def main(n_iter, path_logd=""):
if path_logd is None:
path_logd = ""
if path_logd != "":
path_logd = prepare_logd(path_logd,
redirect_std=False,
timestamp=False)
print("sampling")
smodel = get_model_by_size("small")
xseqs, yseqs = sample_from_model(smodel, n_samples=50, len_seq=50)
# several models will be added
name_nstates = {
"small (1,1,1)": (1, 1, 1),
"med (2,2,2)": (2, 2, 2),
"large (5,5,5)": (5, 5, 5)
}
models = {}
for name, n_states in name_nstates.items():
m = gen_model(n_states)
models[name] = m
print("benchmark training")
benchs = {}
for name, model in models.items():
sec = time_train(model, xseqs, yseqs, n_iter=n_iter)
benchs[name] = sec
print(format_bench_result(benchs, n_iter),
file=open(os.path.join(path_logd, "benchs.txt"), 'w'))
pickle.dump(benchs, open(os.path.join(path_logd, "benchs.pkl"), "wb"))
def main(path_logd="", size="med"):
"""
incremental model selection
model size: med
"""
if path_logd != "":
path_logd = prepare_logd(path_logd)
print(__name__)
print(inspect.getsource(main))
smodel = get_model_by_size(size)
xseqs, yseqs = sample_from_model(smodel, n_samples=200, len_seq=50)
models = incremental_model_selection(xseqs, yseqs, stop_threshold=1e-4, shrink_threshold=1e-2,
max_iter=200, verbose=True, verbose_level=1,
max_n_states=5)
for i in range(min(3, len(models))):
print("{} th model".format(i))
models[i]._print_states()
pickle.dump(models, open(path_logd + "models.pkl", "wb"))
print("detailed model")
models[0]._print_states(verbose_level=2)
def main(path_logd="", size="med"):
"""
incremental model selection
model size: med
"""
if path_logd != "":
path_logd = prepare_logd(path_logd)
print(__name__)
print(inspect.getsource(main))
smodel = get_model_by_size(size)
xseqs, yseqs = sample_from_model(smodel, n_samples=200, len_seq=50)
models = decremental_greedy_selection(xseqs,
yseqs,
stop_threshold=1e-4,
shrink_threshold=1e-2,
max_iter=200,
verbose=True,
verbose_level=1,
max_n_states=5,
sorted=False)
monitor = models[-1].monitor
pickle.dump(monitor, open(path_logd + "monitor.pkl", "wb"))
pickle.dump(models, open(path_logd + "models.pkl", "wb"))
models.sort(key=lambda m: -m._last_score)
print("detailed model")
models[0]._print_states(verbose_level=2)
def main():
"""
PHMM small
sample from small mode and fit on PHMM
"""
print(__name__)
print(inspect.getsource(main))
smodel = small_model()
xseqs, yseqs = sample_from_model(smodel, n_samples=200, len_seq=30)
results = []
max_nins = 5
n_same_setting = 3
for n_xins in range(1, max_nins + 1):
for n_yins in range(1, max_nins + 1):
for i in range(1, n_same_setting + 1):
print("===========")
print('{} th same setting'.format(i))
print('training model of ...')
print('n_xins', n_xins)
print('n_yins', n_yins)
fmodel = FABPHMM(n_match_states=1,
n_xins_states=n_xins,
n_yins_states=n_yins,
shrink_threshold=1e-2,
stop_threshold=1e-4,
shrink=False)
fic = fmodel.fit(xseqs, yseqs, max_iter=500, verbose=True, verbose_level=1)
result = {}
result["n_xins"] = n_xins
result["n_yins"] = n_yins
result["fic"] = fic
result["n_hstates"] = fmodel._n_hstates
result["n_xins"] = n_xins
result["n_yins"] = n_yins
results.append(result)
print("end training with...")
