在给定精度下,美观地打印 NumPy 数组而无需使用科学计数法。

461

我该如何打印格式化的NumPy数组,使其类似于这样:

x = 1.23456
print('%.3f' % x)

如果我想打印 numpy.ndarray 类型的浮点数,它会输出多个小数位数,通常以“科学”格式呈现,即使对于低维数组也很难阅读。但是,numpy.ndarray 显然必须作为字符串打印,即使用 %s。有没有解决方法?


1
这个讨论也可能会引起那些通过谷歌搜索到这里的人的兴趣。 - Foad S. Farimani
14个回答

756

使用numpy.set_printoptions来设置输出的精度:

import numpy as np
x = np.random.random(10)
print(x)
# [ 0.07837821  0.48002108  0.41274116  0.82993414  0.77610352  0.1023732
#   0.51303098  0.4617183   0.33487207  0.71162095]

np.set_printoptions(precision=3)
print(x)
# [ 0.078  0.48   0.413  0.83   0.776  0.102  0.513  0.462  0.335  0.712]

suppress 则抑制了对于小数的科学计数法表示:

y = np.array([1.5e-10, 1.5, 1500])
print(y)
# [  1.500e-10   1.500e+00   1.500e+03]

np.set_printoptions(suppress=True)
print(y)
# [    0.      1.5  1500. ]

要在本地应用打印选项,使用NumPy 1.15.0或更高版本,可以使用numpy.printoptions上下文管理器。 例如,在with-suite内设置precision=3suppress=True:

x = np.random.random(10)
with np.printoptions(precision=3, suppress=True):
    print(x)
    # [ 0.073  0.461  0.689  0.754  0.624  0.901  0.049  0.582  0.557  0.348]

但是在with-suite之外,打印选项会恢复默认设置:

print(x)    
# [ 0.07334334  0.46132615  0.68935231  0.75379645  0.62424021  0.90115836
#   0.04879837  0.58207504  0.55694118  0.34768638]

如果您使用的是较早版本的NumPy,则可以自行创建上下文管理器。例如:

import numpy as np
import contextlib

@contextlib.contextmanager
def printoptions(*args, **kwargs):
    original = np.get_printoptions()
    np.set_printoptions(*args, **kwargs)
    try:
        yield
    finally: 
        np.set_printoptions(**original)

x = np.random.random(10)
with printoptions(precision=3, suppress=True):
    print(x)
    # [ 0.073  0.461  0.689  0.754  0.624  0.901  0.049  0.582  0.557  0.348]

为了防止浮点数末尾的零被去除:

np.set_printoptions现在有一个formatter参数,允许你为每种类型指定格式函数。

np.set_printoptions(formatter={'float': '{: 0.3f}'.format})
print(x)

打印的函数

[ 0.078  0.480  0.413  0.830  0.776  0.102  0.513  0.462  0.335  0.712]

代替

[ 0.078  0.48   0.413  0.83   0.776  0.102  0.513  0.462  0.335  0.712]

1
有没有一种方法可以仅将格式应用于特定的打印语句(而不是设置所有打印语句使用的通用输出格式)? - bph
11
@Hiett:NumPy没有单独设置一个print的打印选项的函数,但是你可以使用上下文管理器来实现类似功能。我已经编辑了上面的帖子来展示我的意思。 - unutbu
2
你的 np.set_printoptions(precision=3) 压制了末尾的零... 你怎么让它们显示成这样 [ 0.078 0.480 0.413 0.830 0.776 0.102 0.513 0.462 0.335 0.712] - Norfeldt
2
@Norfeldt:我已经在上面添加了一种方法来实现这个。 - unutbu
1
这个工作得很好。顺带一提,如果您想要字符串表示而不一定使用 print,您还可以使用 set_printoptions。您只需要调用 numpy 数组实例的 __str__() 方法,就会按照您设置的打印选项获取格式化字符串。 - Jayesh
显示剩余4条评论

73

使用np.array_str仅对单个打印语句应用格式设置。它提供了np.set_printoptions功能的子集。

例如:

In [27]: x = np.array([[1.1, 0.9, 1e-6]] * 3)

In [28]: print(x)
[[  1.10000000e+00   9.00000000e-01   1.00000000e-06]
 [  1.10000000e+00   9.00000000e-01   1.00000000e-06]
 [  1.10000000e+00   9.00000000e-01   1.00000000e-06]]

In [29]: print(np.array_str(x, precision=2))
[[  1.10e+00   9.00e-01   1.00e-06]
 [  1.10e+00   9.00e-01   1.00e-06]
 [  1.10e+00   9.00e-01   1.00e-06]]

In [30]: print(np.array_str(x, precision=2, suppress_small=True))
[[ 1.1  0.9  0. ]
 [ 1.1  0.9  0. ]
 [ 1.1  0.9  0. ]]

4
可能是最简单和高效的选项,因为它不会在printoptions中引入永久更改,也不需要昂贵的循环或with结构。应该将格式化元素的可能性直接集成到numpy中(无法理解为什么不是这种情况)。 - mins

