-
Notifications
You must be signed in to change notification settings - Fork 0
/
OptGridSearchCV.py
179 lines (172 loc) · 7.8 KB
/
OptGridSearchCV.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
'''
An optimized method for GridSearchCV, which iteratively performs grid search
and reduces the span of the parameters after each iteration. Made to make the
life of an engineer less boring.
'''
import numpy as np
from sklearn.model_selection import GridSearchCV
from sklearn.svm import LinearSVC
def optGridSearchCV(classifier, xtrain, ytrain, parameters, reduction_ratio=2,
iter_num=3, scoring='f1_macro', fold_num=5, first_rand=False,
n_jobs=1,verbose=1,only_rand=False, only_brute=False):
'''
The local optimum resides inside the parameters space, with bounds defined
by the min and max of each parameter, thus a recommended way to run this
function, if no prior knowledge exists, is to set the min and max of each
parameter to the corresponding min and max allowed bounds.
<classifier>: initialized classifier object
<xtrain>: features of samples, with shape (n_samples, n_features)
<ytrain>: labels of samples
<parameters>: dictionary of parameters, same with GridSearchCV <params>
type
<reduction_ratio>: the scale of relative reduction of the span of the
number parameters
<iter_num>: number of iterations to take place
<fold_num>: number of folds for CrossValidation
<first_rand> : True to perform random parameter picking (normally
distributed) firstly and then brute parameter picking (using linspace).
If false, the turn of each method changes
<only_rand> : True to perform only random picking
<only_brute> : True to perform only brute picking
'''
def print_params(parameters, preset=''):
'''
print parameters in pandas form, if allowed
'''
try:
from pandas import DataFrame
if isinstance(parameters, list):
params = DataFrame(parameters)
else:
try:
params = DataFrame.from_dict(parameters)
except ValueError:
params = DataFrame([parameters])
print(params)
except ImportError:
print(preset+str(parameters))
def reduce_list(params, best_params):
'''
Reduce parameters list of dictionaries to a parameters dictionary,
which correspots to the <best_params> found by <GridSearchCV>
'''
best_keys = set(best_params.keys())
for count, dic in enumerate(params):
if best_keys == set(dic.keys()):
return dic, count
raise Exception
def update_parameters(prev_parameters, best_parameters, num_of_samples,
rate=2, israndom=True):
'''
Each new parameter has the same number of values as previous one and
its values are inside the bounds set by the min and max values of the
old parameter. Furthermore, best value from the previous paramter
exists inside the new parameter.
<num_of_samples>: dictionary with keys from the best_parameters.
<prev_parameters>: previous parameters, which hold all tested values
<best_parameters>: parameters found to provide the best score (using
GridSearchCV)
<israndom>: whether to perform random or brute method
<rate>: rate of parameters span relative reduction
'''
rate = float(rate)
new_parameters = {}
for key in best_parameters:
if (not isinstance(best_parameters[key], str) and
not isinstance(best_parameters[key], bool) and
not best_parameters[key] is None):
if israndom:
center = best_parameters[key]
std = np.std(prev_parameters[key]) / float(rate)
pick = np.random.normal(loc=center, scale=std,
size=100 * num_of_samples[key])
pick = pick[(pick >=
np.min(prev_parameters[key]))*
(pick <= np.max(prev_parameters[key]))]
new_parameters[key] = pick[
:(num_of_samples[key]-1)]
else:
center = best_parameters[key]
rang = np.max(prev_parameters[
key]) - np.min(prev_parameters[key])
rang = [max(center - rang /
float(rate), min(prev_parameters[key])),
min(center + rang /
float(rate), max(prev_parameters[key]))]
new_parameters[key] = np.linspace(
rang[0], rang[1], num_of_samples[key]-1)
if isinstance(best_parameters[key], int):
new_parameters[key] = new_parameters[key].astype(int)
new_parameters[key] = new_parameters[key].tolist()
new_parameters[key] += [best_parameters[key]]
else:
new_parameters[key] = [best_parameters[key]]
return new_parameters
num_of_samples = {}
if not isinstance(parameters, list):
num_of_samples = {}
for key in parameters:
num_of_samples[key] = len(parameters[key])
best_scores = []
best_params = []
best_estimators = []
rand_flags = [first_rand, not first_rand]
if only_brute:
rand_flags = [False]
if only_rand:
rand_flags = [True]
for it_count in range(iter_num):
for rand_flag in rand_flags:
if verbose==2:
print('Parameters to test on:')
print_params(parameters,'\t')
try:
grids = GridSearchCV(
classifier,
parameters,
scoring=scoring,
cv=fold_num,
n_jobs=n_jobs, verbose=verbose)
grids.fit(xtrain, ytrain)
best_scores.append(grids.best_score_)
best_params.append(grids.best_params_)
best_estimators.append(grids.best_estimator_)
grids_params = grids.best_params_
except ValueError:
print('Invalid parameters')
raise
best_params = parameters
if rand_flag == rand_flags[1]:
print('Iteration Number: ' + str(it_count))
print('\tBest Classifier Params:')
print_params(best_params[-1],'\t\t')
print('\tBest Score:' + str(best_scores[-1]))
if isinstance(parameters, list):
parameters, _ = reduce_list(parameters, grids_params)
for key in parameters:
num_of_samples[key] = len(parameters[key])
if rand_flag == rand_flags[1] and it_count == iter_num - 1:
break
print('Reducing Parameters using '+ ['random' if rand_flag else
'brute'][0] + ' method')
parameters = update_parameters(parameters, grids_params, num_of_samples,
rate=reduction_ratio,
israndom=rand_flag)
return best_params, best_scores, best_estimators
def example():
'''
An example of usage
'''
parameters = [{'C': [1, 10, 100, 1000], 'tol': [0.001, 0.0001],
'class_weight': [None, 'balanced']},
{'C': [1, 10, 100, 1000], 'multi_class': ['crammer_singer'],
'tol': [0.001, 0.0001]}]
xtrain = np.random.random((100, 20))
xtrain[xtrain < 0] = 0
ytrain = (np.random.random(100) > 0.5).astype(int)
lsvc = LinearSVC()
optGridSearchCV(lsvc, xtrain, ytrain, parameters, reduction_ratio=2,
iter_num=3, scoring='f1_macro', fold_num=5, first_rand=False,
n_jobs=4)
if __name__ == '__main__':
example()