Develop Reference

Scipy function that can do np.diff() with compressed sparse column matrix

I want to compute discrete difference of identity matrix.
The code below use numpy and scipy.
import numpy as np
from scipy.sparse import identity
from scipy.sparse import csc_matrix
x = identity(4).toarray()
y = csc_matrix(np.diff(x, n=2))
print(y)
I would like to improve performance or memory usage.
Since identity matrix produce many zeros, it would reduce memory usage to perform calculation in compressed sparse column(csc) format. However, np.diff() does not accept csc format, so converting between csc and normal format using csc_matrix would slow it down a bit.
Normal format
x = identity(4).toarray()
print(x)
[[1. 0. 0. 0.]
[0. 1. 0. 0.]
[0. 0. 1. 0.]
[0. 0. 0. 1.]]
csc format
x = identity(4)
print(x)
(0, 0) 1.0
(1, 1) 1.0
(2, 2) 1.0
(3, 3) 1.0
Thanks

Here is my hacky solution to get the sparse matrix as you want.
L - the length of the original identity matrix,
n - the parameter of np.diff.
In your question they are:
L = 4
n = 2
My code produces the same y as your code, but without the conversions between csc and normal formats.
Your code:
from scipy.sparse import identity, csc_matrix
x = identity(L).toarray()
y = csc_matrix(np.diff(x, n=n))
My code:
from scipy.linalg import pascal
def get_data(n, L):
nums = pascal(n + 1, kind='lower')[-1].astype(float)
minuses_from = n % 2 + 1
nums[minuses_from : : 2] *= -1
return np.tile(nums, L - n)
data = get_data(n, L)
row_ind = (np.arange(n + 1) + np.arange(L - n).reshape(-1, 1)).flatten()
col_ind = np.repeat(np.arange(L - n), n + 1)
y = csc_matrix((data, (row_ind, col_ind)), shape=(L, L - n))
I have noticed that after applying np.diff to the identity matrix n times, the values of the columns are the binomial coefficients with their signs alternating. This is my variable data.
Then I am just constructing the csc_matrix.

Unfortunately, it does not seem that SciPy provides any tools for this kind of sparse matrix manipulation. Regardless, by cleverly manipulating the indices and data of the entries one can emulate np.diff(x,n) in a straightforward fashion.
Given 2D NumPy array (matrix) of dimension MxN, np.diff() multiplies each column (of column index y) with -1 and adds the next column to it (column index y+1). Difference of order k is just the iterative application of k differences of order 1. A difference of order 0 is just the returns the input matrix.
The method below makes use of this, iterateively eliminating duplicate entries by addition through sum_duplicates(), reducing the number of columns by one, and filtering non-valid indices.
def csc_diff(x, n):
'''Emulates np.diff(x,n) for a sparse matrix by iteratively taking difference of order 1'''
assert isinstance(x, csc_matrix) or (isinstance(x, np.ndarray) & len(x.shape) == 2), "Input matrix must be a 2D np.ndarray or csc_matrix."
assert isinstance(n, int) & n >= 0, "Integer n must be larger or equal to 0."
if n >= x.shape[1]:
return csc_matrix(([], ([], [])), shape=(x.shape[0], 0))
if isinstance(x, np.ndarray):
x = csc_matrix(x)
# set-up of data/indices via column-wise difference
if(n > 0):
for k in range(1,n+1):
# extract data/indices of non-zero entries of (current) sparse matrix
M, N = x.shape
idx, idy = x.nonzero()
dat = x.data
# difference: this row (y) * (-1) + next row (y+1)
idx = np.concatenate((idx, idx))
idy = np.concatenate((idy, idy-1))
dat = np.concatenate(((-1)*dat, dat))
# filter valid indices
validInd = (0<=idy) & (idy<N-1)
# x_diff: csc_matrix emulating np.diff(x,1)'s output'
x_diff = csc_matrix((dat[validInd], (idx[validInd], idy[validInd])), shape=(M, N-1))
x_diff.sum_duplicates()
x = x_diff
return x
Moreover, the method outputs an empty csc_matrix of dimension Mx0 when the difference order is larger or equal to the number of columns of the input matrix. This is why the output is identical, see
csc_diff(x, 2).toarray()
> array([[ 1., 0.],
[-2., 1.],
[ 1., -2.],
[ 0., 1.]])
which is identical to
np.diff(x.toarray(), 2)
> array([[ 1., 0.],
[-2., 1.],
[ 1., -2.],
[ 0., 1.]])
This identity holds for other difference orders, too
(csc_diff(x, 0).toarray() == np.diff(x.toarray(), 0)).all()
>True
(csc_diff(x, 3).toarray() == np.diff(x.toarray(), 3)).all()
>True
(csc_diff(x, 13).toarray() == np.diff(x.toarray(), 13)).all()
>True

