python如何用零填充numpy数组

问题:python如何用零填充numpy数组

我想知道如何使用python 2.6.6和numpy版本1.5.0用零填充2D numpy数组。抱歉! 但是这些是我的局限性。因此我不能使用np.pad。例如,我想a用零填充以使其形状匹配b。我想这样做的原因是我可以这样做:

b-a

这样

>>> a
array([[ 1.,  1.,  1.,  1.,  1.],
       [ 1.,  1.,  1.,  1.,  1.],
       [ 1.,  1.,  1.,  1.,  1.]])
>>> b
array([[ 3.,  3.,  3.,  3.,  3.,  3.],
       [ 3.,  3.,  3.,  3.,  3.,  3.],
       [ 3.,  3.,  3.,  3.,  3.,  3.],
       [ 3.,  3.,  3.,  3.,  3.,  3.]])
>>> c
array([[1, 1, 1, 1, 1, 0],
       [1, 1, 1, 1, 1, 0],
       [1, 1, 1, 1, 1, 0],
       [0, 0, 0, 0, 0, 0]])

我能想到的唯一方法是追加,但这看起来很丑。是否有可能使用更清洁的解决方案b.shape

编辑,谢谢MSeiferts的答案。我必须清理一下,这就是我得到的:

def pad(array, reference_shape, offsets):
    """
    array: Array to be padded
    reference_shape: tuple of size of ndarray to create
    offsets: list of offsets (number of elements must be equal to the dimension of the array)
    will throw a ValueError if offsets is too big and the reference_shape cannot handle the offsets
    """

    # Create an array of zeros with the reference shape
    result = np.zeros(reference_shape)
    # Create a list of slices from offset to offset + shape in each dimension
    insertHere = [slice(offsets[dim], offsets[dim] + array.shape[dim]) for dim in range(array.ndim)]
    # Insert the array in the result at the specified offsets
    result[insertHere] = array
    return result

I want to know how I can pad a 2D numpy array with zeros using python 2.6.6 with numpy version 1.5.0. Sorry! But these are my limitations. Therefore I cannot use np.pad. For example, I want to pad a with zeros such that its shape matches b. The reason why I want to do this is so I can do:

b-a

such that

>>> a
array([[ 1.,  1.,  1.,  1.,  1.],
       [ 1.,  1.,  1.,  1.,  1.],
       [ 1.,  1.,  1.,  1.,  1.]])
>>> b
array([[ 3.,  3.,  3.,  3.,  3.,  3.],
       [ 3.,  3.,  3.,  3.,  3.,  3.],
       [ 3.,  3.,  3.,  3.,  3.,  3.],
       [ 3.,  3.,  3.,  3.,  3.,  3.]])
>>> c
array([[1, 1, 1, 1, 1, 0],
       [1, 1, 1, 1, 1, 0],
       [1, 1, 1, 1, 1, 0],
       [0, 0, 0, 0, 0, 0]])

The only way I can think of doing this is appending, however this seems pretty ugly. is there a cleaner solution possibly using b.shape?

Edit, Thank you to MSeiferts answer. I had to clean it up a bit, and this is what I got:

def pad(array, reference_shape, offsets):
    """
    array: Array to be padded
    reference_shape: tuple of size of ndarray to create
    offsets: list of offsets (number of elements must be equal to the dimension of the array)
    will throw a ValueError if offsets is too big and the reference_shape cannot handle the offsets
    """

    # Create an array of zeros with the reference shape
    result = np.zeros(reference_shape)
    # Create a list of slices from offset to offset + shape in each dimension
    insertHere = [slice(offsets[dim], offsets[dim] + array.shape[dim]) for dim in range(array.ndim)]
    # Insert the array in the result at the specified offsets
    result[insertHere] = array
    return result

回答 0

很简单,使用参考形状创建一个包含零的数组:

result = np.zeros(b.shape)
# actually you can also use result = np.zeros_like(b) 
# but that also copies the dtype not only the shape

然后在需要的地方插入数组:

result[:a.shape[0],:a.shape[1]] = a

瞧,您已经填充了它:

print(result)
array([[ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.]])

