scipy.ndimage.label: include error margin - python

After reading an interesting topic on scipy.ndimage.label (Variable area threshold for identifying objects - python), I'd like to include an 'error margin' in the labelling.
In the above linked discussion:
How can the blue dot on top be included, too (let's say it is wrongly disconnected from the orange, biggest, object)?
I found the structure attribute, which should be able to include that dot by changing the array (from np.ones(3,3,3) to anything more than that (I'd like it to be 3D). However, adjusting the 'structure' attribute to a larger array does not seem to work, unfortunately. It either gives an error of dimensions (RuntimeError: structure and input must have equal rank
) or it does not change anything..
Thanks!
this is the code:
labels, nshapes = ndimage.label(a, structure=np.ones((3,3,3)))
in which a is a 3D array.

Here's a possible approach that uses scipy.ndimage.binary_dilation. It is easier to see what is going on in a 2D example, but I'll show how to generalize to 3D at the end.
In [103]: a
Out[103]:
array([[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[1, 1, 0, 0, 1, 0, 0],
[1, 1, 0, 0, 0, 1, 1],
[0, 0, 0, 0, 0, 1, 1],
[1, 1, 1, 0, 0, 0, 0]])
In [104]: from scipy.ndimage import label, binary_dilation
Extend each "shape" by one pixel down and to the right:
In [105]: b = binary_dilation(a, structure=np.array([[0, 0, 0], [0, 1, 1], [0, 1, 1]])).astype(int)
In [106]: b
Out[106]:
array([[0, 0, 0, 1, 1, 0, 0],
[0, 0, 0, 1, 1, 0, 0],
[1, 1, 1, 0, 1, 1, 0],
[1, 1, 1, 0, 1, 1, 1],
[1, 1, 1, 0, 0, 1, 1],
[1, 1, 1, 1, 0, 1, 1]])
Apply label to the padded array:
In [107]: labels, numlabels = label(b)
In [108]: numlabels
Out[108]: 2
In [109]: labels
Out[109]:
array([[0, 0, 0, 1, 1, 0, 0],
[0, 0, 0, 1, 1, 0, 0],
[2, 2, 2, 0, 1, 1, 0],
[2, 2, 2, 0, 1, 1, 1],
[2, 2, 2, 0, 0, 1, 1],
[2, 2, 2, 2, 0, 1, 1]], dtype=int32)
By multiplying a by labels, we get the desired array of labels of a:
In [110]: alab = labels*a
In [111]: alab
Out[111]:
array([[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[2, 2, 0, 0, 1, 0, 0],
[2, 2, 0, 0, 0, 1, 1],
[0, 0, 0, 0, 0, 1, 1],
[2, 2, 2, 0, 0, 0, 0]])
(This assumes that the values in a are 0 or 1. If they are not, you can use alab = labels * (a > 0).)
For a 3D input, you have to change the structure argument to binary_dilation:
struct = np.zeros((3, 3, 3), dtype=int)
struct[1:, 1:, 1:] = 1
b = binary_dilation(a, structure=struct).astype(int)

Related

Python, Numpy. Find values in 2d array and replace neighbors with 1

I have a 10x10 array with zeros and ones.
I would like to:
find the position of each cell with a value of 1.
replace all the neighbors with 1. neighbors= any cell to a n=1 distance (also diagonal).
Example:
array([[0, 1, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 1, 0, 0],
[1, 0, 0, 0, 0],
[0, 0, 0, 1, 1]])
output:
array([[1, 1, 1, 1, 0],
[1, 1, 1, 1, 0],
[1, 1, 1, 1, 0],
[1, 1, 1, 1, 1],
[1, 1, 1, 1, 1]])
I am trying finding indexes but It does not work:
a=np.where(a==1)+1
From other post I also try getting the neighbors with this function:
def n_closest(x,n,d=1):
return x[n[0]-d:n[0]+d+1,n[1]-d:n[1]+d+1]
But this does not work for the edges
Thanks
If you don't mind using scipy, a 2D convolution will solve the problem quickly:
import numpy as np
from scipy import signal
# Input array
X = np.array([[0, 1, 0, 0, 0],
[0, 0, 0, 0, 0],
[0, 0, 1, 0, 0],
[1, 0, 0, 0, 0],
[0, 0, 0, 1, 1]])
# We apply a 2D convolution with a 3x3 kernel and we check which value are bigger than 0.
R = (signal.convolve2d(X,np.ones((3,3)),mode='same')>0).astype(int)
# R = array([[1, 1, 1, 0, 0],
# [1, 1, 1, 1, 0],
# [1, 1, 1, 1, 0],
# [1, 1, 1, 1, 1],
# [1, 1, 1, 1, 1]])
# Finally we extract the index
x,y = np.where(R)

