1. 保存变量

1.1 pickle(泡菜)

pickle库可以指定多个变量保存在.pickle文件中，如果需要保存的变量不是很多，这种方法可以采用。

import pickle

# obj0, obj1, obj2 are created here...
obj0, obj1, obj2 = [1, 2], [2, 3], [3, 4]
# Saving the objects:
# 通过传递protocol = -1到dump()来减少文件大小
with open('test.pickle', 'wb') as f:  # Python 3: open(..., 'wb')
    pickle.dump([obj0, obj1, obj2], f)
f.close()

# Getting back the objects:
with open('test.pickle', 'rb') as f:  # Python 3: open(..., 'rb')
    x0, x1, x2 = pickle.load(f)
print(x0)
f.close()

[1, 2]

1.2 cPickle

cPickle的速度更快，其余和cPickle基无差别。在python3里面，cPickle变成_pickle。具体如下：

Docstring:   Optimized C implementation for the Python pickle module.

import  _pickle as cpickle

# obj0, obj1, obj2 are created here...
obj0, obj1, obj2 = [1, 2], [2, 3], [3, 4]
# Saving the objects:
# 通过传递protocol = -1到dump()来减少文件大小
with open('test.pickle', 'wb') as f:  # Python 3: open(..., 'wb')
    cpickle.dump([obj0, obj1, obj2], f)
f.close()

del x0, x1, x2 
# Getting back the objects:
with open('test.pickle', 'rb') as f:  # Python 3: open(..., 'rb')
    x0, x1, x2 = cpickle.load(f)
print(x0)
f.close()

[1, 2]

1.3 shelve

似乎不支持内建函数等其他的对象，也不是很智能嘛～这样到不如直接用pickle。参考：http://www.php.cn/python-tutorials-410803.html

import shelve

T='Hiya'
val=[1,2,3]

filename='test'
my_shelf = shelve.open(filename,'n') # 'n' for new

for key in dir():
    try:
        my_shelf[key] = globals()[key]
    except:
        #
        # __builtins__, my_shelf, and imported modules can not be shelved.
        #
        print('ERROR shelving: {0}'.format(key))
my_shelf.close()

del val,T

my_shelf = shelve.open(filename)
for key in my_shelf:
    globals()[key]=my_shelf[key]
my_shelf.close()

print(T)
# Hiya
print(val)
# [1, 2, 3]

1.4 dill

pycharm里可以保存文件，but在jupyter里报错，不知道是个啥原因啊～

dump_session(filename='/tmp/session.pkl', main=None, byref=False)
    pickle the current state of __main__ to a file

import dill
# 保存文件
filename = 'globalsave.pkl'
dill.dump_session(filename)

dill.load_session(filename)

2. 保存模型文件

2.1 .model文件

2.1.1 训练模型

import pandas as pd
from sklearn.tree import DecisionTreeClassifier
from sklearn import cross_validation, metrics
from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split

# -- 载入鸢尾花数据集
iris_dataset = load_iris()
# -- 数据处理&训练集、测试集切分
rawdata = pd.DataFrame(iris_dataset['data'], columns=['x0', 'x1', 'x2', 'x3'])
rawlabel = pd.DataFrame(iris_dataset['target'], columns=['label'])
dt_model = DecisionTreeClassifier()

train_X, test_X, train_y, test_y = train_test_split(rawdata,
                                                    rawlabel, test_size=0.3, random_state=0)
dt_model.fit(X=train_X, y=train_y)


print(metrics.classification_report(train_y,
                                    dt_model.predict(X=train_X)))

print(metrics.classification_report(test_y,
                                    dt_model.predict(X=test_X)))

2.1.2 保存&调用模型文件。

from sklearn.externals import joblib

# 模型保存
joblib.dump(dt_model, './Code/dt_model.model')
# 模型载入
dt_model_load = joblib.load('./Code/dt_model.model')

print(metrics.classification_report(test_y,
                                    dt_model_load.predict(X=test_X)))

2.2 pickle文件

发现pickle也可以用，不知道是否性能方面存在一定的问题～

import pickle
with open('dt_model.pickle', 'wb') as f:
    pickle.dump(dt_model, f)
f.close()

import pickle
with open('dt_model.pickle', 'rb') as f:  # Python 3: open(..., 'rb')
    x = pickle.load(f)

2.3 pmml文件

sklearn训练的模型可以保存为pmml文件，似乎可以用java直接调用～后面用到再说，其余不赘。

2018-09-29 于南京紫东创业园

Python-保存对象&模型文件