# -*- encoding:utf-8 -*-
"""封装grid search相关操作模块"""
from __future__ import division
from __future__ import print_function
from __future__ import absolute_import
import logging
import matplotlib.pyplot as plt
import numpy as np
from sklearn.base import ClassifierMixin, RegressorMixin
from sklearn.metrics import explained_variance_score, make_scorer
from ..CoreBu.ABuFixes import signature, six
from ..CoreBu.ABuFixes import GridSearchCV
__author__ = '阿布'
__weixin__ = 'abu_quant'
__all__ = [
'grid_search_init_kwargs',
'grid_search_mul_init_kwargs',
'grid_search_init_n_estimators',
'grid_search_init_n_components',
'grid_search_init_max_depth',
'grid_search_init_n_neighbors'
]
def _scoring_grid(estimator, scoring):
"""
只针对有监督学习过滤无监督学习，对scoring未赋予的情况根据
学习器分类器使用accuracy进行度量，回归器使用可释方差值explained_variance_score，
使用make_scorer对函数进行score封装
:param estimator: 学习器对象
:param scoring: 度量使用的方法，未赋予的情况根据
学习器分类器使用accuracy进行度量，回归器使用explained_variance_score进行度量
:return: scoring
"""
if not isinstance(estimator, (ClassifierMixin, RegressorMixin)):
logging.info('only support supervised learning')
# TODO 无监督学习的scoring度量以及GridSearchCV
return None
if scoring is None:
if isinstance(estimator, ClassifierMixin):
# 分类器使用accuracy
return 'accuracy'
elif isinstance(estimator, RegressorMixin):
# 回归器使用可释方差值explained_variance_score，使用make_scorer对函数进行score封装
"""
make_scorer中通过greater_is_better对返回值进行正负分配
eg: sign = 1 if greater_is_better else -1
"""
return make_scorer(explained_variance_score, greater_is_better=True)
return None
return scoring
def grid_search_init_kwargs(estimator, x, y, param_name, param_range, cv=10, n_jobs=-1, scoring=None, show=True):
"""
对GridSearchCV进行封装，对单个目标关键字参数进行grid search最优参数搜寻
eg：'n_estimators'， 'max_depth'
eg：param_range=np.arange(100, 500, 50))对最优参数进行寻找
eg:
from abupy import AbuML, ml
ttn_abu = AbuML.create_test_more_fiter()
ttn_abu.estimator.random_forest_classifier()
ml.grid_search_init_kwargs(ttn_abu.estimator.clf, ttn_abu.x, ttn_abu.y,
param_name='n_estimators', param_range=np.arange(100, 500, 50))
可找到n_estimators参数最优为：(0.81930415263748602, {'n_estimators': 300})
:param estimator: 学习器对象
:param x: 训练集x矩阵，numpy矩阵
:param y: 训练集y序列，numpy序列
:param param_name: 做为grid的目标关键字参数，eg：'n_estimators'， 'max_depth'
:param param_range: 做为grid的目标关键字参数的grid序列，eg：param_range=np.arange(100, 500, 50))
:param cv: int，GridSearchCV切割训练集测试集参数，默认10
:param n_jobs: 并行执行的进程任务数量，默认-1, 开启与cpu相同数量的进程数
:param scoring: 测试集的度量方法，默认为None, None的情况下分类器使用accuracy进行度量，
回归器使用可释方差值explained_variance_score，使用make_scorer对函数进行score封装
:param show: 是否进行可视化
:return: eg：(0.81930415263748602, {'n_estimators': 300})
"""
if not isinstance(param_name, six.string_types):
# param_name参数需要是字符串类型
logging.info('param_name is str, not {}, eg: \'n_estimators\''.format(param_name))
return None, None
# 根据分类回归得到非None的score
scoring = _scoring_grid(estimator, scoring)
if scoring is None:
# 如果_scoring_grid返回的结果仍然是None, 说明无监督学习，暂时不支持
return None, None
# 获取学习器的init函数，使用getattr
init = getattr(estimator.__class__.__init__, 'deprecated_original', estimator.__class__.__init__)
# 获取函数签名
init_signature = signature(init)
"""
eg：init_signature
['self', 'base_estimator', 'n_estimators', 'max_samples', 'max_features', 'bootstrap',
'bootstrap_features', 'oob_score', 'warm_start', 'n_jobs', 'random_state', 'verbose']
"""
if param_name not in init_signature.parameters.keys():
# 如果需要grid的参数param_name不在init函数签名中，打log，返回
logging.info('check init signature {} not in **kwargs\ninit_signature:{}'.format(
param_name, init_signature.parameters.keys()))
return None, None
param_grid = {param_name: param_range}
grid = GridSearchCV(estimator, param_grid, cv=cv, scoring=scoring, n_jobs=n_jobs)
print('start grid search please wait...')