How to set the measurement matrix of opencv kalman filter depending on the measurement dimensions [OpenCV+Python]

I am working on a tracking application where I use the opencv kalman filter to validate my current measurement of the position. I use the code from this question:
At first I calculate velocity (v) and accelearation (a) of my moving object at (x, y). These 4 values are used as my kalman state. I initiate the kalman filter as follows:
(np.eye(n,m) generates the identity matrix with dimensions nxm):
def initKalman(init_state, fps):
kalman = cv.KalmanFilter(4, 4, 2)
kalman.transitionMatrix = np.array([[1., 0., 1/fps, 0.],
[0., 1., 0., 1/fps],
[0., 0., 1., 0.],
[0, 0., 0., 1.]])
kalman.measurementMatrix = 1. * np.eye(2, 4)
kalman.measurementNoiseCov = 1e-3 * np.eye(2, 2)
kalman.processNoiseCov = 1e-5 * np.eye(4, 4)
kalman.errorCovPost = 1e-1 * np.eye(4, 4)
kalman.statePost = init_state.reshape(4, 1)
return kalman
kinematics = np.array((velocity, acceleration), dtype=np.float32)
kalman_state = np.concatenate((point, kinematics))
kalman_filter = initKalman(kalman_state, fps = 15)
During operation the correction is done as follows:
def correct_kalman(kalman, state):
measurement = (np.dot(kalman.measurementNoiseCov, np.random.randn(2, 1))).reshape(-1)
measurement = np.dot(kalman.measurementMatrix, state) + measurement
return kalman.correct(measurement)
kinematics = np.array((velocity, acceleration), dtype=np.float32)
kalman_state = np.concatenate((point, kinematics))
correct_kalman(kalman_filter, kalman_state)
It seems to work witch is great, but im trying to understand why. In my understanding it shouldn't work because in correct_kalman() the velocity and acceleration are ommited in this code line:
measurement = np.dot(kalman.measurementMatrix, state) + measurement
because the measurementmatrix is just 2 x 4. (In fact, if I set acceleration and speed to 0, the behavior of the filter does not change.)
For Example take the kalman_state = np.array([10., 20., 25., 75.]) and calculate the dot product with the measurementMatrix = 1. * np.eye(2, 4)
then measurement = np.dot(kalman.measurementMatrix, kalman_state) is just
>>> measurement
array([10., 20.])
v and a are gone.
So I changed my measurementMatrix and my measurementNoiseCov to 4 x 4 dimensionality and adjusted my correction acordingly by using np.random.randn(4, 1) but now the kalman filter is way to sluggish and falls behind the measurement.
Why is the first approach working if v and a are not used?
How can I change the measurement matrix in a more targeted way than just iteratively adjusting the values?
Thanks for the help!

Weighted sum of an input inside the network

I have a network with multiple inputs and I split out the first 10 inputs and calculate the weighted sum, and then concatenate it with the rest of the input:
first = Lambda(lambda z: z[:, 0:11])(d_inputs)
wsum_first = Lambda(calcWSumF)(first )
d_input = concatenate([d_inputs, wsum_first], axis=-1)
with the function defined as:
w_vec = K.constant(np.array([range(10)]*64).reshape(10, 64)) # batch size is 64
def calcWSumF(x):
y = K.dot(w_vec, x)
y = K.expand_dims(y, -1)
return y
I want a constant vector to be used to calculate the weighted sum of the first part of the input. The concatenation doesn't work because the shapes don't match. How can I implement this correctly?

You can write this much better using K.sum and only a vector containing the coefficients. Further, there is no need to use a fixed batch size (it can be any number):
def calcWSumF(x, idx):
w_vec = K.constant(np.arange(idx))
y = K.sum(x[:, 0:idx] * w_vec, axis=-1, keepdims=True)
return y
d_inputs = Input((15,))
wsum_first = Lambda(calcWSumF, arguments={'idx': 10})(d_inputs)
d_input = concatenate([d_inputs, wsum_first], axis=-1)
model = Model(d_inputs, d_input)
model.predict(np.arange(15).reshape(1, 15))
# output:
array([[ 0., 1., 2., 3., 4., 5., 6., 7., 8., 9., 10.,
11., 12., 13., 14., 285.]], dtype=float32)
# Note: 0*0 + 1*1 + 2*2 + ... + 9*9 = 285
Note that, to make it more general, we have added another argument (idx) to the lambda function which specifies how many of the elements from the beginning we would like to consider.

