Objective
I have a soup of triangle polygons. I want to retrieve the largest median as vector for each triangle.
State of work
Starting point:
Array of points (n,3) , e.g. [x,y,z]
Array of triangle point indices (n, 3) referencing the array of points above, e.g. [[0,1,2],[2,3,4]...]
I combine both two one single matrix containing the real 3D point coordinates. Then I calculate the median vectors and their lengths.
/Edit : I updated the code to my current version of it
def calcMedians(polygon):
# C -> AB = C-(A + 0.5(B-A))
# B -> AC = B - (A + 0.5(C-A))
# A -> BC = A - (B
dim = np.shape(polygon)
medians = np.zeros((dim[0],3,2,dim[1]))
medians[:,0,0] = polygon[:,2]
medians[:,0,1] = polygon[:,0] + 0.5*(polygon[:,1]-polygon[:,0])
medians[:,1,0] = polygon[:,1]
medians[:,1,1] = polygon[:,0] + 0.5*(polygon[:,2]-polygon[:,0])
medians[:,2,0] = polygon[:,0]
medians[:,2,1] = polygon[:,1] + 0.5*(polygon[:,2]-polygon[:,1])
m1 = np.linalg.norm(medians[:,0,0]-medians[:,0,1],axis=1)
m2 = np.linalg.norm(medians[:,1,0]-medians[:,1,1],axis=1)
m3 = np.linalg.norm(medians[:,2,0]-medians[:,2,1],axis=1)
medianlengths = np.vstack((m1,m2,m3)).T
maxlengths = np.argmax(medianlengths,axis=1)
final = np.zeros((dim[0],2,dim[1]))
dim = np.shape(medians)
for i in range(0,dim[0]):
idx = maxlengths[i]
final[i] = medians[i,idx]
return final
Now I am creating the final median vector matrix using an empty matrix first. The lengths are calculated using np.linalg.norm and collected in a matrix. For this matrix, the argmax method is used to identify to target median vector.
Problem
Old:However, I am somehow confused by the dimensionality and currently not able to get this to work or to understand if the result is correct.
Does somebody know how to do this correctly and/or if this approach is efficient?
My target would be a construct of the 3 medians in form of [n_polygons, 3( due to 3 medians), 2 (start and end point), 3 (xyz)]
Using the max lengths information, i would like to reduce it to [n_polygons, 2 (start and end point), 3 (xyz)]
Using this improvised for loop in the function, I can create the output. But there has to be a more "clean" matrix method to it. Using medians[:,maxlengths,:,:] leads to a shape of [4,n_polygons,2,3] instead of [n_polygons,2,3] and I do not understand why.
Example image for medians of two triangles:
Unfortunately, I don't have a large exemplary data set but I guess that this can be generated quite quickly. The example data set from the picture shown above is:
polygons = np.array([[0,1,2],[0,3,2]])
points = np.array([[0,0],
[1,0],
[1,1],
[0,1]])
polygons3d = points[polygons[:,:]]
The longest median is for the shortest triangle side. Look here and rewrite median length formula as
M[i] = Sqrt(2(a^2+b^2+c^2)-3*side[i]^2) / 2
So you can simplify calculations a bit using only side lengths (perhaps you already have them)
Concerning 3D coordinates - just use projection on any coordinate plane not perpendicular to your point plane - ignore one dimension (choose dimension with the lowest value range)
Related
I have two two-dimensional arrays loaded into NumPy, both of which are 80i x 80j in size. I'm looking to do a moving window polyfit calculation across these arrays, I've nailed down how to conduct the polyfit but am stuck on the specific moving window approach I'm looking to accomplish. I'm aiming for:
1) At each index of Array 1 (a1), code isolates all the values of the index & its closest 8 neighbors into a separate 1D array, and repeats over the same window at Array 2 (a2)
2) With these two new arrays isolated, perform linear regression line-fitting using NumPy's polyfit in the following approach: model = np.polyfit(a1slice, a2slice, 1)
3) Cast the resulting regression coefficient and intercept (example output doing this manually: array([-0.02114911, 10.02127152])) to the same index of two other arrays, where model[0] would be placed into the first new array and model[1] into the second new array at this index.
4) The code then moves sequentially to the next index and performs steps 1-3 again, or a1(i+1, j+0, etc.)