grid.fit(x, y)
if show:
if hasattr(grid, 'cv_results_'):
# 0.18之后的版本有cv_results_，拿出每一次的训练score的mean形成grid_scores
cv_results = grid.cv_results_
grid_scores = cv_results['mean_test_score']
else:
# 0.18之前的版本
cv_results = grid.grid_scores_
grid_scores = [result.mean_validation_score for result in cv_results]
"""
cv_results中每一个元素为_CVScoreTuple namedtuple对象，如下所示：
class _CVScoreTuple (namedtuple('_CVScoreTuple',
('parameters',
'mean_validation_score',
'cv_validation_scores'))):
"""
# FIXME 这里假定了所有param_range的元素类型都是数值类型，需要判定，并且根据情况是否需要排序
plt.plot(param_range, grid_scores)
# 把最好的红圈标记出来
plt.plot(grid.best_params_[param_name], grid.best_score_, 'ro', markersize=12, markeredgewidth=1.5,
markerfacecolor='None', markeredgecolor='r')
plt.title('search {}, best is {}'.format(param_name, grid.best_params_[param_name]))
plt.show()
return grid.best_score_, grid.best_params_
def grid_search_mul_init_kwargs(estimator, x, y, param_grid, cv=10, n_jobs=-1, scoring=None, show=True):
"""
对GridSearchCV进行封装，对多个目标关键字参数进行grid search最优参数搜寻
eg:
from abupy import AbuML, ml
ttn_abu = AbuML.create_test_more_fiter()
ttn_abu.estimator.random_forest_classifier()
param_grid = {'max_depth': np.arange(2, 5), 'n_estimators': np.arange(100, 300, 50)}
ml.grid_search_mul_init_kwargs(ttn_abu.estimator.clf, ttn_abu.x, ttn_abu.y, param_grid=param_grid)
out: (0.81593714927048255, {'max_depth': 4, 'n_estimators': 250})
:param estimator: 学习器对象
:param x: 训练集x矩阵，numpy矩阵
:param y: 训练集y序列，numpy序列
:param param_grid: eg：param_grid = {'max_depth': np.arange(2, 5), 'n_estimators': np.arange(100, 300, 50)}
:param cv: int，GridSearchCV切割训练集测试集参数，默认10
:param n_jobs: 并行执行的进程任务数量，默认-1, 开启与cpu相同数量的进程数
:param scoring: 测试集的度量方法，默认为None, None的情况下分类器使用accuracy进行度量，
回归器使用可释方差值explained_variance_score，使用make_scorer对函数进行score封装
:param show: 是否进行可视化
:return: eg: (0.81593714927048255, {'max_depth': 4, 'n_estimators': 250})
"""
if not isinstance(param_grid, dict):
# param_grid参数是dict对象
logging.info('param_grid is dict object, not {}}'.format(param_grid))
return None, None
# 根据分类回归得到非None的score
scoring = _scoring_grid(estimator, scoring)
if scoring is None:
# 如果_scoring_grid返回的结果仍然是None, 说明无监督学习，暂时不支持
return None, None
for param_name in param_grid.keys():
# 迭代每一个key，即每一个关键字参数，看init方法中的签名是否存在该关键字
# 获取学习器的init函数，使用getattr
init = getattr(estimator.__class__.__init__, 'deprecated_original', estimator.__class__.__init__)
# 获取函数签名
init_signature = signature(init)
if param_name not in init_signature.parameters.keys():
# 如果需要grid的参数param_name不在init函数签名中，打log，返回
logging.info('check init signature {} not in **kwargs\ninit_signature:{}'.format(
param_name, init_signature.parameters.keys()))
return None, None
grid = GridSearchCV(estimator, param_grid, cv=cv, scoring=scoring, n_jobs=n_jobs)
print('start grid search please wait...')
