr(models[0:10])
gmm_mgr2 = GmmMgr(models[10:20])
# Hmm setup
# Make one Hmm with 4 states and order 2 (self loop, forward 1)
num_states = 4
seed(0)
hmm0 = make_forward_hmm(gmm_mgr1, num_states, 2, exact=True)
hmm_mgr = HmmMgr((hmm0, ))
# TrainingGraph setup
gb = GraphBuilder()
node_id0 = gb.new_node((0, 0))
gr0 = FrozenGraph(gb)
tg0 = TrainingGraph(gr0, hmm_mgr, dict())
valid, ret = validate_training_graph(tg0, gmm_mgr1, hmm_mgr, obs_list, 1,
gmm_mgr2)
return ret
def test1(num_obs):
# 1 node contains a 4-node order-2 Hmm
dimension = 2
obs_gen = make_data_generator(dimension)
obs_list = [obs_gen.next() for i in xrange(num_obs)]
# GmmMgr setup
num_models = 20
models = make_standard_gmms(dimension, num_models)
gmm_mgr1 = GmmMgr(models[0:10])
gmm_mgr2 = GmmMgr(models[10:20])
# Hmm setup
# Make one Hmm with 4 states and order 2 (self loop, forward 1)
num_states = 4
seed(0)
hmm0 = make_forward_hmm(gmm_mgr1, num_states, 2, exact=True)
hmm_mgr = HmmMgr((hmm0,))
# TrainingGraph setup
gb = GraphBuilder()
node_id0 = gb.new_node((0, 0))
gr0 = FrozenGraph(gb)
tg0 = TrainingGraph(gr0, hmm_mgr, dict())
valid, ret = validate_training_graph(tg0, gmm_mgr1, hmm_mgr, obs_list, 1, gmm_mgr2)
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 _sequence_to_linear_graph(s):
gb = GraphBuilder()
start = gb.new_node()
for item in s:
end = gb.new_node()
gb.new_arc(start, end, item)
start = end
return FrozenGraph(gb)
def __init__(self, sendee=None, sending=True, label=None):
self._sendee = None
self._sending = sending
if sendee is not None:
self.set_sendee(sendee)
self._label = label if label is not None else type(self).__name__
gb = GraphBuilder()
node = gb.new_node(self._label)
self._graph = FrozenGraph(gb)
def _test11():
# A reduced version of test10
ret = ""
# GmmMgr setup
num_states = 2
dimension = 2
models = []
for i in xrange(num_states):
dm = DummyModel(dimension, 1.0)
models.append(dm)
gmm_mgr = GmmMgr(models)
gb = GraphBuilder()
node_id0 = gb.new_node((0, 0))
node_id1 = gb.new_node((1, 1))
node_id2 = gb.new_node((2, 1))
node_id3 = gb.new_node((3, 1))
node_id4 = gb.new_node((4, 2))
# The topology here is slightly complex than the previous example
arc_id = gb.new_arc(node_id0, node_id1)
arc_id = gb.new_arc(node_id1, node_id4)
arc_id = gb.new_arc(node_id0, node_id2)
arc_id = gb.new_arc(node_id2, node_id3)
arc_id = gb.new_arc(node_id3, node_id4)
arc_id = gb.new_arc(node_id2, node_id4)
gr0 = FrozenGraph(gb)
# Make two Hmms with 3 states and order 2 (self loop, forward 1)
# The models in the middle are special and can skip.
seed(0)
hmm0 = make_forward_hmm(gmm_mgr, num_states, order=2, exact=False)
hmm1 = Hmm(1)
trans = array(((0.0, 0.5, 0.5), (0.0, 0.5, 0.5), (0.0, 0.0, 0.0)))
hmm1.build_model(gmm_mgr, (0, ), 1, 1, trans)
hmm2 = make_forward_hmm(gmm_mgr, num_states, order=2, exact=True)
hmm_mgr = HmmMgr((hmm0, hmm1, hmm2))
spd = {}
spd[(0, 1)] = (0.4, )
spd[(0, 2)] = (0.6, )
spd[(2, 3)] = (0.4, )
spd[(2, 4)] = (0.6, )
tg0 = TrainingGraph(gr0, hmm_mgr, split_prob_dict=spd)
if do_display:
tg0.dot_display()
tg0.dot_display(expand_hmms=True)
with DebugPrint("bwt_ctsh") if True else DebugPrint():
result_hmm = tg0.convert_to_standalone_hmm()
ret += "\n\n========= TG CONVERTED TO Hmm =========\n\n" + result_hmm.to_string(
full=True)
return ret
def test5(num_obs, do_display=False):