print(result)
print()
sys.stdout.flush()
results.sort(key=lambda r: - r["fic"])
print(results)
def main():
"""
PHMM small
sample from small mode and fit on PHMM
"""
print(__name__)
print(inspect.getsource(main))
smodel = small_model()
fmodel = FABPHMM(n_match_states=1,
n_xins_states=10,
n_yins_states=10,
shrink_threshold=1e-2,
stop_threshold=1e-6)
xseqs, yseqs = sample_from_model(smodel, n_samples=200, len_seq=30)
fmodel.fit(xseqs, yseqs, max_iter=10000, verbose=True)
print("end with n_hstates {}".format(fmodel._n_hstates))
def main():
"""
sample from small model and fit to PHMM
start from true props
"""
print(__name__)
print(inspect.getsource(main))
smodel = small_model()
fmodel = PHMM(n_match_states=1,
n_xins_states=1,
n_yins_states=1,
initprob=smodel._initprob,
transprob=smodel._transprob,
emitprob=smodel._emitprob)
xseqs, yseqs = sample_from_model(smodel, n_samples=200, len_seq=30)
fmodel.fit(xseqs, yseqs, max_iter=10000, verbose=True)
def main():
"""
PHMM med
sample from med mode and fit on FABPHMM without shrinkage
"""
print(__name__)
print(inspect.getsource(main))
smodel = med3_model()
xseqs, yseqs = sample_from_model(smodel, n_samples=200, len_seq=50)
fab_small = FABPHMM(n_match_states=1,
n_xins_states=1,
n_yins_states=1,
shrink_threshold=1e-2,
stop_threshold=1e-4,
shrink=False)
fab_med = FABPHMM(n_match_states=1,
n_xins_states=2,
n_yins_states=2,
shrink_threshold=1e-2,
stop_threshold=1e-4,
shrink=False)
phmm_small = PHMM(n_match_states=1,
n_xins_states=1,
n_yins_states=1,
stop_threshold=1e-4)
phmm_med = PHMM(n_match_states=1,
n_xins_states=2,
n_yins_states=2,
stop_threshold=1e-4)
fic_small = fab_small.fit(xseqs, yseqs, max_iter=500, verbose=True, verbose_level=1)
ll_small = fab_small.score(xseqs, yseqs, type="ll")
fic_med = fab_med.fit(xseqs, yseqs, max_iter=500, verbose=True, verbose_level=1)
ll_med = fab_med.score(xseqs, yseqs, type="ll")
phmm_small.fit(xseqs, yseqs, max_iter=500, verbose=True, verbose_level=1)
ll_phmms = phmm_small.score(xseqs, yseqs)
phmm_med.fit(xseqs, yseqs, max_iter=500, verbose=True, verbose_level=1)
ll_phmmm = phmm_med.score(xseqs, yseqs)
ll_org = smodel.score(xseqs, yseqs)
smodel.fit(xseqs, yseqs)
ll_org_fit = smodel.score(xseqs, yseqs)
print("dump fab_small")
fab_small._print_states(verbose_level=2)
print("dump fab_med")
fab_med._print_states(verbose_level=2)
print("dump phmm_small")
phmm_small._print_states(verbose_level=2)
print("dump phmm_med")
phmm_med._print_states(verbose_level=2)
print("fic_small", fic_small)
print("ll_small", ll_small)
print("fic_med", fic_med)
print("ll_med", ll_med)
print("ll_phmms", ll_phmms)
print("ll_phmmm", ll_phmmm)
print("ll_org", ll_org)
print("ll_org", ll_org_fit)
def main():
"""
PHMM med
sample from med mode and fit on FABPHMM without shrinkage
"""
print(__name__)
print(inspect.getsource(main))
smodel = med_model()
xseqs, yseqs = sample_from_model(smodel, n_samples=200, len_seq=50)
results = []
max_nins = 3
max_nmat = 3
n_same_setting = 1
for n_xins in range(1, max_nins + 1):
for n_yins in range(1, max_nins + 1):
for n_match in range(1, max_nmat + 1):
for i in range(1, n_same_setting + 1):
print("===========")
print('{} th same setting'.format(i))
print('training model of ...')
print('n_xins', n_xins)
print('n_yins', n_yins)
print('n_match', n_match)
fmodel = FABPHMM(n_match_states=n_match,
n_xins_states=n_xins,
n_yins_states=n_yins,
shrink_threshold=1e-2,
stop_threshold=1e-4,
shrink=False)
fic = fmodel.fit(xseqs,
yseqs,
max_iter=500,
verbose=True,
verbose_level=1)
result = {}
result["n_xins"] = n_xins
result["n_yins"] = n_yins
result['n_match'] = n_match
result["fic"] = fic
result["n_hstates"] = fmodel._n_hstates
result["model"] = fmodel
results.append(result)
print("end training with...")
print(result)
print()
sys.stdout.flush()
results.sort(key=lambda r: -r["fic"])
for i, r in enumerate(results):
print("{}th result".format(i))
print(r)
r["model"]._print_states(verbose_level=2)