如果您定义应该在左上方插入元素的位置,也可以使其更通用一些

result = np.zeros_like(b)
x_offset = 1  # 0 would be what you wanted
y_offset = 1  # 0 in your case
result[x_offset:a.shape[0]+x_offset,y_offset:a.shape[1]+y_offset] = a
result

array([[ 0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  1.,  1.,  1.,  1.,  1.],
       [ 0.,  1.,  1.,  1.,  1.,  1.],
       [ 0.,  1.,  1.,  1.,  1.,  1.]])

但请注意,偏移量不要超过允许的范围。例如x_offset = 2,这将失败。


如果您有任意数量的维,则可以定义切片列表以插入原始数组。我发现有趣的是可以玩一下,并创建了一个填充函数,该函数可以填充(偏移)任意形状的数组,只要数组和引用的维数相同且偏移量不太大即可。

def pad(array, reference, offsets):
    """
    array: Array to be padded
    reference: Reference array with the desired shape
    offsets: list of offsets (number of elements must be equal to the dimension of the array)
    """
    # Create an array of zeros with the reference shape
    result = np.zeros(reference.shape)
    # Create a list of slices from offset to offset + shape in each dimension
    insertHere = [slice(offset[dim], offset[dim] + array.shape[dim]) for dim in range(a.ndim)]
    # Insert the array in the result at the specified offsets
    result[insertHere] = a
    return result

和一些测试用例:

import numpy as np

# 1 Dimension
a = np.ones(2)
b = np.ones(5)
offset = [3]
pad(a, b, offset)

# 3 Dimensions

a = np.ones((3,3,3))
b = np.ones((5,4,3))
offset = [1,0,0]
pad(a, b, offset)

Very simple, you create an array containing zeros using the reference shape:

result = np.zeros(b.shape)
# actually you can also use result = np.zeros_like(b) 
# but that also copies the dtype not only the shape

and then insert the array where you need it:

result[:a.shape[0],:a.shape[1]] = a

and voila you have padded it:

print(result)
array([[ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.]])

You can also make it a bit more general if you define where your upper left element should be inserted

result = np.zeros_like(b)
x_offset = 1  # 0 would be what you wanted
y_offset = 1  # 0 in your case
result[x_offset:a.shape[0]+x_offset,y_offset:a.shape[1]+y_offset] = a
result

array([[ 0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  1.,  1.,  1.,  1.,  1.],
       [ 0.,  1.,  1.,  1.,  1.,  1.],
       [ 0.,  1.,  1.,  1.,  1.,  1.]])

but then be careful that you don’t have offsets bigger than allowed. For x_offset = 2 for example this will fail.


If you have an arbitary number of dimensions you can define a list of slices to insert the original array. I’ve found it interesting to play around a bit and created a padding function that can pad (with offset) an arbitary shaped array as long as the array and reference have the same number of dimensions and the offsets are not too big.

def pad(array, reference, offsets):
    """
    array: Array to be padded
    reference: Reference array with the desired shape
    offsets: list of offsets (number of elements must be equal to the dimension of the array)
    """
    # Create an array of zeros with the reference shape
    result = np.zeros(reference.shape)
    # Create a list of slices from offset to offset + shape in each dimension
    insertHere = [slice(offset[dim], offset[dim] + array.shape[dim]) for dim in range(a.ndim)]
    # Insert the array in the result at the specified offsets
    result[insertHere] = a
    return result

And some test cases:

import numpy as np

# 1 Dimension
a = np.ones(2)
b = np.ones(5)
offset = [3]
pad(a, b, offset)

# 3 Dimensions

a = np.ones((3,3,3))
b = np.ones((5,4,3))
offset = [1,0,0]
pad(a, b, offset)

回答 1

NumPy 1.7.0(numpy.pad添加时)现在已经很老了(它于2013年发布),因此即使问题要求使用不使用该功能的方法,我也认为了解使用可以实现该功能很有用numpy.pad

实际上很简单:

>>> import numpy as np
>>> a = np.array([[ 1.,  1.,  1.,  1.,  1.],
...               [ 1.,  1.,  1.,  1.,  1.],
...               [ 1.,  1.,  1.,  1.,  1.]])
>>> np.pad(a, [(0, 1), (0, 1)], mode='constant')
array([[ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.]])