# A test in which one of the HMMs has a transition from an input directly to
# an output, so it can behave as an epsilon. This node is between two other
# nodes in a linear arrangement.
# Data generator setup and data generation
dimension = 2
obs_gen = make_data_generator(dimension)
obs_list = [obs_gen.next() for i in xrange(num_obs)]
# GmmMgr setup
num_models = 20
models = make_standard_gmms(dimension, num_models)
gmm_mgr1 = GmmMgr(models[0:10])
gmm_mgr2 = GmmMgr(models[10:20])
# Hmm setup
# Make two Hmms with 2 states and order 2 (self loop, forward 1) The model
# in the middle is special in that it can skip directly from the input state
# to the output state.
seed(0)
num_states = 2
hmm0 = make_forward_hmm(gmm_mgr1, num_states, 2, exact=False)
hmm1 = Hmm(1)
trans = array(((0.0, 0.5, 0.5), (0.0, 0.5, 0.5), (0.0, 0.0, 0.0)))
hmm1.build_model(gmm_mgr1, (0, ), 1, 1, trans)
hmm2 = make_forward_hmm(gmm_mgr1, num_states, 2, exact=False)
hmm_mgr = HmmMgr((hmm0, hmm1, hmm2))
# TrainingGraph setup
gb = GraphBuilder()
node_id0 = gb.new_node((0, 0))
node_id1 = gb.new_node((1, 1))
# node_id2 = gb.new_node((2,2))
arc_id = gb.new_arc(node_id0, node_id1)
# arc_id = gb.new_arc(node_id1, node_id2)
gr0 = FrozenGraph(gb)
tg0 = TrainingGraph(gr0, hmm_mgr, split_prob_dict=dict())
if do_display:
tg0.dot_display()
tg0.dot_display(expand_hmms=True)
valid, ret = validate_training_graph(tg0, gmm_mgr1, hmm_mgr, obs_list, 1,
gmm_mgr2)
return ret
def test5(num_obs, do_display=False):
# A test in which one of the HMMs has a transition from an input directly to
# an output, so it can behave as an epsilon. This node is between two other
# nodes in a linear arrangement.
# Data generator setup and data generation
dimension = 2
obs_gen = make_data_generator(dimension)
obs_list = [obs_gen.next() for i in xrange(num_obs)]
# GmmMgr setup
num_models = 20
models = make_standard_gmms(dimension, num_models)
gmm_mgr1 = GmmMgr(models[0:10])
gmm_mgr2 = GmmMgr(models[10:20])
# Hmm setup
# Make two Hmms with 2 states and order 2 (self loop, forward 1) The model
# in the middle is special in that it can skip directly from the input state
# to the output state.
seed(0)
num_states = 2
hmm0 = make_forward_hmm(gmm_mgr1, num_states, 2, exact=False)
hmm1 = Hmm(1)
trans = array(((0.0, 0.5, 0.5), (0.0, 0.5, 0.5), (0.0, 0.0, 0.0)))
hmm1.build_model(gmm_mgr1, (0,), 1, 1, trans)
hmm2 = make_forward_hmm(gmm_mgr1, num_states, 2, exact=False)
hmm_mgr = HmmMgr((hmm0, hmm1, hmm2))
# TrainingGraph setup
gb = GraphBuilder()
node_id0 = gb.new_node((0, 0))
node_id1 = gb.new_node((1, 1))
# node_id2 = gb.new_node((2,2))
arc_id = gb.new_arc(node_id0, node_id1)
# arc_id = gb.new_arc(node_id1, node_id2)
gr0 = FrozenGraph(gb)
tg0 = TrainingGraph(gr0, hmm_mgr, split_prob_dict=dict())
if do_display:
tg0.dot_display()
tg0.dot_display(expand_hmms=True)
valid, ret = validate_training_graph(tg0, gmm_mgr1, hmm_mgr, obs_list, 1, gmm_mgr2)
return ret
def test3(num_obs):
# Each of the 4 nodes contains a 4 (or 6)-node order-3 Hmm; the nodes are connected in a
# diamond pattern
dimension = 2
obs_gen = make_data_generator(dimension)
obs_list = [obs_gen.next() for i in xrange(num_obs)]
# GmmMgr setup
num_states = 4
num_models = 20
models = make_standard_gmms(dimension, num_models)
gmm_mgr1 = GmmMgr(models[0:10])
gmm_mgr2 = GmmMgr(models[10:20])
# Hmm setup
# Make four Hmms with 4 (or 6) states and order 3 (self loop, forward 1, forward 2)
seed(0)
hmm0 = make_forward_hmm(gmm_mgr1, num_states, 3, exact=True)
# NB: the asymetry between the two successors is a key part of this test; otherwise,
# there are no differences between the transition probs going to these successors,
# which is the tricky case
hmm1 = make_forward_hmm(gmm_mgr1, num_states + 2, 3, exact=True)
hmm2 = make_forward_hmm(gmm_mgr1, num_states, 3, exact=True)
hmm3 = make_forward_hmm(gmm_mgr1, num_states, 3, exact=True)
hmm_mgr = HmmMgr((hmm0, hmm1, hmm2, hmm3))
# TrainingGraph setup
gb = GraphBuilder()
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, 3))
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)
