def get_hyper_parameter(cls):
### Specific function to handle the fact that I don't want ngram != 1 IF analyzer = word ###
res = hp.HyperComposition([
(
0.5,
hp.HyperCrossProduct({
"ngram_range": 1,
"analyzer": "word",
"min_df": [1, 0.001, 0.01, 0.05],
"max_df": [0.999, 0.99, 0.95],
"tfidf": [True, False],
}),
),
(
0.5,
hp.HyperCrossProduct({
"ngram_range":
hp.HyperRangeBetaInt(
start=1, end=5, alpha=2, beta=1
), # 1 = 1.5% ; 2 = 12% ; 3 = 25% ; 4 = 37% ; 5 = 24%
"analyzer":
hp.HyperChoice(("char", "char_wb")),
"min_df": [1, 0.001, 0.01, 0.05],
"max_df": [0.999, 0.99, 0.95],
"tfidf": [True, False],
}),
),
])
return res
def get_hyper_parameter(cls):
if cls.klass is None:
raise ValueError("I need a klass")
if cls.hyper is not None:
return cls.hyper
all_params = list(get_init_parameters(cls.klass).keys())
all_hyper = {}
for p in all_params:
if p in cls.custom_hyper:
all_hyper[p] = cls.custom_hyper[p]
elif p in cls.default_hyper:
all_hyper[p] = cls.default_hyper[p]
return hp.HyperCrossProduct(all_hyper) # TODO : fix seed here
def get_hyper_parameter(cls):
""" specific function to handle dependency between hyper-parameters : bagging_fraction AND bagging_freq """
res = hp.HyperComposition([
##################
### No Bagging ###
##################
# * bagging_freq == 0
# * bagging_fraction == 1.0
# * no random forest here : 'booting_type' != 'rf'
(
0.5,
hp.HyperCrossProduct({
"boosting_type": ["gbdt", "dart"],
"learning_rate":
hp.HyperLogRangeFloat(0.0001, 0.1),
"max_depth":
hp.HyperChoice([
-1, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
20, 25, 30, 50, 100
]),
"n_estimators":
hp.HyperComposition([
(0.50, hp.HyperRangeInt(start=25, end=175, step=25)),
(0.25, hp.HyperRangeInt(start=200, end=900, step=100)),
(0.25, hp.HyperRangeInt(start=1000,
end=10000,
step=100)),
]),
"colsample_bytree":
hp.HyperRangeBetaFloat(start=0.1, end=1, alpha=3,
beta=1), # Mean = 0.75
"min_child_samples":
hp.HyperRangeInt(2, 50),
"num_leaves":
hp.HyperRangeInt(10, 200),
"bagging_fraction": [1.0],
"bagging_freq": [0],
"n_jobs": [1],
}),
),
###############
### Bagging ###
###############
# * bagging_freq = 1
# * bagging_fraction < 1
(
0.5,
hp.HyperCrossProduct({
"boosting_type": ["rf", "gbdt", "dart"],
"learning_rate":
hp.HyperLogRangeFloat(0.0001, 0.1),
"max_depth":
hp.HyperChoice([
-1, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
20, 25, 30, 50, 100
]),
"n_estimators":
hp.HyperComposition([
(0.50, hp.HyperRangeInt(start=25, end=175, step=25)),
(0.25, hp.HyperRangeInt(start=200, end=900, step=100)),
(0.25, hp.HyperRangeInt(start=1000,
end=10000,
step=100)),
]),
"colsample_bytree":
hp.HyperRangeBetaFloat(start=0.1, end=1, alpha=3,
beta=1), # Mean = 0.75
"min_child_samples":
hp.HyperRangeInt(2, 50),
"num_leaves":
hp.HyperRangeInt(10, 200),
"bagging_fraction":
hp.HyperRangeBetaFloat(start=0.1, end=1, alpha=3, beta=1),
"bagging_freq": [1],
"n_jobs": [1],
}),
),
])
return res