我正在编写一个相当简单的函数来执行在日志空间中应用梯形规则的集成。
我想添加axis参数来实现类似于numpy.trapz
函数的功能,但我对如何正确实现它有点困惑。
不可播放的功能如下所示:
import numpy as np
def logtrapz(y, x):
logx = np.log(x)
dlogx = np.diff(logx)
logy = np.log(y)
dlogy = np.diff(logy)
b = dlogx + dlogy
a = np.exp(logx + logy)
dF = a[:-1] * (np.exp(b) - 1)/b * dlogx
return np.sum(dF)
这适用于一维输入。
我认为解决方案在于numpy.expand_dims
,但我不确定如何实现它
为了说明互动环节中的slice
探索:
In [216]: slice(None)
Out[216]: slice(None, None, None)
In [217]: slice??
Init signature: slice(self, /, *args, **kwargs)
Docstring:
slice(stop)
slice(start, stop[, step])
Create a slice object. This is used for extended slicing (e.g. a[0:10:2]).
Type: type
Subclasses:
In [218]: np.s_[:]
Out[218]: slice(None, None, None)
我没有看过np.trapz
代码,但我知道其他numpy
函数经常构造索引元组,当它们需要是axis
一般。
例如,3d数组的通用索引:
In [221]: arr = np.arange(24).reshape(2,3,4)
In [223]: idx = [slice(None) for _ in range(3)]
In [224]: idx
Out[224]: [slice(None, None, None), slice(None, None, None), slice(None, None, None)]
In [225]: idx[1]=1
In [226]: idx
Out[226]: [slice(None, None, None), 1, slice(None, None, None)]
In [227]: tuple(idx)
Out[227]: (slice(None, None, None), 1, slice(None, None, None))
In [228]: arr[tuple(idx)] # arr[:,1,:]
Out[228]:
array([[ 4, 5, 6, 7],
[16, 17, 18, 19]])
In [229]: idx[2]=2
In [230]: arr[tuple(idx)] # arr[:,1,2]
Out[230]: array([ 6, 18])
我解决了这个复制numpy.trapz
中使用的方法。这有点令人费解,但效果很好。
对于未来的读者,上述功能的可播放版本是
import numpy as np
def logtrapz(y, x, axis=-1):
x = np.asanyarray(x)
logx = np.log(x)
if x.ndim == 1:
dlogx = np.diff(logx)
# reshape to correct shape
shape1 = [1]*y.ndim
shape1[axis] = dlogx.shape[0]
shape2 = [1]*y.ndim
shape2[axis] = logx.shape[0]
dlogx = dlogx.reshape(shape1)
logx = logx.reshape(shape2)
else:
dlogx = np.diff(x, axis=axis)
nd = y.ndim
slice1 = [slice(None)]*nd
slice2 = [slice(None)]*nd
slice1[axis] = slice(None, -1)
slice2[axis] = slice(1, None)
slice1 = tuple(slice1)
slice2 = tuple(slice2)
logy = np.log(y)
dlogy = logy[slice2] - logy[slice1]
b = dlogx + dlogy
a = np.exp(logx + logy)
dF = a[slice1] * (np.exp(b) - 1)/b * dlogx
np.sum(dF, axis=axis)
为了实现“可广播性”,采用reshape
和slice
的混合,明确地创建具有所需输出形状的“形状”矢量。
我认为这可以通过更短更简洁的方式实现,但显然这是在numpy本身实现的方式。