valid, ret = validate_training_graph(tg0, gmm_mgr1, hmm_mgr, obs_list, 1, gmm_mgr2)
return ret
def _test11():
# A reduced version of test10
ret = ""
# GmmMgr setup
num_states = 2
dimension = 2
models = []
for i in xrange(num_states):
dm = DummyModel(dimension, 1.0)
models.append(dm)
gmm_mgr = GmmMgr(models)
gb = GraphBuilder()
node_id0 = gb.new_node((0, 0))
node_id1 = gb.new_node((1, 1))
node_id2 = gb.new_node((2, 1))
node_id3 = gb.new_node((3, 1))
node_id4 = gb.new_node((4, 2))
# The topology here is slightly complex than the previous example
arc_id = gb.new_arc(node_id0, node_id1)
arc_id = gb.new_arc(node_id1, node_id4)
arc_id = gb.new_arc(node_id0, node_id2)
arc_id = gb.new_arc(node_id2, node_id3)
arc_id = gb.new_arc(node_id3, node_id4)
arc_id = gb.new_arc(node_id2, node_id4)
gr0 = FrozenGraph(gb)
# Make two Hmms with 3 states and order 2 (self loop, forward 1)
# The models in the middle are special and can skip.
seed(0)
hmm0 = make_forward_hmm(gmm_mgr, num_states, order=2, exact=False)
hmm1 = Hmm(1)
trans = array(((0.0, 0.5, 0.5), (0.0, 0.5, 0.5), (0.0, 0.0, 0.0)))
hmm1.build_model(gmm_mgr, (0,), 1, 1, trans)
hmm2 = make_forward_hmm(gmm_mgr, num_states, order=2, exact=True)
hmm_mgr = HmmMgr((hmm0, hmm1, hmm2))
spd = {}
spd[(0, 1)] = (0.4,)
spd[(0, 2)] = (0.6,)
spd[(2, 3)] = (0.4,)
spd[(2, 4)] = (0.6,)
tg0 = TrainingGraph(gr0, hmm_mgr, split_prob_dict=spd)
if do_display:
tg0.dot_display()
tg0.dot_display(expand_hmms=True)
with DebugPrint("bwt_ctsh") if True else DebugPrint():
result_hmm = tg0.convert_to_standalone_hmm()
ret += "\n\n========= TG CONVERTED TO Hmm =========\n\n" + result_hmm.to_string(full=True)
return ret
def _test9():
# Like test8, but now HMMs have multiple inputs and outputs.
ret = ""
# GmmMgr setup
num_states = 3
dimension = 2
models = []
for i in xrange(num_states):
dm = DummyModel(dimension, 1.0)
models.append(dm)
gmm_mgr = GmmMgr(models)
gb = GraphBuilder()
node_id0 = gb.new_node((0, 0))
node_id1 = gb.new_node((1, 1))
node_id2 = gb.new_node((2, 1))
node_id3 = gb.new_node((3, 1))
node_id4 = gb.new_node((4, 1))
node_id5 = gb.new_node((5, 2))
arc_id = gb.new_arc(node_id0, node_id1)
arc_id = gb.new_arc(node_id1, node_id2)
arc_id = gb.new_arc(node_id2, node_id3)
arc_id = gb.new_arc(node_id3, node_id4)
arc_id = gb.new_arc(node_id4, node_id5)
gr0 = FrozenGraph(gb)
# Make two Hmms with 3 states and order 3 (self loop, forward 1, forward 2)
# The models in the middle are special and can skip directly
seed(0)
hmm0 = make_forward_hmm(gmm_mgr, num_states, order=3, exact=True)
hmm1 = Hmm(1)
trans = array(((0.0, 0.0, 0.0, 0.5, 0.5, 0.0,
0.0), (0.0, 0.0, 0.0, 0.5, 0.0, 0.5,
0.0), (0.0, 0.0, 0.0, 0.5, 0.0, 0.0,
0.5), (0.0, 0.0, 0.0, 0.5, 0.35, 0.1, 0.05),
(0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0), (0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0), (0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0)))
hmm1.build_model(gmm_mgr, (0, ), 3, 3, trans)
hmm2 = make_forward_hmm(gmm_mgr, num_states, order=3, exact=True)
hmm_mgr = HmmMgr((hmm0, hmm1, hmm2))
with DebugPrint("bwt_vrfy") if False else DebugPrint():
tg0 = TrainingGraph(gr0, hmm_mgr, split_prob_dict=dict())
result_hmm = tg0.convert_to_standalone_hmm()
ret += "\n\n========= TG CONVERTED TO Hmm =========\n\n" + result_hmm.to_string(
full=True)
return ret
def __init__(self, stream):
docgen = YamldataGenerator(stream)
reader = YamldataReader(docgen, stream_type=self.STREAM_TYPE, stream_version=self.STREAM_VERSION, header_only=True)
self.runtime_objects = IndexedObjectSet(docgen)
self.dataflow_graph = FrozenGraph(SerializedGraphTables(docgen))
# until we've sorted out the yaml issues
# make a linear graph with arc labels indexing into runtime_objects
builder = GraphBuilder()
startid = builder.new_node_label_is_id()
for arc_label, obj in enumerate(self.runtime_objects):
endid = builder.new_node_label_is_id()
arcid = builder.new_arc(startid, endid, arc_label)
startid = endid
self.graph = FrozenGraph(builder)
for i in self.runtime_objects:
self.last = i[-1]
def _test9():
