class ModelRepresentationBase(_AbstractModelRepresentation):
""" class just to store the default HyperParameters """
default_hyper = {
"n_components":
hp.HyperRangeFloat(start=0.1, end=1, step=0.05),
# Forest like estimators
"n_estimators":
hp.HyperComposition([
(0.75, hp.HyperRangeInt(start=25, end=175, step=25)),
(0.25, hp.HyperRangeInt(start=200, end=1000, step=100)),
]),
"max_features":
hp.HyperComposition([(0.25, ["sqrt", "auto"]),
(0.75,
hp.HyperRangeBetaFloat(start=0,
end=1,
alpha=3,
beta=1))]),
"max_depth":
hp.HyperChoice([
None, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 20, 25,
30, 50, 100
]),
"min_samples_split":
hp.HyperRangeBetaInt(start=2, end=100, alpha=1, beta=5),
# Linear model
"C":
hp.HyperLogRangeFloat(start=0.00001, end=10, n=50),
"alpha":
hp.HyperLogRangeFloat(start=0.00001, end=10, n=50),
# CV
"analyzer":
hp.HyperChoice(["word", "char", "char_wb"]),
"penalty": ["l1", "l2"],
"random_state": [
123
], # So that every for every model with a random_state attribute, it will be passed and fix
"drop_used_columns": [True],
"drop_unused_columns": [True]
}
# This dictionnary is used to specify the default hyper-parameters that are used during the random search phase
# They will be used if :
# * the model has a paramters among that list
# * the parameters is not specified within the class (within 'custom_hyper')
default_default_hyper = {
"random_state": 123,
"drop_used_columns": True,
"drop_unused_columns": True
}
# This dictionnary is used to specify the default hyper-parameters that are used during the default model phase
# They will be used if :
# * the model has a paramters among that list
# * the default parameters is not specified within the class (withing 'default_parameters')
depends_on = ()
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.HyperRangeInt(start=1, end=4),
"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
class Char2VecVectorizer_TextEncoder(ModelRepresentationBase):
klass = Char2VecVectorizer
category = StepCategories.TextEncoder
type_of_variable = TypeOfVariables.TEXT
custom_hyper = {
"size": hp.HyperRangeInt(50, 300, step=10),
"window": [3, 5, 7],
"ngram": hp.HyperRangeInt(2, 6),
"same_embedding_all_columns": [True, False],
"text_preprocess": [None, "default", "digit", "nltk"],
}
type_of_model = None
use_y = False
class Text_TruncatedSVD_DimensionReduction(ModelRepresentationBase):
klass = TruncatedSVDWrapper
category = StepCategories.TextDimensionReduction
custom_hyper = {"n_components": hp.HyperRangeInt(10, 500, step=5)}
type_of_variable = TypeOfVariables.TEXT
type_of_model = None
use_y = False
custom_hyper = {"drop_used_columns": [True, False]}
class ModelRepresentationBase(_AbstractModelRepresentation):
""" class just to store the default HyperParameters """
default_hyper = {
"n_components":
hp.HyperRangeFloat(start=0.1, end=1, step=0.05),
# Forest like estimators
"n_estimators":
hp.HyperComposition([
(0.75, hp.HyperRangeInt(start=25, end=175, step=25)),
(0.25, hp.HyperRangeInt(start=200, end=1000, step=100)),
]),
"max_features":
hp.HyperComposition([(0.25, ["sqrt", "auto"]),
(0.75,
hp.HyperRangeBetaFloat(start=0,
end=1,
alpha=3,
beta=1))]),
"max_depth":
hp.HyperChoice([
None, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 20, 25,
30, 50, 100
]),
"min_samples_split":
hp.HyperRangeBetaInt(start=1, end=100, alpha=1, beta=5),
# Linear model
"C":
hp.HyperLogRangeFloat(start=0.00001, end=10, n=50),
"alpha":
hp.HyperLogRangeFloat(start=0.00001, end=10, n=50),
# CV
"analyzer":
hp.HyperChoice(["word", "char", "char_wb"]),
"penalty": ["l1", "l2"],
"random_state": [
123
], # So that every for every model with a random_state attribute, it will be passed and fix
"columns_to_encode": ["--object--"]
}
class NumericalEncoder_CatEncoder(ModelRepresentationBase):
klass = NumericalEncoder
category = StepCategories.CategoryEncoder
type_of_variable = (TypeOfVariables.CAT, TypeOfVariables.NUM)
custom_hyper = {
"encoding_type": ["dummy", "num"],
"min_nb_observations": hp.HyperRangeInt(2, 20)
}
type_of_model = None
use_y = False
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