在这种情况下,我使用0的默认值mode='constant'。但是也可以通过显式传递它来指定它:

>>> np.pad(a, [(0, 1), (0, 1)], mode='constant', constant_values=0)
array([[ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.]])

以防第二个参数([(0, 1), (0, 1)])令人困惑:每个列表项(在本例中为元组)都对应于一个维度,并且其中的每个项都表示(第一个元素)之前之后(第二个元素)的填充。所以:

[(0, 1), (0, 1)]
         ^^^^^^------ padding for second dimension
 ^^^^^^-------------- padding for first dimension

  ^------------------ no padding at the beginning of the first axis
     ^--------------- pad with one "value" at the end of the first axis.

在这种情况下,第一轴和第二轴的填充相同,因此也可以只传入2元组:

>>> np.pad(a, (0, 1), mode='constant')
array([[ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.]])

如果前后的填充相同,甚至可以省略该元组(尽管在这种情况下不适用):

>>> np.pad(a, 1, mode='constant')
array([[ 0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  1.,  1.,  1.,  1.,  1.,  0.],
       [ 0.,  1.,  1.,  1.,  1.,  1.,  0.],
       [ 0.,  1.,  1.,  1.,  1.,  1.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.]])

或者,如果前后的填充相同但轴的填充不同,则也可以在内部元组中省略第二个参数:

>>> np.pad(a, [(1, ), (2, )], mode='constant')
array([[ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  1.,  1.,  1.,  1.,  1.,  0.,  0.],
       [ 0.,  0.,  1.,  1.,  1.,  1.,  1.,  0.,  0.],
       [ 0.,  0.,  1.,  1.,  1.,  1.,  1.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.]])

但是我倾向于始终使用显式的,因为这样做很容易犯错(当NumPys的期望与您的意图有所不同时):

>>> np.pad(a, [1, 2], mode='constant')
array([[ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  1.,  1.,  1.,  1.,  1.,  0.,  0.],
       [ 0.,  1.,  1.,  1.,  1.,  1.,  0.,  0.],
       [ 0.,  1.,  1.,  1.,  1.,  1.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.]])

在这里,NumPy认为您希望在每个轴前填充1个元素,在每个轴后填充2个元素!即使您打算用轴1中的1个元素和轴2中的2个元素填充。

我使用元组列表进行填充,请注意,这只是“我的约定”,您也可以使用列表列表或元组的元组,甚至数组的元组。NumPy只是检查参数的长度(如果没有长度)和每个项目的长度(或者如果有长度)!

NumPy 1.7.0 (when numpy.pad was added) is pretty old now (it was released in 2013) so even though the question asked for a way without using that function I thought it could be useful to know how that could be achieved using numpy.pad.

It’s actually pretty simple:

>>> import numpy as np
>>> a = np.array([[ 1.,  1.,  1.,  1.,  1.],
...               [ 1.,  1.,  1.,  1.,  1.],
...               [ 1.,  1.,  1.,  1.,  1.]])
>>> np.pad(a, [(0, 1), (0, 1)], mode='constant')
array([[ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.]])

In this case I used that 0 is the default value for mode='constant'. But it could also be specified by passing it in explicitly:

>>> np.pad(a, [(0, 1), (0, 1)], mode='constant', constant_values=0)
array([[ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.]])

Just in case the second argument ([(0, 1), (0, 1)]) seems confusing: Each list item (in this case tuple) corresponds to a dimension and item therein represents the padding before (first element) and after (second element). So:

[(0, 1), (0, 1)]
         ^^^^^^------ padding for second dimension
 ^^^^^^-------------- padding for first dimension

  ^------------------ no padding at the beginning of the first axis
     ^--------------- pad with one "value" at the end of the first axis.

In this case the padding for the first and second axis are identical, so one could also just pass in the 2-tuple:

>>> np.pad(a, (0, 1), mode='constant')
array([[ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 1.,  1.,  1.,  1.,  1.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.]])

In case the padding before and after is identical one could even omit the tuple (not applicable in this case though):

>>> np.pad(a, 1, mode='constant')
array([[ 0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  1.,  1.,  1.,  1.,  1.,  0.],
       [ 0.,  1.,  1.,  1.,  1.,  1.,  0.],
       [ 0.,  1.,  1.,  1.,  1.,  1.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.]])