# Like test8, but now HMMs have multiple inputs and outputs.
ret = ""
# GmmMgr setup
num_states = 3
dimension = 2
models = []
for i in xrange(num_states):
dm = DummyModel(dimension, 1.0)
models.append(dm)
gmm_mgr = GmmMgr(models)
gb = GraphBuilder()
node_id0 = gb.new_node((0, 0))
node_id1 = gb.new_node((1, 1))
node_id2 = gb.new_node((2, 1))
node_id3 = gb.new_node((3, 1))
node_id4 = gb.new_node((4, 1))
node_id5 = gb.new_node((5, 2))
arc_id = gb.new_arc(node_id0, node_id1)
arc_id = gb.new_arc(node_id1, node_id2)
arc_id = gb.new_arc(node_id2, node_id3)
arc_id = gb.new_arc(node_id3, node_id4)
arc_id = gb.new_arc(node_id4, node_id5)
gr0 = FrozenGraph(gb)
# Make two Hmms with 3 states and order 3 (self loop, forward 1, forward 2)
# The models in the middle are special and can skip directly
seed(0)
hmm0 = make_forward_hmm(gmm_mgr, num_states, order=3, exact=True)
hmm1 = Hmm(1)
trans = array(
(
(0.0, 0.0, 0.0, 0.5, 0.5, 0.0, 0.0),
(0.0, 0.0, 0.0, 0.5, 0.0, 0.5, 0.0),
(0.0, 0.0, 0.0, 0.5, 0.0, 0.0, 0.5),
(0.0, 0.0, 0.0, 0.5, 0.35, 0.1, 0.05),
(0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0),
(0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0),
(0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0),
)
)
hmm1.build_model(gmm_mgr, (0,), 3, 3, trans)
hmm2 = make_forward_hmm(gmm_mgr, num_states, order=3, exact=True)
hmm_mgr = HmmMgr((hmm0, hmm1, hmm2))
with DebugPrint("bwt_vrfy") if False else DebugPrint():
tg0 = TrainingGraph(gr0, hmm_mgr, split_prob_dict=dict())
result_hmm = tg0.convert_to_standalone_hmm()
ret += "\n\n========= TG CONVERTED TO Hmm =========\n\n" + result_hmm.to_string(full=True)
return ret
def __init__(self, processors, sendee=None, sending=True):
super(ChainProcessor, self).__init__(sendee, sending=sending)
processors = tuple(processors)
if not processors:
raise ValueError("expected at least one element in processors chain, got zero")
# front of the chain, where we push stuff
self.head = processors[0]