How to define an array with all possible combinations of numbers

I want to define an array with a given number of columns (let's say n=5) and in each cell of the array, the value can be either 0 or 1. And I would like to create all possibilities of ones and zeros, which means, that each row would represent one possible vector with n elements.
In other words, I want the table to look like this:
I know that create the vector of ones and zeros is quite easy but how can I ensure that the vectors would not repeat in the table and that there will be all possible combinations included (If my math is correct the table should have 2**5 = 32 rows)
How can I do it in Python? Thank you very much
Easy with itertools:
itertools.product(*[[0, 1]] * 3)
results in
[(0, 0, 0),
(0, 0, 1),
(0, 1, 0),
(0, 1, 1),
(1, 0, 0),
(1, 0, 1),
(1, 1, 0),
(1, 1, 1)]
You could generate all the numbers up to 32, and convert each to binary representation using bit shifts.
combs = [[(n >> p) & 1 for p in range(4, -1, -1)] for n in range(32)]
which gives combs as:
[
[0, 0, 0, 0, 0],
[0, 0, 0, 0, 1],
[0, 0, 0, 1, 0],
[0, 0, 0, 1, 1],
[0, 0, 1, 0, 0],
[0, 0, 1, 0, 1],
[0, 0, 1, 1, 0],
[0, 0, 1, 1, 1],
[0, 1, 0, 0, 0],
[0, 1, 0, 0, 1],
[0, 1, 0, 1, 0],
[0, 1, 0, 1, 1],
[0, 1, 1, 0, 0],
[0, 1, 1, 0, 1],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 1],
[1, 0, 0, 0, 0],
[1, 0, 0, 0, 1],
[1, 0, 0, 1, 0],
[1, 0, 0, 1, 1],
[1, 0, 1, 0, 0],
[1, 0, 1, 0, 1],
[1, 0, 1, 1, 0],
[1, 0, 1, 1, 1],
[1, 1, 0, 0, 0],
[1, 1, 0, 0, 1],
[1, 1, 0, 1, 0],
[1, 1, 0, 1, 1],
[1, 1, 1, 0, 0],
[1, 1, 1, 0, 1],
[1, 1, 1, 1, 0],
[1, 1, 1, 1, 1]
]
Alternatively, you could use a recursive generation function:
def gimme_combs(n):
if n == 1: return [[0], [1]]
lower_combs = gimme_combs(n - 1)
return [[0] + c for c in lower_combs] + \
[[1] + c for c in lower_combs]
which would give the same result when called with:
combs = gimme_combs(5)