46

Unutbu提供了非常全面的答案(他们也得到了我的+1),但这里有一个低技术的选择:

>>> x=np.random.randn(5)
>>> x
array([ 0.25276524,  2.28334499, -1.88221637,  0.69949927,  1.0285625 ])
>>> ['{:.2f}'.format(i) for i in x]
['0.25', '2.28', '-1.88', '0.70', '1.03']

作为一个函数(使用format()语法进行格式化):
def ndprint(a, format_string ='{0:.2f}'):
    print [format_string.format(v,i) for i,v in enumerate(a)]

使用方法:

>>> ndprint(x)
['0.25', '2.28', '-1.88', '0.70', '1.03']

>>> ndprint(x, '{:10.4e}')
['2.5277e-01', '2.2833e+00', '-1.8822e+00', '6.9950e-01', '1.0286e+00']

>>> ndprint(x, '{:.8g}')
['0.25276524', '2.283345', '-1.8822164', '0.69949927', '1.0285625']

数组的索引可以在格式字符串中访问:
>>> ndprint(x, 'Element[{1:d}]={0:.2f}')
['Element[0]=0.25', 'Element[1]=2.28', 'Element[2]=-1.88', 'Element[3]=0.70', 'Element[4]=1.03']

20
FYI,Numpy 1.15(发布日期待定)将 包括用于在本地设置打印选项的上下文管理器。这意味着以下内容将与接受的答案中的对应示例相同,而无需编写自己的上下文管理器。例如,使用他们的示例:
x = np.random.random(10)
with np.printoptions(precision=3, suppress=True):
    print(x)
    # [ 0.073  0.461  0.689  0.754  0.624  0.901  0.049  0.582  0.557  0.348]

17

在今天的numpy版本中,使得将结果作为字符串轻松获取的宝石被隐藏在denis的答案中: np.array2string

>>> import numpy as np
>>> x=np.random.random(10)
>>> np.array2string(x, formatter={'float_kind':'{0:.3f}'.format})
'[0.599 0.847 0.513 0.155 0.844 0.753 0.920 0.797 0.427 0.420]'

这是两个回答中的一个,实际上满足了要求,没有重复造轮子。 - undefined

11

多年后,下面还有另一个。但是对于日常使用,我只需要

np.set_printoptions( threshold=20, edgeitems=10, linewidth=140,
    formatter = dict( float = lambda x: "%.3g" % x ))  # float arrays %.3g

''' printf( "... %.3g ... %.1f  ...", arg, arg ... ) for numpy arrays too

Example:
    printf( """ x: %.3g   A: %.1f   s: %s   B: %s """,
                   x,        A,        "str",  B )

If `x` and `A` are numbers, this is like `"format" % (x, A, "str", B)` in python.
If they're numpy arrays, each element is printed in its own format:
    `x`: e.g. [ 1.23 1.23e-6 ... ]  3 digits
    `A`: [ [ 1 digit after the decimal point ... ] ... ]
with the current `np.set_printoptions()`. For example, with
    np.set_printoptions( threshold=100, edgeitems=3, suppress=True )
only the edges of big `x` and `A` are printed.
`B` is printed as `str(B)`, for any `B` -- a number, a list, a numpy object ...

`printf()` tries to handle too few or too many arguments sensibly,
but this is iffy and subject to change.

How it works:
numpy has a function `np.array2string( A, "%.3g" )` (simplifying a bit).
`printf()` splits the format string, and for format / arg pairs
    format: % d e f g
    arg: try `np.asanyarray()`
-->  %s  np.array2string( arg, format )
Other formats and non-ndarray args are left alone, formatted as usual.

Notes:

`printf( ... end= file= )` are passed on to the python `print()` function.

Only formats `% [optional width . precision] d e f g` are implemented,
not `%(varname)format` .

%d truncates floats, e.g. 0.9 and -0.9 to 0; %.0f rounds, 0.9 to 1 .
%g is the same as %.6g, 6 digits.
%% is a single "%" character.

The function `sprintf()` returns a long string. For example,
    title = sprintf( "%s  m %g  n %g  X %.3g",
                    __file__, m, n, X )
    print( title )
    ...
    pl.title( title )

Module globals:
_fmt = "%.3g"  # default for extra args
_squeeze = np.squeeze  # (n,1) (1,n) -> (n,) print in 1 line not n

See also:
http://docs.scipy.org/doc/numpy/reference/generated/numpy.set_printoptions.html
http://docs.python.org/2.7/library/stdtypes.html#string-formatting

'''
# https://dev59.com/eHE85IYBdhLWcg3wNgrk


#...............................................................................
from __future__ import division, print_function
import re
import numpy as np

__version__ = "2014-02-03 feb denis"

_splitformat = re.compile( r'''(
    %
    (?<! %% )  # not %%
    -? [ \d . ]*  # optional width.precision
    \w
    )''', re.X )
    # ... %3.0f  ... %g  ... %-10s ...
    # -> ['...' '%3.0f' '...' '%g' '...' '%-10s' '...']
    # odd len, first or last may be ""