# create chain of processors, linking each processor's send function to
# each successor's process() function....
gb = GraphBuilder()
nodes, starts, ends = gb.add_graph(processors[0].graph)
assert len(starts) == 1 and len(ends) == 1
if len(processors) > 1:
succ_iter = iter(processors)
succ_iter.next()
for pred, succ in izip(processors, succ_iter):
pred.set_sendee(succ.process)
nodes, new_starts, new_ends = gb.add_graph(succ.graph)
gb.new_arc(ends[0], new_starts[0])
starts, ends = new_starts, new_ends
assert len(starts) == 1 and len(ends) == 1
assert succ == processors[-1]
# set up the list that will collect what the final element pushes
self.collector = list()
processors[-1].set_sendee(self.collector.append)
self._graph = FrozenGraph(gb)
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 _sequence_to_linear_graph(s):
gb = GraphBuilder()
start = gb.new_node()
for item in s:
end = gb.new_node()
gb.new_arc(start, end, item)
start = end
return FrozenGraph(gb)
def test3(num_obs):
# Each of the 4 nodes contains a 4 (or 6)-node order-3 Hmm; the nodes are connected in a
# diamond pattern
dimension = 2
obs_gen = make_data_generator(dimension)
obs_list = [obs_gen.next() for i in xrange(num_obs)]
# GmmMgr setup
num_states = 4
num_models = 20
models = make_standard_gmms(dimension, num_models)
gmm_mgr1 = GmmMgr(models[0:10])
gmm_mgr2 = GmmMgr(models[10:20])
# Hmm setup
# Make four Hmms with 4 (or 6) states and order 3 (self loop, forward 1, forward 2)
seed(0)
hmm0 = make_forward_hmm(gmm_mgr1, num_states, 3, exact=True)
# NB: the asymetry between the two successors is a key part of this test; otherwise,
# there are no differences between the transition probs going to these successors,
# which is the tricky case
hmm1 = make_forward_hmm(gmm_mgr1, num_states + 2, 3, exact=True)
hmm2 = make_forward_hmm(gmm_mgr1, num_states, 3, exact=True)
hmm3 = make_forward_hmm(gmm_mgr1, num_states, 3, exact=True)
hmm_mgr = HmmMgr((hmm0, hmm1, hmm2, hmm3))
# TrainingGraph setup
gb = GraphBuilder()
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, 3))
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)
valid, ret = validate_training_graph(tg0, gmm_mgr1, hmm_mgr, obs_list, 1,
gmm_mgr2)
return ret
def _lattice_nbest_align(l1, l2, cost_fn=None, substring_cost=False):
if not (isinstance(l1, FrozenGraph) and isinstance(l2, FrozenGraph)):
raise ValueError("lattice_nbest_align needs two FrozenGraph arguments")
if cost_fn is None:
cost_fn = _std_cost
else:
cost_fn = _make_safe_cost_fn(cost_fn)
l1_canon = l1.get_canonical_DAG()
l2_canon = l2.get_canonical_DAG()
len1 = l1_canon.get_num_arcs()
len2 = l2_canon.get_num_arcs()
assert l1_canon.is_lattice()
assert l2_canon.is_lattice()
terms = l1_canon.get_terminals()
lat_start1 = terms[0][0]
lat_end1 = terms[1][0]
terms = l2_canon.get_terminals()
lat_start2 = terms[0][0]
lat_end2 = terms[1][0]
