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 _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 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 _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 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
raise IOError("No global options found in %s" % (filename,)) opts = result['options'] if 'models' not in result: raise IOError("No models found in %s!" % (filename,)) models = result['models'] if 'vecsize' not in opts: raise IOError("No vecsize option found in %s" % (filename,)) dim = opts['vecsize'] if 'covar' not in opts: covar_type = GaussianMixtureModel.DIAGONAL_COVARIANCE else: if opts['covar'] not in covar_map: raise IOError("Unknown covar option %s found in %s" % (opts['covar'], filename,)) covar_type = covar_map[opts['covar']] dim = opts['vecsize'] hmm_mgr = HmmMgr(dim) gmm_mgr = GmmMgr(dim) hmms = [] names = [] unnamed_index = 0 for label, m in models: assert label == 'HMM' dc and dc("m = \n%s" % (pformat(m),)) dc and dc("m.keys() = \n%s" % (m.keys(),)) if m.hasattr.decl: name = m.decl else: name = ("UnnamedModel%d" % unnamed_index) unnamed_index += 1 n = m.numstates - 2 # HTK numstates counts virtual entry and exit states hmm = Hmm(n, log_domain)
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