Or if the padding before and after is identical but different for the axis, you could also omit the second argument in the inner tuples:

>>> np.pad(a, [(1, ), (2, )], mode='constant')
array([[ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  1.,  1.,  1.,  1.,  1.,  0.,  0.],
       [ 0.,  0.,  1.,  1.,  1.,  1.,  1.,  0.,  0.],
       [ 0.,  0.,  1.,  1.,  1.,  1.,  1.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.]])

However I tend to prefer to always use the explicit one, because it’s just to easy to make mistakes (when NumPys expectations differ from your intentions):

>>> np.pad(a, [1, 2], mode='constant')
array([[ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  1.,  1.,  1.,  1.,  1.,  0.,  0.],
       [ 0.,  1.,  1.,  1.,  1.,  1.,  0.,  0.],
       [ 0.,  1.,  1.,  1.,  1.,  1.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.]])

Here NumPy thinks you wanted to pad all axis with 1 element before and 2 elements after each axis! Even if you intended it to pad with 1 element in axis 1 and 2 elements for axis 2.

I used lists of tuples for the padding, note that this is just “my convention”, you could also use lists of lists or tuples of tuples, or even tuples of arrays. NumPy just checks the length of the argument (or if it doesn’t have a length) and the length of each item (or if it has a length)!


回答 2

我了解您的主要问题是您需要计算,d=b-a但数组的大小不同。无需中间填充c

您可以解决此问题而无需填充:

import numpy as np

a = np.array([[ 1.,  1.,  1.,  1.,  1.],
              [ 1.,  1.,  1.,  1.,  1.],
              [ 1.,  1.,  1.,  1.,  1.]])

b = np.array([[ 3.,  3.,  3.,  3.,  3.,  3.],
              [ 3.,  3.,  3.,  3.,  3.,  3.],
              [ 3.,  3.,  3.,  3.,  3.,  3.],
              [ 3.,  3.,  3.,  3.,  3.,  3.]])

d = b.copy()
d[:a.shape[0],:a.shape[1]] -=  a

print d

输出:

[[ 2.  2.  2.  2.  2.  3.]
 [ 2.  2.  2.  2.  2.  3.]
 [ 2.  2.  2.  2.  2.  3.]
 [ 3.  3.  3.  3.  3.  3.]]

I understand that your main problem is that you need to calculate d=b-a but your arrays have different sizes. There is no need for an intermediate padded c

You can solve this without padding:

import numpy as np

a = np.array([[ 1.,  1.,  1.,  1.,  1.],
              [ 1.,  1.,  1.,  1.,  1.],
              [ 1.,  1.,  1.,  1.,  1.]])

b = np.array([[ 3.,  3.,  3.,  3.,  3.,  3.],
              [ 3.,  3.,  3.,  3.,  3.,  3.],
              [ 3.,  3.,  3.,  3.,  3.,  3.],
              [ 3.,  3.,  3.,  3.,  3.,  3.]])

d = b.copy()
d[:a.shape[0],:a.shape[1]] -=  a

print d

Output:

[[ 2.  2.  2.  2.  2.  3.]
 [ 2.  2.  2.  2.  2.  3.]
 [ 2.  2.  2.  2.  2.  3.]
 [ 3.  3.  3.  3.  3.  3.]]

回答 3

如果需要向数组添加1s的范围:

>>> mat = np.zeros((4,4), np.int32)
>>> mat
array([[0, 0, 0, 0],
       [0, 0, 0, 0],
       [0, 0, 0, 0],
       [0, 0, 0, 0]])
>>> mat[0,:] = mat[:,0] = mat[:,-1] =  mat[-1,:] = 1
>>> mat
array([[1, 1, 1, 1],
       [1, 0, 0, 1],
       [1, 0, 0, 1],
       [1, 1, 1, 1]])

In case you need to add a fence of 1s to an array:

>>> mat = np.zeros((4,4), np.int32)
>>> mat
array([[0, 0, 0, 0],
       [0, 0, 0, 0],
       [0, 0, 0, 0],
       [0, 0, 0, 0]])
>>> mat[0,:] = mat[:,0] = mat[:,-1] =  mat[-1,:] = 1
>>> mat
array([[1, 1, 1, 1],
       [1, 0, 0, 1],
       [1, 0, 0, 1],
       [1, 1, 1, 1]])