# The lattices to be aligned, l1 and l2, have labels on the arcs;
# any labels on their nodes will be ignored. We begin by finding
# a topological ordering of the lattice arcs, so that each arc is
# assigned a non-negative index. We construct a 2-D lattice, g,
# with nodes representing pairs of arcs in the l1 and l2. An
# additional row and column on the top and left of g represent an
# initial position prior consuming any tokens from l1 and l2,
# respectively. The arcs in g are labeled with triples giving the
# orginal labels from l1 and l2 (or None if the arc's start node
# is on the left side or top row) and the local cost of the edit
# represented by the arc. There are two slightly tricky parts.
# First, because the alignment is done on lattices, the cost
# lattice may have links which cross several rows or columns,
# depending on the adjacencies of l1 and l2. Second, there's an
# offset of 1 between the arc numbering for l1 and l2 and the node
# numbering in g, because of the initial row and column. Thus,
# the node in g corresponding to the arc pair <a1, a2> is at
# indices <a1+1, a2+1>. One egregious (but very handy) abuse of
# this arrangement is the occasional use of -1 as an ersatz arcID,
# which will be converted to a 0 index into g. Finally, because
# the graph iterpath function requires a single start and end
# node, we tie all the potential end nodes in g to a special end
# node with "ground" arcs. A node in g is a potential end node if
# the pair of arcs it represents are each incident on the terminal
# node of their respective lattices.
gb = GraphBuilder()
# Initialize cost lattice nodes
node_array = [[gb.new_node() for j in range(len2 + 1)]
for i in range(len1 + 1)]
# We use this single node to tie all end nodes together
end_node = gb.new_node()
# Add first row of arcs
for i in xrange(len1):
start, end, x = l1_canon.get_arc(i)
insert_cost = cost_fn(x, None)
if end == lat_end1 and len2 == 0:
gb.new_arc(node_array[i + 1][0], end_node, (None, None, 0))
# print "Added ground arc for l1 arc from %d to %d with label %s" % (start, end, x)
pred_arcs = l1_canon.get_node_in_arcs(start)
if start == lat_start1:
pred_arcs.append(-1)
for arc in pred_arcs:
gb.new_arc(node_array[arc + 1][0], node_array[i + 1][0],
(x, None, insert_cost))
# print ("Processed l1 arc from %d to %d with label %s - added %d arcs"
# % (start, end, x, len(pred_arcs)))
# Add first column of arcs
for j in xrange(len2):
start, end, y = l2_canon.get_arc(j)
delete_cost = 0 if substring_cost else cost_fn(None, y)
if end == lat_end2 and len1 == 0:
gb.new_arc(node_array[0][j + 1], end_node, (None, None, 0))
# print "Added ground arc for l1 arc from %d to %d with label %s" % (start, end, y)
pred_arcs = l2_canon.get_node_in_arcs(start)
if start == lat_start2:
pred_arcs.append(-1)
for arc in pred_arcs:
gb.new_arc(node_array[0][arc + 1], node_array[0][j + 1],
(None, y, delete_cost))
# print ("Processed l1 arc from %d to %d with label %s - added %d arcs"
# % (start, end, y, len(pred_arcs)))
# Construct remainder of cost lattice
for i in xrange(len1):
for j in xrange(len2):
start1, end1, x = l1_canon.get_arc(i)
start2, end2, y = l2_canon.get_arc(j)
pred_arcs1 = l1_canon.get_node_in_arcs(start1)
pred_arcs2 = l2_canon.get_node_in_arcs(start2)
if start1 == lat_start1:
pred_arcs1.append(-1)
if start2 == lat_start2:
pred_arcs2.append(-1)
insert_cost = cost_fn(x, None)
delete_cost = 0 if (substring_cost and i == len1 - 1) else cost_fn(
None, y)
subst_cost = cost_fn(x, y)
num_added = 0
if end1 == lat_end1 and end2 == lat_end2:
gb.new_arc(node_array[i + 1][j + 1], end_node, (None, None, 0))
# print "Added ground arc for l1,l1 arcs with labels %s, %s" % (x,y)
for arc in pred_arcs1:
gb.new_arc(node_array[arc + 1][j + 1],
node_array[i + 1][j + 1], (x, None, insert_cost))
num_added += 1
for arc in pred_arcs2:
gb.new_arc(node_array[i + 1][arc + 1],
node_array[i + 1][j + 1], (None, y, delete_cost))
num_added += 1
for arc1 in pred_arcs1:
for arc2 in pred_arcs2:
gb.new_arc(node_array[arc1 + 1][arc2 + 1],
node_array[i + 1][j + 1], (x, y, subst_cost))
num_added += 1
# print ("Processed l1 arc from %d to %d with label %s "
# "and l1 arc from %d to %d with label %s - added %d arcs"
# % (start1, end1, x, start2, end2, y, num_added))
g = FrozenGraph(gb)
assert g.is_lattice()
# print "g.get_terminals() = %s, %s" % g.get_terminals()
def graph_cost_fn(label):
return label[2]
def iter_helper(path):
arc_labels = [path[2].get_arc_label(arc) for arc in path[1][:-1]]
return (path[0], tuple(arc_labels))
return imap(iter_helper,
g.iterpaths(graph_cost_fn, node_array[0][0], end_node))
def _test10():
# Like test9, but now HMMs are arranged in a diamond pattern so inter-HMM
# probabilities come into play
ret = ""
# GmmMgr setup
num_states = 3
dimension = 2
models = []
for i in xrange(num_states):
dm = DummyModel(dimension, 1.0)
models.append(dm)
gmm_mgr = GmmMgr(models)
gb = GraphBuilder()
node_id0 = gb.new_node((0, 0))
node_id1 = gb.new_node((1, 1))
node_id2 = gb.new_node((2, 1))
node_id3 = gb.new_node((3, 1))
node_id4 = gb.new_node((4, 1))
node_id5 = gb.new_node((5, 2))
# The topology here is more complex than previous examples
arc_id = gb.new_arc(node_id0, node_id1)
arc_id = gb.new_arc(node_id1, node_id5)
arc_id = gb.new_arc(node_id0, node_id2)
arc_id = gb.new_arc(node_id2, node_id3)
arc_id = gb.new_arc(node_id3, node_id4)
arc_id = gb.new_arc(node_id3, node_id5)
arc_id = gb.new_arc(node_id4, node_id5)
gr0 = FrozenGraph(gb)