Implementing gradient operator in Python

I'm working on a Computer Vision system and this is giving me a serious headache. I'm having trouble re-implementing an old gradient operator more efficiently, I'm working with numpy and openCV2.
This is what I had:
def gradientX(img):
rows, cols = img.shape
out = np.zeros((rows,cols))
for y in range(rows-1):
Mr = img[y]
Or = out[y]
Or[0] = Mr[1] - Mr[0]
for x in xrange(1, cols - 2):
Or[x] = (Mr[x+1] - Mr[x-1])/2.0
Or[cols-1] = Mr[cols-1] - Mr[cols-2]
return out
def gradient(img):
return [gradientX(img), (gradientX(img.T).T)]
I've tried using numpy's gradient operator but the result is not the same
For this input
array([[ 3, 4, 5],
[255, 0, 12],
[ 25, 15, 200]])
Using my gradient returns
[array([[ 1., 0., 1.],
[-255., 0., 12.],
[ 0., 0., 0.]]),
array([[ 252., -4., 0.],
[ 0., 0., 0.],
[-230., 15., 0.]])]
While using numpy's np.gradient returns
[array([[ 252. , -4. , 7. ],
[ 11. , 5.5, 97.5],
[-230. , 15. , 188. ]]),
array([[ 1. , 1. , 1. ],
[-255. , -121.5, 12. ],
[ -10. , 87.5, 185. ]])]
There are cleary some similarities between the results but they're definitely not the same. So I'm missing something here or the two operators aren't mean to produce the same results. In that case, I wanted to know how to re-implement my gradientX function so it doesn't use that awful looking double loop for traversing the 2-d array using mostly numpy's potency.

I've been working a bit more on this just to find that my mistake.
I was skipping last row and last column when iterating. As #wflynny noted, the result was identical except for a row and a column of zeros.
Provided this, the result could not be the same as np.gradient, but with that change, the results are identical, so there's no need to find any other numpy implementation for this.
Answering my own question, a good numpy's implementation for my gradient algorithm would be
import numpy as np
def gradientX(img):
return np.gradient(img)[::-1]
I'm also posting the working code, just because it shows how numpy's gradient operator works
def computeMatXGradient(img):
rows, cols = img.shape
out = np.zeros((rows,cols))
for y in range(rows):
Mr = img[y]
Or = out[y]
Or[0] = float(Mr[1]) - float(Mr[0])
for x in xrange(1, cols - 1):
Or[x] = (float(Mr[x+1]) - float(Mr[x-1]))/2.0
Or[cols-1] = float(Mr[cols-1]) - float(Mr[cols-2])
return out

Python get get average of neighbours in matrix with na value

I have very large matrix, so dont want to sum by going through each row and column.
a = [[1,2,3],[3,4,5],[5,6,7]]
def neighbors(i,j,a):
return [a[i][j-1], a[i][(j+1)%len(a[0])], a[i-1][j], a[(i+1)%len(a)][j]]
[[np.mean(neighbors(i,j,a)) for j in range(len(a[0]))] for i in range(len(a))]
This code works well for 3x3 or small range of matrix, but for large matrix like 2k x 2k this is not feasible. Also this does not work if any of the value in matrix is missing or it's like na
This code works well for 3x3 or small range of matrix, but for large matrix like 2k x 2k this is not feasible. Also this does not work if any of the value in matrix is missing or it's like na. If any of the neighbor values is na then skip that neighbour in getting the average