回答 4

我知道我有点晚了,但是如果您想执行相对填充(aka边缘填充),可以通过以下方法实现它。请注意,分配的第一个实例将导致零填充,因此您可以将其用于零填充和相对填充(这是将原始数组的边值复制到填充数组中的地方)。

def replicate_padding(arr):
    """Perform replicate padding on a numpy array."""
    new_pad_shape = tuple(np.array(arr.shape) + 2) # 2 indicates the width + height to change, a (512, 512) image --> (514, 514) padded image.
    padded_array = np.zeros(new_pad_shape) #create an array of zeros with new dimensions
    
    # perform replication
    padded_array[1:-1,1:-1] = arr        # result will be zero-pad
    padded_array[0,1:-1] = arr[0]        # perform edge pad for top row
    padded_array[-1, 1:-1] = arr[-1]     # edge pad for bottom row
    padded_array.T[0, 1:-1] = arr.T[0]   # edge pad for first column
    padded_array.T[-1, 1:-1] = arr.T[-1] # edge pad for last column
    
    #at this point, all values except for the 4 corners should have been replicated
    padded_array[0][0] = arr[0][0]     # top left corner
    padded_array[-1][0] = arr[-1][0]   # bottom left corner
    padded_array[0][-1] = arr[0][-1]   # top right corner 
    padded_array[-1][-1] = arr[-1][-1] # bottom right corner

    return padded_array

复杂度分析:

对此的最佳解决方案是numpy的pad方法。在平均运行5次之后,具有相对填充的np.pad仅8%比上面定义的函数好。这表明这是相对填充和零填充的最佳方法。


#My method, replicate_padding
start = time.time()
padded = replicate_padding(input_image)
end = time.time()
delta0 = end - start

#np.pad with edge padding
start = time.time()
padded = np.pad(input_image, 1, mode='edge')
end = time.time()
delta = end - start


print(delta0) # np Output: 0.0008790493011474609 
print(delta)  # My Output: 0.0008130073547363281
print(100*((delta0-delta)/delta)) # Percent difference: 8.12316715542522%

I know I’m a bit late to this, but in case you wanted to perform relative padding (aka edge padding), here’s how you can implement it. Note that the very first instance of assignment results in zero-padding, so you can use this for both zero-padding and relative padding (this is where you copy the edge values of the original array into the padded array).

def replicate_padding(arr):
    """Perform replicate padding on a numpy array."""
    new_pad_shape = tuple(np.array(arr.shape) + 2) # 2 indicates the width + height to change, a (512, 512) image --> (514, 514) padded image.
    padded_array = np.zeros(new_pad_shape) #create an array of zeros with new dimensions
    
    # perform replication
    padded_array[1:-1,1:-1] = arr        # result will be zero-pad
    padded_array[0,1:-1] = arr[0]        # perform edge pad for top row
    padded_array[-1, 1:-1] = arr[-1]     # edge pad for bottom row
    padded_array.T[0, 1:-1] = arr.T[0]   # edge pad for first column
    padded_array.T[-1, 1:-1] = arr.T[-1] # edge pad for last column
    
    #at this point, all values except for the 4 corners should have been replicated
    padded_array[0][0] = arr[0][0]     # top left corner
    padded_array[-1][0] = arr[-1][0]   # bottom left corner
    padded_array[0][-1] = arr[0][-1]   # top right corner 
    padded_array[-1][-1] = arr[-1][-1] # bottom right corner

    return padded_array

Complexity Analysis:

The optimal solution for this is numpy’s pad method. After averaging for 5 runs, np.pad with relative padding is only 8% better than the function defined above. This shows that this is fairly an optimal method for relative and zero-padding padding.


#My method, replicate_padding
start = time.time()
padded = replicate_padding(input_image)
end = time.time()
delta0 = end - start

#np.pad with edge padding
start = time.time()
padded = np.pad(input_image, 1, mode='edge')
end = time.time()
delta = end - start


print(delta0) # np Output: 0.0008790493011474609 
print(delta)  # My Output: 0.0008130073547363281
print(100*((delta0-delta)/delta)) # Percent difference: 8.12316715542522%