# Make two Hmms with 3 states and order 3 (self loop, forward 1, forward 2)
# The models in the middle are special and can skip.
seed(0)
hmm0 = make_forward_hmm(gmm_mgr, num_states, order=3, exact=True)
hmm1 = Hmm(1)
trans = array(
(
(0.0, 0.0, 0.0, 0.5, 0.5, 0.0, 0.0),
(0.0, 0.0, 0.0, 0.5, 0.0, 0.5, 0.0),
(0.0, 0.0, 0.0, 0.5, 0.0, 0.0, 0.5),
(0.0, 0.0, 0.0, 0.5, 0.35, 0.1, 0.05),
(0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0),
(0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0),
(0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0),
)
)
hmm1.build_model(gmm_mgr, (0,), 3, 3, trans)
hmm2 = make_forward_hmm(gmm_mgr, num_states, order=3, exact=True)
hmm_mgr = HmmMgr((hmm0, hmm1, hmm2))
spd = {}
spd[(0, 1)] = (0.4, 0.3, 0.8)
spd[(0, 2)] = (0.6, 0.7, 0.2)
spd[(3, 4)] = (0.4, 0.3, 0.8)
spd[(3, 5)] = (0.6, 0.7, 0.2)
tg0 = TrainingGraph(gr0, hmm_mgr, split_prob_dict=spd)
with DebugPrint("bwt_ctsh") if True else DebugPrint():
result_hmm = tg0.convert_to_standalone_hmm()
ret += "\n\n========= TG CONVERTED TO Hmm =========\n\n" + result_hmm.to_string(full=True)
return ret
def _lattice_nbest_align(l1, l2, cost_fn = None, substring_cost = False):
if not (isinstance(l1, FrozenGraph) and isinstance(l2, FrozenGraph)):
raise ValueError("lattice_nbest_align needs two FrozenGraph arguments")
if cost_fn is None:
cost_fn = _std_cost
else:
cost_fn = _make_safe_cost_fn(cost_fn)
l1_canon = l1.get_canonical_DAG()
l2_canon = l2.get_canonical_DAG()
len1 = l1_canon.get_num_arcs()
len2 = l2_canon.get_num_arcs()
assert l1_canon.is_lattice()
assert l2_canon.is_lattice()
terms = l1_canon.get_terminals()
lat_start1 = terms[0][0]
lat_end1 = terms[1][0]
terms = l2_canon.get_terminals()
lat_start2 = terms[0][0]
lat_end2 = terms[1][0]
# The lattices to be aligned, l1 and l2, have labels on the arcs;
# any labels on their nodes will be ignored. We begin by finding
# a topological ordering of the lattice arcs, so that each arc is
# assigned a non-negative index. We construct a 2-D lattice, g,
# with nodes representing pairs of arcs in the l1 and l2. An
# additional row and column on the top and left of g represent an
# initial position prior consuming any tokens from l1 and l2,
# respectively. The arcs in g are labeled with triples giving the
# orginal labels from l1 and l2 (or None if the arc's start node
# is on the left side or top row) and the local cost of the edit
# represented by the arc. There are two slightly tricky parts.
# First, because the alignment is done on lattices, the cost
# lattice may have links which cross several rows or columns,
# depending on the adjacencies of l1 and l2. Second, there's an
# offset of 1 between the arc numbering for l1 and l2 and the node
# numbering in g, because of the initial row and column. Thus,
# the node in g corresponding to the arc pair <a1, a2> is at
# indices <a1+1, a2+1>. One egregious (but very handy) abuse of
# this arrangement is the occasional use of -1 as an ersatz arcID,
# which will be converted to a 0 index into g. Finally, because
# the graph iterpath function requires a single start and end
# node, we tie all the potential end nodes in g to a special end
# node with "ground" arcs. A node in g is a potential end node if
# the pair of arcs it represents are each incident on the terminal
# node of their respective lattices.
gb = GraphBuilder()
# Initialize cost lattice nodes
node_array = [[gb.new_node() for j in range(len2+1)] for i in range(len1+1)]
# We use this single node to tie all end nodes together
end_node = gb.new_node()
# Add first row of arcs
for i in xrange(len1):
start, end, x = l1_canon.get_arc(i)
insert_cost = cost_fn(x, None)
if end == lat_end1 and len2 == 0:
gb.new_arc(node_array[i+1][0], end_node, (None, None, 0))
# print "Added ground arc for l1 arc from %d to %d with label %s" % (start, end, x)
pred_arcs = l1_canon.get_node_in_arcs(start)
if start == lat_start1:
pred_arcs.append(-1)
for arc in pred_arcs:
gb.new_arc(node_array[arc+1][0], node_array[i+1][0], (x, None, insert_cost))
# print ("Processed l1 arc from %d to %d with label %s - added %d arcs"
# % (start, end, x, len(pred_arcs)))
# Add first column of arcs
for j in xrange(len2):