Setting indicators based in index per row in numpy

I am looking for an efficient way to set a indicators from zero to a known number (which differs for each row).
e.g.
a =
array([[1, 1, 1, 0, 0, 0, 0, 0, 0],
[1, 0, 0, 0, 0, 0, 0, 0, 0],
[1, 1, 1, 1, 1, 1, 0, 0, 0],
[1, 1, 0, 0, 0, 0, 0, 0, 0],
[1, 1, 1, 1, 1, 1, 1, 1, 0]])
and I know the vector with the index when a goes from 1 to zero.
b = [3, 1, 6, 2, 8]
Rather than filling all the rows of a using a for-loop, I want to know if there is a fast way to set these indicators.
Use outer-comparison on ranged array vs. b -
In [16]: ncols = 9
In [17]: b
Out[17]: [3, 1, 6, 2, 8]
In [19]: np.greater.outer(b,np.arange(ncols)).view('i1')
Out[19]:
array([[1, 1, 1, 0, 0, 0, 0, 0, 0],
[1, 0, 0, 0, 0, 0, 0, 0, 0],
[1, 1, 1, 1, 1, 1, 0, 0, 0],
[1, 1, 0, 0, 0, 0, 0, 0, 0],
[1, 1, 1, 1, 1, 1, 1, 1, 0]], dtype=int8)
Other similar ways to express the same -
(np.asarray(b)[:,None] > np.arange(ncols)).view('i1')
(np.asarray(b)[:,None] > np.arange(ncols)).astype(int)
With b being an array, simplifies further, as we can skip the array conversion with np.asarray(b).
Simplest way I can think of is:
result=[]
for row in array:
result.append(row.tolist().index(0))
print(result)
[3, 1, 6, 2, 8]
The reason this works is, that list has a method called index, which tells the first occurrence of a specific item in the list. So I am iterating over this 2-dimentional array, converting each of it to list and using index of 0 on each.
You can store these values into another list and append to it for each row and that's it.
You can use broadcasting to do an outer comparison:
b = np.asarray([3, 1, 6, 2, 8])
a = (np.arange(b.max() + 1) < b[:, None]).astype(int)
# array([[1, 1, 1, 0, 0, 0, 0, 0, 0],
# [1, 0, 0, 0, 0, 0, 0, 0, 0],
# [1, 1, 1, 1, 1, 1, 0, 0, 0],
# [1, 1, 0, 0, 0, 0, 0, 0, 0],
# [1, 1, 1, 1, 1, 1, 1, 1, 0]])

Numpy trouble vectorizing certain kind of aggregation

I am having difficulty in vectorizing the below operation:
# x.shape = (a,)
# y.shape = (a, b)
# x and y are ordered over a.
# Want to combine x, y into z.shape(num_unique_x, b)
# Below works and illustrates intent but is iterative
z = np.zeros((num_unique_x, b))
for i in range(a):
z[x[i], y[i, :]] += 1
Your use of num_unique_x, and the size of z suggests that this is a case where x and y have repeats, and that some of the z will be larger than 1. In which case we need to use np.add.at. But to set that up I'd have review its documentation, and possibly test some alternatives.
But first a no-repeats case
In [522]: x=np.arange(6)
In [523]: y=np.arange(3)+x[:,None]
In [524]: y
Out[524]:
array([[0, 1, 2],
[1, 2, 3],
[2, 3, 4],
[3, 4, 5],
[4, 5, 6],
[5, 6, 7]])
See why I ask for a diagnostic example. I'm guessing as to possible values. I have to make a z with more than 3 columns.
In [529]: z=np.zeros((6,8),dtype=int)
In [530]: for i in range(6):
...: z[x[i],y[i,:]]+=1
In [531]: z
Out[531]:
array([[1, 1, 1, 0, 0, 0, 0, 0],
[0, 1, 1, 1, 0, 0, 0, 0],
[0, 0, 1, 1, 1, 0, 0, 0],
[0, 0, 0, 1, 1, 1, 0, 0],
[0, 0, 0, 0, 1, 1, 1, 0],
[0, 0, 0, 0, 0, 1, 1, 1]])
The vectorized equivalent
In [532]: z[x[:,None],y]
Out[532]:
array([[1, 1, 1],
[1, 1, 1],
[1, 1, 1],
[1, 1, 1],
[1, 1, 1],
[1, 1, 1]])
In [533]: z[x[:,None],y] += 1
In [534]: z
Out[534]:
array([[2, 2, 2, 0, 0, 0, 0, 0],
[0, 2, 2, 2, 0, 0, 0, 0],
[0, 0, 2, 2, 2, 0, 0, 0],
[0, 0, 0, 2, 2, 2, 0, 0],
[0, 0, 0, 0, 2, 2, 2, 0],
[0, 0, 0, 0, 0, 2, 2, 2]])
The corresponding add.at expression is
In [538]: np.add.at(z,(x[:,None],y),1)
In [539]: z
Out[539]:
array([[3, 3, 3, 0, 0, 0, 0, 0],
[0, 3, 3, 3, 0, 0, 0, 0],
[0, 0, 3, 3, 3, 0, 0, 0],
[0, 0, 0, 3, 3, 3, 0, 0],
[0, 0, 0, 0, 3, 3, 3, 0],
[0, 0, 0, 0, 0, 3, 3, 3]])
So that works for this no-repeats case.
For repeats in x:
In [542]: x1=np.array([0,1,1,2,3,5])
In [543]: z1=np.zeros((6,8),dtype=int)
In [544]: np.add.at(z1,(x1[:,None],y),1)
In [545]: z1
Out[545]:
array([[1, 1, 1, 0, 0, 0, 0, 0],
[0, 1, 2, 2, 1, 0, 0, 0],
[0, 0, 0, 1, 1, 1, 0, 0],
[0, 0, 0, 0, 1, 1, 1, 0],
[0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 1, 1, 1]])
Without add.at we miss the 2s.
In [546]: z2=np.zeros((6,8),dtype=int)
In [547]: z2[x1[:,None],y] += 1
In [548]: z2
Out[548]:
array([[1, 1, 1, 0, 0, 0, 0, 0],
[0, 1, 1, 1, 1, 0, 0, 0],
[0, 0, 0, 1, 1, 1, 0, 0],
[0, 0, 0, 0, 1, 1, 1, 0],
[0, 0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 1, 1, 1]])