I've provided a graphical example of what I'm trying achieve for two random index selections across Array 1 and the calculation across the index's eight nearest neighbors, should this make the desired result easier to understand:
I've written a function to get the window, but there is a troublesome edge case when the index is on the edge and one distance away from the corner. You'll probably want to modify the function to get exactly what you're looking for.
def get_neighbors(arr, origin, num_neighbors = 8):
coords = np.array([[i,j] for (i,j),value in np.ndenumerate(arr)]).reshape(arr.shape + (2,))
distances = np.linalg.norm(coords - origin, axis = -1)
neighbor_limit = np.sort(distances.ravel())[num_neighbors]
window = np.where(distances <= neighbor_limit)
exclude_window = np.where(distances > neighbor_limit)
return window, exclude_window, distances
test = np.zeros((5, 5))
plt.close('all')
cases = [[0,0], [0, 1], [0, 2]]
for case in cases:
window, exclude, distances = get_neighbors(test, case)
distances[exclude] = 0
plt.figure()
plt.imshow(distances)
Image Outputs:
I have several dataframes which each contain two columns of x and y values, so each row represents a point on a curve. The different dataframes then represent contours on a map. I have another series of data points (fewer in number), and I'd like to see which contour they are closest to on average.
I would like to establish the distance from each datapoint to each point on the curve, with sqrt(x^2+y^2) - sqrt(x_1^2 + y_1^2), add them up for each point on the curve. The trouble is that there are several thousand points on the curve, and there are only a few dozen datapoints to assess, so I can't simply put these in columns next to each other.
I think I need to cycle through the datapoints, checking the sqdistance between them and each point in the curve.
I don't know whether there is an easy function or module that can do this.
Thanks in advance!
Edit: Thanks for the comments. #Alexander: I've tried the vectorize function, as follows, with a sample dataset. I'm actually using contours which comprise several thousand datapoints, and the dataset to compare against are 100+, so I'd like to be able to automate as much as possible. I'm currently able to create a distance measurement from the first datapoint against my contour, but I would ideally like to cycle through j as well. When I try it, it comes up with an error:
import numpy as np
from numpy import vectorize
import pandas as pd
from pandas import DataFrame
df1 = {'X1':['1', '2', '2', '3'], 'Y1':['2', '5', '7', '9']}
df1 = DataFrame(df1, columns=['X1', 'Y1'])
df2 = {'X2':['3', '5', '6'], 'Y2':['10', '15', '16']}
df2 = DataFrame(df2, columns=['X2', 'Y2'])
df1=df1.astype(float)
df2=df2.astype(float)
Distance=pd.DataFrame()
i = range(0, len(df1))
j = range(0, len(df2))
def myfunc(x1, y1, x2, y2):
return np.sqrt((x2-x1)**2+np.sqrt(y2-y1)**2)
vfunc=np.vectorize(myfunc)
Distance['Distance of Datapoint j to Contour']=vfunc(df1.iloc[i] ['X1'], df1.iloc[i]['Y1'], df2.iloc[0]['X2'], df2.iloc[0]['Y2'])
Distance['Distance of Datapoint j to Contour']=vfunc(df1.iloc[i] ['X1'], df1.iloc[i]['Y1'], df2.iloc[1]['X2'], df2.iloc[1]['Y2'])
Distance
General idea
The "curve" is actually a polygon with a lot's of points. There definetly some libraries to calculate the distance between the polygon and the point. But generally it will be something like:
Calculate "approximate distance" to whole polygon, e.g. to the bounding box of a polygon (from point to 4 line segments), or to the center of bounding box
calculate distances to the lines of a polygon. If you have too many points then as an extra step "resolution" of a polygon might be reduced.
Smallest found distance is the distance from point to the polygon.
repeat for each point and each polygon
Existing solutions
Some libraries already can do that:
shapely question, shapely Geo-Python docs
Using shapely in geopandas to calculate distance
scipy.spatial.distance: scipy can be used to calculate distance between arbitrary number of points
numpy.linalg.norm(point1-point2): some answers propose different ways to calculate distance using numpy. Some even show performance benchmarks
sklearn.neighbors: not really about curves and distances to them, but can be used if you want to check "to which area point is most likely related"
And you can always calculate distances yourself using D(x1, y1, x2, y2) = sqrt((x₂-x₁)² + (y₂-y₁)²) and search for best combination of points that gives minimal distance
Example:
# get distance from points of 1 dataset to all the points of another dataset
from scipy.spatial import distance
d = distance.cdist(df1.to_numpy(), df2.to_numpy(), 'euclidean')
print(d)
# Results will be a matrix of all possible distances:
# [[ D(Point_df1_0, Point_df2_0), D(Point_df1_0, Point_df2_1), D(Point_df1_0, Point_df2_2)]
# [ D(Point_df1_1, Point_df2_0), D(Point_df1_1, Point_df2_1), D(Point_df1_1, Point_df2_2)]
# [ D(Point_df1_3, Point_df2_0), D(Point_df1_2, Point_df2_1), D(Point_df1_2, Point_df2_2)]
# [ D(Point_df1_3, Point_df2_0), D(Point_df1_3, Point_df2_1), D(Point_df1_3, Point_df2_2)]]
[[ 8.24621125 13.60147051 14.86606875]
[ 5.09901951 10.44030651 11.70469991]
[ 3.16227766 8.54400375 9.8488578 ]
[ 1. 6.32455532 7.61577311]]
What to do next is up to you. For example as a metric of "general distance between curves" you can:
Pick smallest values in each row and each column (if you skip some columns/rows, then you might end up with candidate that "matches only a part of contour), and calculate their median: np.median(np.hstack([np.amin(d, axis) for axis in range(len(d.shape))])).
