def check_min_data(trial, params, **args):
min_data = trial.suggest_int("min_data", 5, 100)
params = dict(params, min_data=min_data)
score = check(trial, params, **args)
acc = human.humanacc(trial.user_attrs["acc"])
logger.debug("- min_data trial %d: %s [min_data=%s]" %
(trial.number, acc, params["min_data"]))
return score
def check_regular(trial, params, **args):
lambda_l1 = trial.suggest_float("lambda_l1", 1e-8, 10.0)
lambda_l2 = trial.suggest_float("lambda_l2", 1e-8, 10.0)
params = dict(params, lambda_l1=lambda_l1, lambda_l2=lambda_l2)
score = check(trial, params, **args)
acc = human.humanacc(trial.user_attrs["acc"])
logger.debug("- regular trial %d: %s [l1=%s, l2=%s]" %
(trial.number, acc, params["lambda_l1"], params["lambda_l2"]))
return score
def train(f_train,
f_test,
d_tmp="optuna-tmp",
phases="l:b:m:r",
iters=100,
timeout=None,
init_params=None,
usebar=True,
min_leaves=256,
max_leaves=32768):
(xs, ys) = trains.load(f_train)
dtrain = lgb.Dataset(xs, label=ys)
testd = trains.load(f_test) if f_test != f_train else (xs, ys)
os.system('mkdir -p "%s"' % d_tmp)
redirect.module("optuna", os.path.join(d_tmp, "optuna.log"))
params = dict(lgbooster.DEFAULTS)
if init_params: params.update(init_params)
pos = sum(ys)
neg = len(ys) - pos
#params["scale_pos_weight"] = neg / pos
params["is_unbalance"] = "true" if neg != pos else "false"
phases = phases.split(":")
if "m" in phases:
params["feature_pre_filter"] = "false"
timeout = timeout / len(phases) if timeout else None
iters = iters // len(phases) if iters else None
args = dict(dtrain=dtrain,
testd=testd,
d_tmp=d_tmp,
iters=iters,
timeout=timeout,
usebar=usebar,
min_leaves=min_leaves,
max_leaves=max_leaves)
if init_params is not None:
f_mod = os.path.join(d_tmp, "init.lgb")
(score, acc, dur) = model(params, dtrain, testd, f_mod,
"[init]" if usebar else None)
best = (score, acc, f_mod, dur)
logger.debug("- initial model: %s" % human.humanacc(acc))
else:
best = (-1, None, None)
for phase in phases:
trial = PHASES[phase](params=params, **args)
if trial.user_attrs["score"] > best[0]:
best = tuple(trial.user_attrs[x]
for x in ["score", "acc", "model", "time"])
params.update(trial.params)
#if "num_leaves_base" in params:
# params["num_leaves"] = round(2**(params["num_leaves_base"]/2))
# del params["num_leaves_base"]
return best + (params, pos, neg)
def check_leaves(trial, params, min_leaves, max_leaves, **args):
#num_leaves_base = trial.suggest_int('num_leaves_base', 16, 31)
#num_leaves = round(2**(num_leaves_base/2))
num_leaves = trial.suggest_int('num_leaves', min_leaves, max_leaves)
params = dict(params, num_leaves=num_leaves)
score = check(trial, params, **args)
acc = human.humanacc(trial.user_attrs["acc"])
logger.debug("- leaves trial %d: %s [num_leaves=%s]" %
(trial.number, acc, params["num_leaves"]))
return score
def check_bagging(trial, params, **args):
bagging_freq = trial.suggest_int("bagging_freq", 1, 7)
bagging_fraction = min(
trial.suggest_float("bagging_fraction", 0.4, 1.0 + 1e-12), 1.0)
params = dict(params,
bagging_freq=bagging_freq,
bagging_fraction=bagging_fraction)
score = check(trial, params, **args)
acc = human.humanacc(trial.user_attrs["acc"])
logger.debug("- bagging trial %d: %s [freq=%s, frac=%s]" %
(trial.number, acc, params["bagging_freq"],
params["bagging_fraction"]))
return score
def lgbtune(f_train,
f_test,
d_tmp="optuna-tmp",
phases="l:b:m:r",
iters=None,
timeout=3600,
init_params={},
min_leaves=256,
max_leaves=32768):
(_, acc, f_mod, _, params, _, _) = train(f_train, f_test, d_tmp, phases,
iters, timeout, init_params, True,
min_leaves, max_leaves)
logger.info("")
logger.info("Best model params: %s" % str(params))
logger.info("Best model accuracy: %s" % human.humanacc(acc))
logger.info("Best model file: %s" % f_mod)