How can I use a Matrix as a dataset on PyBran?

I´m using pybrain in order to train a simple neural network in which the input is going to be a 7x5 Matrix.
The following are the inputs:
A = [[0, 0, 1, 0, 0],
[0, 1, 1, 0, 0],
[0, 1, 0, 1, 0],
[0, 1, 0, 1, 0],
[1, 1, 1, 1, 1],
[1, 0, 0, 0, 1],
[1, 0, 0, 0, 1]]
E = [[1, 1, 1, 1, 1],
[1, 0, 0, 0, 0],
[1, 0, 0, 0, 0],
[1, 1, 1, 1, 0],
[1, 0, 0, 0, 0],
[1, 0, 0, 0, 0],
[1, 1, 1, 1, 1]]
I = [[0, 0, 1, 0, 0],
[0, 0, 1, 0, 0],
[0, 0, 1, 0, 0],
[0, 0, 1, 0, 0],
[0, 0, 1, 0, 0],
[0, 0, 1, 0, 0],
[0, 0, 1, 0, 0]]
O = [[1, 1, 1, 1, 0],
[1, 0, 0, 0, 1],
[1, 0, 0, 0, 1],
[1, 0, 0, 0, 1],
[1, 0, 0, 0, 1],
[1, 0, 0, 0, 1],
[1, 1, 1, 1, 0]]
U = [[1, 0, 0, 0, 1],
[1, 0, 0, 0, 1],
[1, 0, 0, 0, 1],
[1, 0, 0, 0, 1],
[1, 0, 0, 0, 1],
[0, 1, 0, 0, 1],
[0, 0, 1, 1, 0]]
I thought writing something like:
ds = SupervisedDataSet(1, 1)
ds.addSample((A), ("A",))
might work, but I´m getting:
ValueError: cannot copy sequence with size 7 to array axis with dimension 1
Is there any way I can give this datasets to pyBrain?
First you have to know that SupervisedDataSet works with list, so you will need to convert the 2D arrays into a list. You can do it with something like this:
def convertToList (matrix):
list = [ y for x in matrix for y in x]
return list
Then you will need to give the new list to the method SupervisedDataSet.
Also if you would like to use that info to make the network you should use some number to identify the letter like A = 1, E = 2, I = 3, O = 4, U = 5. So to do this, the second parameter for SupervisedDataSet should be just a number 1. In this way you are saying something like "For a list with 35 elements use these numbers to identify a single number".
Finally your code should look like this:
ds = SupervisedDataSet(35, 1)
A2 = convertToList(A)
ds.addSample(A2, (1,))
E2 = convertToList(E)
ds.addSample(E2, (2,))
I2 = convertToList(I)
ds.addSample(I2, (3,))
O2 = convertToList(O)
ds.addSample(O2, (4,))
U2 = convertToList(U)
ds.addSample(U2, (5,))
Hope this could help.

Categories

Resources