start, end, y = l2_canon.get_arc(j)
delete_cost = 0 if substring_cost else cost_fn(None, y)
if end == lat_end2 and len1 == 0:
gb.new_arc(node_array[0][j+1], end_node, (None, None, 0))
# print "Added ground arc for l1 arc from %d to %d with label %s" % (start, end, y)
pred_arcs = l2_canon.get_node_in_arcs(start)
if start == lat_start2:
pred_arcs.append(-1)
for arc in pred_arcs:
gb.new_arc(node_array[0][arc+1], node_array[0][j+1], (None, y, delete_cost))
# print ("Processed l1 arc from %d to %d with label %s - added %d arcs"
# % (start, end, y, len(pred_arcs)))
# Construct remainder of cost lattice
for i in xrange(len1):
for j in xrange(len2):
start1, end1, x = l1_canon.get_arc(i)
start2, end2, y = l2_canon.get_arc(j)
pred_arcs1 = l1_canon.get_node_in_arcs(start1)
pred_arcs2 = l2_canon.get_node_in_arcs(start2)
if start1 == lat_start1:
pred_arcs1.append(-1)
if start2 == lat_start2:
pred_arcs2.append(-1)
insert_cost = cost_fn(x, None)
delete_cost = 0 if (substring_cost and i == len1 - 1) else cost_fn(None, y)
subst_cost = cost_fn(x, y)
num_added = 0
if end1 == lat_end1 and end2 == lat_end2:
gb.new_arc(node_array[i+1][j+1], end_node, (None, None, 0))
# print "Added ground arc for l1,l1 arcs with labels %s, %s" % (x,y)
for arc in pred_arcs1:
gb.new_arc(node_array[arc+1][j+1], node_array[i+1][j+1], (x, None, insert_cost))
num_added += 1
for arc in pred_arcs2:
gb.new_arc(node_array[i+1][arc+1], node_array[i+1][j+1], (None, y, delete_cost))
num_added += 1
for arc1 in pred_arcs1:
for arc2 in pred_arcs2:
gb.new_arc(node_array[arc1+1][arc2+1], node_array[i+1][j+1], (x, y, subst_cost))
num_added += 1
# print ("Processed l1 arc from %d to %d with label %s "
# "and l1 arc from %d to %d with label %s - added %d arcs"
# % (start1, end1, x, start2, end2, y, num_added))
g = FrozenGraph(gb)
assert g.is_lattice()
# print "g.get_terminals() = %s, %s" % g.get_terminals()
def graph_cost_fn(label): return label[2]
def iter_helper(path):
arc_labels = [path[2].get_arc_label(arc) for arc in path[1][:-1]]
return(path[0], tuple(arc_labels))
return imap(iter_helper,
g.iterpaths(graph_cost_fn, node_array[0][0], end_node))
def _test10():
# Like test9, but now HMMs are arranged in a diamond pattern so inter-HMM
# probabilities come into play
ret = ""
# GmmMgr setup
num_states = 3
dimension = 2
models = []
for i in xrange(num_states):
dm = DummyModel(dimension, 1.0)
models.append(dm)
gmm_mgr = GmmMgr(models)
gb = GraphBuilder()
node_id0 = gb.new_node((0, 0))
node_id1 = gb.new_node((1, 1))
node_id2 = gb.new_node((2, 1))
node_id3 = gb.new_node((3, 1))
node_id4 = gb.new_node((4, 1))
node_id5 = gb.new_node((5, 2))
# The topology here is more complex than previous examples
arc_id = gb.new_arc(node_id0, node_id1)
arc_id = gb.new_arc(node_id1, node_id5)
arc_id = gb.new_arc(node_id0, node_id2)
arc_id = gb.new_arc(node_id2, node_id3)
arc_id = gb.new_arc(node_id3, node_id4)
arc_id = gb.new_arc(node_id3, node_id5)
arc_id = gb.new_arc(node_id4, node_id5)
gr0 = FrozenGraph(gb)
# Make two Hmms with 3 states and order 3 (self loop, forward 1, forward 2)
# The models in the middle are special and can skip.
seed(0)
hmm0 = make_forward_hmm(gmm_mgr, num_states, order=3, exact=True)
hmm1 = Hmm(1)
trans = array(((0.0, 0.0, 0.0, 0.5, 0.5, 0.0,
0.0), (0.0, 0.0, 0.0, 0.5, 0.0, 0.5,
0.0), (0.0, 0.0, 0.0, 0.5, 0.0, 0.0,
0.5), (0.0, 0.0, 0.0, 0.5, 0.35, 0.1, 0.05),
(0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0), (0.0, 0.0, 0.0, 0.0, 0.0, 0.0,
0.0), (0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0)))
hmm1.build_model(gmm_mgr, (0, ), 3, 3, trans)
hmm2 = make_forward_hmm(gmm_mgr, num_states, order=3, exact=True)
hmm_mgr = HmmMgr((hmm0, hmm1, hmm2))
spd = {}
spd[(0, 1)] = (0.4, 0.3, 0.8)
spd[(0, 2)] = (0.6, 0.7, 0.2)
spd[(3, 4)] = (0.4, 0.3, 0.8)
spd[(3, 5)] = (0.6, 0.7, 0.2)
tg0 = TrainingGraph(gr0, hmm_mgr, split_prob_dict=spd)
with DebugPrint("bwt_ctsh") if True else DebugPrint():
result_hmm = tg0.convert_to_standalone_hmm()
ret += "\n\n========= TG CONVERTED TO Hmm =========\n\n" + result_hmm.to_string(
full=True)
return ret