将字符串(类别)的数组从pandas数据帧转换为int数组

时间:2011-10-18 20:16:13

标签: python numpy pandas

我正在尝试做与previous question非常相似的事情,但我遇到了错误。 我有一个包含功能和标签的pandas数据框我需要进行一些转换以将功能和标签变量发送到机器学习对象中:

import pandas
import milk
from scikits.statsmodels.tools import categorical

然后我有:

trainedData=bigdata[bigdata['meta']<15]
untrained=bigdata[bigdata['meta']>=15]
#print trainedData
#extract two columns from trainedData
#convert to numpy array
features=trainedData.ix[:,['ratio','area']].as_matrix(['ratio','area'])
un_features=untrained.ix[:,['ratio','area']].as_matrix(['ratio','area'])
print 'features'
print features[:5]
##label is a string:single, touching,nuclei,dust
print 'labels'

labels=trainedData.ix[:,['type']].as_matrix(['type'])
print labels[:5]
#convert single to 0, touching to 1, nuclei to 2, dusts to 3
#
tmp=categorical(labels,drop=True)
targets=categorical(labels,drop=True).argmax(1)
print targets

输出控制台首先产生:

features
[[ 0.38846334  0.97681855]
[ 3.8318634   0.5724734 ]
[ 0.67710876  1.01816444]
[ 1.12024943  0.91508699]
[ 7.51749674  1.00156707]]
labels
[[single]
[touching]
[single]
[single]
[nuclei]]

我遇到了以下错误:

Traceback (most recent call last):
File "/home/claire/Applications/ProjetPython/projet particule et objet/karyotyper/DAPI-Trainer02-MILK.py", line 83, in <module>
tmp=categorical(labels,drop=True)
File "/usr/local/lib/python2.6/dist-packages/scikits.statsmodels-0.3.0rc1-py2.6.egg/scikits/statsmodels/tools/tools.py", line 206, in categorical
tmp_dummy = (tmp_arr[:,None]==data).astype(float)
AttributeError: 'bool' object has no attribute 'astype'

是否可以将数据框中的类别变量'type'转换为int? 'type'可以取值'single','touching','nuclei','dusts',我需要转换为int值0,1,2,3。

4 个答案:

答案 0 :(得分:18)

之前的答案已经过时,所以这里有一个解决方案,用于将字符串映射到适用于Pandas版本0.18.1的数字。

对于系列赛:

In [1]: import pandas as pd
In [2]: s = pd.Series(['single', 'touching', 'nuclei', 'dusts',
                       'touching', 'single', 'nuclei'])
In [3]: s_enc = pd.factorize(s)
In [4]: s_enc[0]
Out[4]: array([0, 1, 2, 3, 1, 0, 2])
In [5]: s_enc[1]
Out[5]: Index([u'single', u'touching', u'nuclei', u'dusts'], dtype='object')

对于DataFrame:

In [1]: import pandas as pd
In [2]: df = pd.DataFrame({'labels': ['single', 'touching', 'nuclei', 
                       'dusts', 'touching', 'single', 'nuclei']})
In [3]: catenc = pd.factorize(df['labels'])
In [4]: catenc
Out[4]: (array([0, 1, 2, 3, 1, 0, 2]), 
        Index([u'single', u'touching', u'nuclei', u'dusts'],
        dtype='object'))
In [5]: df['labels_enc'] = catenc[0]
In [6]: df
Out[4]:
         labels  labels_enc
    0    single           0
    1  touching           1
    2    nuclei           2
    3     dusts           3
    4  touching           1
    5    single           0
    6    nuclei           2

答案 1 :(得分:11)

如果你有一个字符串或其他对象的向量,并且你想给它分类标签,你可以使用Factor类(在pandas命名空间中提供):

In [1]: s = Series(['single', 'touching', 'nuclei', 'dusts', 'touching', 'single', 'nuclei'])

In [2]: s
Out[2]: 
0    single
1    touching
2    nuclei
3    dusts
4    touching
5    single
6    nuclei
Name: None, Length: 7

In [4]: Factor(s)
Out[4]: 
Factor:
array([single, touching, nuclei, dusts, touching, single, nuclei], dtype=object)
Levels (4): [dusts nuclei single touching]

该系数具有属性labelslevels

In [7]: f = Factor(s)

In [8]: f.labels
Out[8]: array([2, 3, 1, 0, 3, 2, 1], dtype=int32)

In [9]: f.levels
Out[9]: Index([dusts, nuclei, single, touching], dtype=object)

这适用于1D矢量,因此不确定它是否可以立即应用于您的问题,但请查看。

BTW我建议你在statsmodels和/或scikit-learn邮件列表上提出这些问题,因为我们大多数人都不是SO用户。

答案 2 :(得分:6)

我正在回答Pandas 0.10.1的问题。 Factor.from_array似乎可以解决问题。

>>> s = pandas.Series(['a', 'b', 'a', 'c', 'a', 'b', 'a'])
>>> s
0    a
1    b
2    a
3    c
4    a
5    b
6    a
>>> f = pandas.Factor.from_array(s)
>>> f
Categorical: 
array([a, b, a, c, a, b, a], dtype=object)
Levels (3): Index([a, b, c], dtype=object)
>>> f.labels
array([0, 1, 0, 2, 0, 1, 0])
>>> f.levels
Index([a, b, c], dtype=object)

答案 3 :(得分:0)

因为这些都不适用于尺寸&gt; 1,我使一些代码适用于任何numpy数组维度:

def encode_categorical(array):
    d = {key: value for (key, value) in zip(np.unique(array), np.arange(len(u)))}
    shape = array.shape
    array = array.ravel()
    new_array = np.zeros(array.shape, dtype=np.int)
    for i in range(len(array)):
        new_array[i] = d[array[i]]
    return new_array.reshape(shape)