Or you can calculate mean value of:
all the distances: np.median(d)
of "smallest 2/3 of distances": np.median(d[d<np.percentile(d, 66, interpolation='higher')])
of "smallest distances that cover at least each rows and each columns":
for min_value in np.sort(d, None):
chosen_indices = d<=min_value
if np.all(np.hstack([np.amax(chosen_indices, axis) for axis in range(len(chosen_indices.shape))])):
break
similarity = np.median(d[chosen_indices])
Or maybe you can use different type of distance from the begining (e.g. "correlation distance" looks promising to your task)
Maybe use "Procrustes analysis, a similarity test for two data sets" together with distances.
Maybe you can use minkowski distance as a similarity metric.
Alternative approach
Alternative approach would be to use some "geometry" library to compare areas of concave hulls:
Build concave hulls for contours and for "candidate datapoints" (not easy, but possible: using shapely , using concaveman). But if you are sure that your contours are already ordered and without overlapping segments, then you can directly build polygons from those points without need for concave hull.
Use "intersection area" minus "non-common area" as a metric of similarity (shapely can be used for that):
Non-common area is: union - intersection or simply "symmetric difference"
Final metric: intersection.area - symmetric_difference.area (intersection, area)
This approach might be better than processing distances in some situations, for example:
You want to prefer "fewer points covering whole area" over "huge amount of very close points that cover only half of the area"
It's more obvious way to compare candidates with different number of points
But it has it's disadvantages too (just draw some examples on paper and experiment to find them)
Other ideas:
instead of using polygons or concave hull you can:
build a linear ring from your points and then use contour.buffer(some_distance). This way you ignore "internal area" of the contour and only compare contour itself (with tolerance of some_distance). Distance between centroids (or double of that) may be used as value for some_distance
You can build polygons/lines from segments using ops.polygonize
instead of using intersection.area - symmetric_difference.area you can:
Snap one object to another, and then compare snapped object to original
Before comparing real objects you can compare "simpler" versions of the objects to filter out obvious mismatches:
For example you can check if boundaries of objects intersect
Or you can simplify geometries before comparing them
For the distance, you need to change your formula to
def getDistance(x, y, x_i, y_i):
return sqrt((x_i -x)^2 + (y_i - y)^2)
with (x,y) being your datapoint and (x_i, y_i) being a point from the curve.
Consider using NumPy for vectorization. Explicitly looping through your data points will most likely be less efficient, depending on your use case, it might however be quick enough. (If you need to run it on a regular basis, I think vectorization will easily outspeed the explicit way) This could look something like this:
import numpy as np # Universal abbreviation for the module
datapoints = np.random.rand(3,2) # Returns a vector with randomized entries of size 3x2 (Imagine it as 3 sets of x- and y-values
contour1 = np.random.rand(1000, 2) # Other than the size (which is 1000x2) no different than datapoints
contour2 = np.random.rand(1000, 2)
contour3 = np.random.rand(1000, 2)
def squareDistanceUnvectorized(datapoint, contour):
retVal = 0.
print("Using datapoint with values x:{}, y:{}".format(datapoint[0], datapoint[1]))
lengthOfContour = np.size(contour, 0) # This gets you the number of lines in the vector
for pointID in range(lengthOfContour):
squaredXDiff = np.square(contour[pointID,0] - datapoint[0])
squaredYDiff = np.square(contour[pointID,1] - datapoint[1])
retVal += np.sqrt(squaredXDiff + squaredYDiff)
retVal = retVal / lengthOfContour # As we want the average, we are dividing the sum by the element count
return retVal
if __name__ == "__main__":
noOfDatapoints = np.size(datapoints,0)
contID = 0
for currentDPID in range(noOfDatapoints):
dist1 = squareDistanceUnvectorized(datapoints[currentDPID,:], contour1)
dist2 = squareDistanceUnvectorized(datapoints[currentDPID,:], contour2)
dist3 = squareDistanceUnvectorized(datapoints[currentDPID,:], contour3)
if dist1 > dist2 and dist1 > dist3:
contID = 1
elif dist2 > dist1 and dist2 > dist3:
contID = 2
elif dist3 > dist1 and dist3 > dist2:
contID = 3
else:
contID = 0
if contID == 0:
print("Datapoint {} is inbetween two contours".format(currentDPID))
else:
print("Datapoint {} is closest to contour {}".format(currentDPID, contID))
Okay, now moving on to vector-land.