grid.fit(x, y)
if show:
if hasattr(grid, 'cv_results_'):
# 0.18之后的版本有cv_results_，拿出每一次的训练score的mean形成grid_scores
cv_results = grid.cv_results_
grid_scores = cv_results['mean_test_score']
"""
eg：grid_scores
[0.77890011223344557, 0.79349046015712688, 0.77553310886644222, 0.77441077441077444,
0.80920314253647585, 0.80920314253647585, 0.80808080808080807, 0.81032547699214363,
0.80695847362514028, 0.81144781144781142, 0.80471380471380471, 0.81593714927048255]
"""
grid_params = cv_results['params']
"""
eg：grid_params
[{'max_depth': 2, 'n_estimators': 100}, {'max_depth': 2, 'n_estimators': 150},
{'max_depth': 2, 'n_estimators': 200}, {'max_depth': 2, 'n_estimators': 250},
{'max_depth': 3, 'n_estimators': 100}, {'max_depth': 3, 'n_estimators': 150},
{'max_depth': 3, 'n_estimators': 200}, {'max_depth': 3, 'n_estimators': 250},
{'max_depth': 4, 'n_estimators': 100}, {'max_depth': 4, 'n_estimators': 150},
{'max_depth': 4, 'n_estimators': 200}, {'max_depth': 4, 'n_estimators': 250}]
"""
else:
cv_results = grid.grid_scores_
"""
cv_results中每一个元素为_CVScoreTuple namedtuple对象，如下所示：
class _CVScoreTuple (namedtuple('_CVScoreTuple',
('parameters',
'mean_validation_score',
'cv_validation_scores'))):
"""
grid_scores = [result.mean_validation_score for result in cv_results]
grid_params = [result.parameters for result in cv_results]
# 与grid_search_init_kwargs不同可视化grid_scores绘制曲线y，x只使用index
plt.plot(grid_scores)
cmap = plt.get_cmap('jet', len(grid_scores))
cmap.set_under('gray')
for grid_index in np.arange(0, len(grid_scores)):
# 迭代每一个分数，绘制点在曲线上根据分数用颜色区分，使用label进行标注
plt.scatter(grid_index, grid_scores[grid_index], s=50, cmap=cmap,
vmin=np.min(grid_scores),
vmax=np.max(grid_scores),
label='{}: {:.2f}'.format(grid_params[grid_index], grid_scores[grid_index]))
plt.title('best params is {}'.format(grid.best_params_))
# 将label标注文字绘制在外面
plt.legend(bbox_to_anchor=(1.05, 1), loc=2, borderaxespad=0.)
plt.show()
return grid.best_score_, grid.best_params_
def grid_search_init_n_estimators(estimator, x, y, n_estimators_range=None, cv=10, n_jobs=-1,
scoring=None, show=True):
"""
封装grid search特定的'n_estimators'关键字参数最优搜索，
为AbuMLCreater中_estimators_prarms_best提供callback函数，
具体阅读
AbuMLCreater._estimators_prarms_best()
+ AbuMLCreater.random_forest_classifier_best()
eg:
from abupy import AbuML, ml
ttn_abu = AbuML.create_test_more_fiter()
ttn_abu.estimator.random_forest_classifier()
ml.grid_search_init_n_estimators(ttn_abu.estimator.clf, ttn_abu.x, ttn_abu.y)
:param estimator: 学习器对象
:param x: 训练集x矩阵，numpy矩阵
:param y: 训练集y序列，numpy序列
:param n_estimators_range: 默认None, None则会使用 n_estimators_range = np.arange(50, 500, 10)
:param cv: int，GridSearchCV切割训练集测试集参数，默认10
:param n_jobs: 并行执行的进程任务数量，默认-1, 开启与cpu相同数量的进程数
:param scoring: 测试集的度量方法，默认为None, None的情况下分类器使用accuracy进行度量，
回归器使用可释方差值explained_variance_score，使用make_scorer对函数进行score封装
:param show: 是否进行可视化
:return: eg: (0.82154882154882158, {'n_estimators': 310})
"""
if n_estimators_range is None:
n_estimators_range = np.arange(50, 500, 10)
return grid_search_init_kwargs(estimator, x, y, 'n_estimators', n_estimators_range,