_fmt = "%.3g"  # default for extra args
_squeeze = np.squeeze  # (n,1) (1,n) -> (n,) print in 1 line not n

#...............................................................................
def printf( format, *args, **kwargs ):
    print( sprintf( format, *args ), **kwargs )  # end= file=

printf.__doc__ = __doc__


def sprintf( format, *args ):
    """ sprintf( "text %.3g text %4.1f ... %s ... ", numpy arrays or ... )
        %[defg] array -> np.array2string( formatter= )
    """
    args = list(args)
    if not isinstance( format, basestring ):
        args = [format] + args
        format = ""

    tf = _splitformat.split( format )  # [ text %e text %f ... ]
    nfmt = len(tf) // 2
    nargs = len(args)
    if nargs < nfmt:
        args += (nfmt - nargs) * ["?arg?"]
    elif nargs > nfmt:
        tf += (nargs - nfmt) * [_fmt, " "]  # default _fmt

    for j, arg in enumerate( args ):
        fmt = tf[ 2*j + 1 ]
        if arg is None \
        or isinstance( arg, basestring ) \
        or (hasattr( arg, "__iter__" ) and len(arg) == 0):
            tf[ 2*j + 1 ] = "%s"  # %f -> %s, not error
            continue
        args[j], isarray = _tonumpyarray(arg)
        if isarray  and fmt[-1] in "defgEFG":
            tf[ 2*j + 1 ] = "%s"
            fmtfunc = (lambda x: fmt % x)
            formatter = dict( float_kind=fmtfunc, int=fmtfunc )
            args[j] = np.array2string( args[j], formatter=formatter )
    try:
        return "".join(tf) % tuple(args)
    except TypeError:  # shouldn't happen
        print( "error: tf %s  types %s" % (tf, map( type, args )))
        raise


def _tonumpyarray( a ):
    """ a, isarray = _tonumpyarray( a )
        ->  scalar, False
            np.asanyarray(a), float or int
            a, False
    """
    a = getattr( a, "value", a )  # cvxpy
    if np.isscalar(a):
        return a, False
    if hasattr( a, "__iter__" )  and len(a) == 0:
        return a, False
    try:
        # map .value ?
        a = np.asanyarray( a )
    except ValueError:
        return a, False
    if hasattr( a, "dtype" )  and a.dtype.kind in "fi":  # complex ?
        if callable( _squeeze ):
            a = _squeeze( a )  # np.squeeze
        return a, True
    else:
        return a, False


#...............................................................................
if __name__ == "__main__":
    import sys

    n = 5
    seed = 0
        # run this.py n= ...  in sh or ipython
    for arg in sys.argv[1:]:
        exec( arg )
    np.set_printoptions( 1, threshold=4, edgeitems=2, linewidth=80, suppress=True )
    np.random.seed(seed)

    A = np.random.exponential( size=(n,n) ) ** 10
    x = A[0]

    printf( "x: %.3g  \nA: %.1f  \ns: %s  \nB: %s ",
                x,         A,         "str",   A )
    printf( "x %%d: %d", x )
    printf( "x %%.0f: %.0f", x )
    printf( "x %%.1e: %.1e", x )
    printf( "x %%g: %g", x )
    printf( "x %%s uses np printoptions: %s", x )

    printf( "x with default _fmt: ", x )
    printf( "no args" )
    printf( "too few args: %g %g", x )
    printf( x )
    printf( x, x )
    printf( None )
    printf( "[]:", [] )
    printf( "[3]:", [3] )
    printf( np.array( [] ))
    printf( [[]] )  # squeeze

9

以下是我使用的内容,非常简单易懂:

print(np.vectorize("%.2f".__mod__)(sparse))

8
NumPy 数组具有方法 round(precision),该方法返回一个新的 NumPy 数组,其中的元素按照精度进行四舍五入。
import numpy as np

x = np.random.random([5,5])
print(x.round(3))

1
当我将数组传递给matplotlib的ylabel时,这对我很有用,谢谢。 - Hans

6

惊讶地发现没有提到around方法 - 这意味着不需要改变打印选项。

import numpy as np

x = np.random.random([5,5])
print(np.around(x,decimals=3))

Output:
[[0.475 0.239 0.183 0.991 0.171]
 [0.231 0.188 0.235 0.335 0.049]
 [0.87  0.212 0.219 0.9   0.3  ]
 [0.628 0.791 0.409 0.5   0.319]
 [0.614 0.84  0.812 0.4   0.307]]

1
但这意味着每次想要打印它时都要创建数组的第二个副本,这可能会阻止某些使用情况。 - Brandon Rhodes

2

我经常希望不同的列具有不同的格式。以下是我如何使用一些不同的格式打印简单的二维数组,通过将NumPy数组的(切片)转换为元组:

import numpy as np
dat = np.random.random((10,11))*100  # Array of random values between 0 and 100
print(dat)                           # Lines get truncated and are hard to read
for i in range(10):
    print((4*"%6.2f"+7*"%9.4f") % tuple(dat[i,:]))

网页内容由stack overflow 提供, 点击上面的
可以查看英文原文,
原文链接