Shot #1
This assumes you are looking to get sliding windowed average values in an input array with a window of 3 x 3 and considering only the north-west-east-south neighborhood elements.
For such a case, signal.convolve2d with an appropriate kernel could be used. At the end, you need to divide those summations by the number of ones in kernel, i.e. kernel.sum() as only those contributed to the summations. Here's the implementation -
import numpy as np
from scipy import signal
# Inputs
a = [[1,2,3],[3,4,5],[5,6,7],[4,8,9]]
# Convert to numpy array
arr = np.asarray(a,float)
# Define kernel for convolution
kernel = np.array([[0,1,0],
[1,0,1],
[0,1,0]])
# Perform 2D convolution with input data and kernel
out = signal.convolve2d(arr, kernel, boundary='wrap', mode='same')/kernel.sum()
Shot #2
This makes the same assumptions as in shot #1, except that we are looking to find average values in a neighborhood of only zero elements with the intention to replace them with those average values.
Approach #1: Here's one way to do it using a manual selective convolution approach -
import numpy as np
# Convert to numpy array
arr = np.asarray(a,float)
# Pad around the input array to take care of boundary conditions
arr_pad = np.lib.pad(arr, (1,1), 'wrap')
R,C = np.where(arr==0) # Row, column indices for zero elements in input array
N = arr_pad.shape[1] # Number of rows in input array
offset = np.array([-N, -1, 1, N])
idx = np.ravel_multi_index((R+1,C+1),arr_pad.shape)[:,None] + offset
arr_out = arr.copy()
arr_out[R,C] = arr_pad.ravel()[idx].sum(1)/4
Sample input, output -
In [587]: arr
Out[587]:
array([[ 4., 0., 3., 3., 3., 1., 3.],
[ 2., 4., 0., 0., 4., 2., 1.],
[ 0., 1., 1., 0., 1., 4., 3.],
[ 0., 3., 0., 2., 3., 0., 1.]])
In [588]: arr_out
Out[588]:
array([[ 4. , 3.5 , 3. , 3. , 3. , 1. , 3. ],
[ 2. , 4. , 2. , 1.75, 4. , 2. , 1. ],
[ 1.5 , 1. , 1. , 1. , 1. , 4. , 3. ],
[ 2. , 3. , 2.25, 2. , 3. , 2.25, 1. ]])
To take care of the boundary conditions, there are other options for padding. Look at numpy.pad for more info.
Approach #2: This would be a modified version of convolution based approach listed earlier in Shot #1. This is same as that earlier approach, except that at the end, we selectively replace
the zero elements with the convolution output. Here's the code -
import numpy as np
from scipy import signal
# Inputs
a = [[1,2,3],[3,4,5],[5,6,7],[4,8,9]]
# Convert to numpy array
arr = np.asarray(a,float)
# Define kernel for convolution
kernel = np.array([[0,1,0],
[1,0,1],
[0,1,0]])
# Perform 2D convolution with input data and kernel
conv_out = signal.convolve2d(arr, kernel, boundary='wrap', mode='same')/kernel.sum()
# Initialize output array as a copy of input array
arr_out = arr.copy()
# Setup a mask of zero elements in input array and
# replace those in output array with the convolution output
mask = arr==0
arr_out[mask] = conv_out[mask]
Remarks: Approach #1 would be the preferred way when you have fewer number of zero elements in input array, otherwise go with Approach #2.

This is an appendix to comments under #Divakar's answer (rather than an independent answer).
Out of curiosity I tried different 'pseudo' convolutions against the scipy convolution. The fastest one was the % (modulus) wrapping one, which surprised me: obviously numpy does something clever with its indexing, though obviously not having to pad will save time.
fn3 -> 9.5ms, fn1 -> 21ms, fn2 -> 232ms
import timeit
setup = """
import numpy as np
from scipy import signal
N = 1000
M = 750
P = 5 # i.e. small number -> bigger proportion of zeros
a = np.random.randint(0, P, M * N).reshape(M, N)
arr = np.asarray(a,float)"""
fn1 = """
arr_pad = np.lib.pad(arr, (1,1), 'wrap')
R,C = np.where(arr==0)
N = arr_pad.shape[1]
offset = np.array([-N, -1, 1, N])
idx = np.ravel_multi_index((R+1,C+1),arr_pad.shape)[:,None] + offset
arr[R,C] = arr_pad.ravel()[idx].sum(1)/4"""
fn2 = """
kernel = np.array([[0,1,0],
[1,0,1],
[0,1,0]])
conv_out = signal.convolve2d(arr, kernel, boundary='wrap', mode='same')/kernel.sum()
mask = arr == 0.0
arr[mask] = conv_out[mask]"""
fn3 = """
R,C = np.where(arr == 0.0)
arr[R, C] = (arr[(R-1)%M,C] + arr[R,(C-1)%N] + arr[R,(C+1)%N] + arr[(R+1)%M,C]) / 4.0
"""
print(timeit.timeit(fn1, setup, number = 100))
print(timeit.timeit(fn2, setup, number = 100))
print(timeit.timeit(fn3, setup, number = 100))

Using numpy and scipy.ndimage, you can apply a "footprint" that defines where you look for the neighbours of each element and apply a function to those neighbours:
import numpy as np
import scipy.ndimage as ndimage
# Getting neighbours horizontally and vertically,
# not diagonally
footprint = np.array([[0,1,0],
[1,0,1],
[0,1,0]])
a = [[1,2,3],[3,4,5],[5,6,7]]
# Need to make sure that dtype is float or the
# mean won't be calculated correctly
a_array = np.array(a, dtype=float)
# Can specify that you want neighbour selection to
# wrap around at the borders
ndimage.generic_filter(a_array, np.mean,
footprint=footprint, mode='wrap')
Out[36]:
array([[ 3.25, 3.5 , 3.75],
[ 3.75, 4. , 4.25],
[ 4.25, 4.5 , 4.75]])