cv=cv, n_jobs=n_jobs, scoring=scoring, show=show)
def grid_search_init_max_depth(estimator, x, y, max_depth_range=None, cv=10, n_jobs=-1,
scoring=None, show=True):
"""
封装grid search特定的'n_components'关键字参数最优搜索，
为AbuMLCreater中_estimators_prarms_best提供callback函数
具体阅读
AbuMLCreater._estimators_prarms_best()
+ AbuMLCreater.decision_tree_classifier_best()
:param estimator: 学习器对象
:param x: 训练集x矩阵，numpy矩阵
:param y: 训练集y序列，numpy序列
:param max_depth_range: 默认None, None则会使用:
max_depth_range = np.arange(2, np.maximum(10, int(x.shape[1]) - 1), 1)
:param cv: int，GridSearchCV切割训练集测试集参数，默认10
:param n_jobs: 并行执行的进程任务数量，默认-1, 开启与cpu相同数量的进程数
:param scoring: 测试集的度量方法，默认为None, None的情况下分类器使用accuracy进行度量，
回归器使用可释方差值explained_variance_score，使用make_scorer对函数进行score封装
:param show: 是否进行可视化
:return: eg: (0.82154882154882158, {'max_depth': 3})
"""
if max_depth_range is None:
max_depth_range = np.arange(2, np.maximum(10, int(x.shape[1]) - 1), 1)
return grid_search_init_kwargs(estimator, x, y, 'max_depth', max_depth_range,
cv=cv, n_jobs=n_jobs, scoring=scoring, show=show)
def grid_search_init_n_neighbors(estimator, x, y, n_neighbors_range=None, cv=10, n_jobs=-1,
scoring=None, show=True):
"""
封装grid search特定的'n_components'关键字参数最优搜索，
为AbuMLCreater中_estimators_prarms_best提供callback函数
具体阅读
AbuMLCreater._estimators_prarms_best()
+ AbuMLCreater.knn_classifier_best()
:param estimator: 学习器对象
:param x: 训练集x矩阵，numpy矩阵
:param y: 训练集y序列，numpy序列
:param n_neighbors_range: 默认None, None则会使用:
n_estimators_range = np.arange(2, np.maximum(10, int(x.shape[1]) - 1), 1)
:param cv: int，GridSearchCV切割训练集测试集参数，默认10
:param n_jobs: 并行执行的进程任务数量，默认-1, 开启与cpu相同数量的进程数
:param scoring: 测试集的度量方法，默认为None, None的情况下分类器使用accuracy进行度量，
回归器使用可释方差值explained_variance_score，使用make_scorer对函数进行score封装
:param show: 是否进行可视化
:return: eg: (0.82154882154882158, {'n_components': 10})
"""
if n_neighbors_range is None:
# 邻居投票者控制在1-np.minimum(26, 总数的1/3）
n_neighbors_range = np.arange(1, np.minimum(26, int(x.shape[0] / 3)), 1)
return grid_search_init_kwargs(estimator, x, y, 'n_neighbors', n_neighbors_range,
cv=cv, n_jobs=n_jobs, scoring=scoring, show=show)
def grid_search_init_n_components(estimator, x, y, n_components_range=None, cv=10, n_jobs=-1,
scoring=None, show=True):
"""
封装grid search特定的'n_components'关键字参数最优搜索，
为AbuMLCreater中_estimators_prarms_best提供callback函数，
具体阅读AbuMLCreater._estimators_prarms_best()
:param estimator: 学习器对象
:param x: 训练集x矩阵，numpy矩阵
:param y: 训练集y序列，numpy序列
:param n_components_range: 默认None, None则会使用:
n_estimators_range = np.arange(2, np.maximum(10, int(x.shape[1]) - 1), 1)
:param cv: int，GridSearchCV切割训练集测试集参数，默认10
:param n_jobs: 并行执行的进程任务数量，默认-1, 开启与cpu相同数量的进程数
:param scoring: 测试集的度量方法，默认为None, None的情况下分类器使用accuracy进行度量，回归器使用
回归器使用可释方差值explained_variance_score，使用make_scorer对函数进行score封装
:param show: 是否进行可视化
:return: eg: (0.82154882154882158, {'n_components': 10})
"""
if n_components_range is None:
n_components_range = np.arange(2, np.maximum(10, int(x.shape[1]) - 1), 1)
return grid_search_init_kwargs(estimator, x, y, 'n_components', n_components_range,
cv=cv, n_jobs=n_jobs, scoring=scoring, show=show)