I have taken the liberty to adjust this part to what I think is your dataset. Try it and let me know if it works.
import numpy as np
import pandas as pd
# Generate 1000 points (2-dim Vector) with random values between 0 and 1. Make them strings afterwards.
# This is the first contour
random2Ddata1 = np.random.rand(1000,2)
listOfX1 = [str(x) for x in random2Ddata1[:,0]]
listOfY1 = [str(y) for y in random2Ddata1[:,1]]
# Do the same for a second contour, except that we de-center this 255 units into the first dimension
random2Ddata2 = np.random.rand(1000,2)+[255,0]
listOfX2 = [str(x) for x in random2Ddata2[:,0]]
listOfY2 = [str(y) for y in random2Ddata2[:,1]]
# After this step, our 'contours' are basically two blobs of datapoints whose centers are approx. 255 units apart.
# Generate a set of 4 datapoints and make them a Pandas-DataFrame
datapoints = {'X': ['0.5', '0', '255.5', '0'], 'Y': ['0.5', '0', '0.5', '-254.5']}
datapoints = pd.DataFrame(datapoints, columns=['X', 'Y'])
# Do the same for the two contours
contour1 = {'Xf': listOfX1, 'Yf': listOfY1}
contour1 = pd.DataFrame(contour1, columns=['Xf', 'Yf'])
contour2 = {'Xf': listOfX2, 'Yf': listOfY2}
contour2 = pd.DataFrame(contour2, columns=['Xf', 'Yf'])
# We do now have 4 datapoints.
# - The first datapoint is basically where we expect the mean of the first contour to be.
# Contour 1 consists of 1000 points with x, y- values between 0 and 1
# - The second datapoint is at the origin. Its distances should be similar to the once of the first datapoint
# - The third datapoint would be the result of shifting the first datapoint 255 units into the positive first dimension
# - The fourth datapoint would be the result of shifting the first datapoint 255 units into the negative second dimension
# Transformation into numpy array
# First the x and y values of the data points
dpArray = ((datapoints.values).T).astype(np.float)
c1Array = ((contour1.values).T).astype(np.float)
c2Array = ((contour2.values).T).astype(np.float)
# This did the following:
# - Transform the datapoints and contours into numpy arrays
# - Transpose them afterwards so that if we want all x values, we can write var[0,:] instead of var[:,0].
# A personal preference, maybe
# - Convert all the values into floats.
# Now, we iterate through the contours. If you have a lot of them, putting them into a list beforehand would do the job
for contourid, contour in enumerate([c1Array, c2Array]):
# Now for the datapoints
for _index, _value in enumerate(dpArray[0,:]):
# The next two lines do vectorization magic.
# First, we square the difference between one dpArray entry and the contour x values.
# You might notice that contour[0,:] returns an 1x1000 vector while dpArray[0,_index] is an 1x1 float value.
# This works because dpArray[0,_index] is broadcasted to fit the size of contour[0,:].
dx = np.square(dpArray[0,_index] - contour[0,:])
# The same happens for dpArray[1,_index] and contour[1,:]
dy = np.square(dpArray[1,_index] - contour[1,:])
# Now, we take (for one datapoint and one contour) the mean value and print it.
# You could write it into an array or do basically anything with it that you can imagine
distance = np.mean(np.sqrt(dx+dy))
print("Mean distance between contour {} and datapoint {}: {}".format(contourid+1, _index+1, distance))
# But you want to be able to call this... so here we go, generating a function out of it!
def getDistanceFromDatapointsToListOfContoursFindBetterName(datapoints, listOfContourDataFrames):
""" Takes a DataFrame with points and a list of different contours to return the average distance for each combination"""
dpArray = ((datapoints.values).T).astype(np.float)
listOfContours = []
for item in listOfContourDataFrames:
listOfContours.append(((item.values).T).astype(np.float))
retVal = np.zeros((np.size(dpArray,1), len(listOfContours)))
for contourid, contour in enumerate(listOfContours):
for _index, _value in enumerate(dpArray[0,:]):
dx = np.square(dpArray[0,_index] - contour[0,:])
dy = np.square(dpArray[1,_index] - contour[1,:])
distance = np.mean(np.sqrt(dx+dy))
print("Mean distance between contour {} and datapoint {}: {}".format(contourid+1, _index+1, distance))
retVal[_index, contourid] = distance
return retVal
# And just to see that it is, indeed, returning the same results, run it once
getDistanceFromDatapointsToListOfContoursFindBetterName(datapoints, [contour1, contour2])
I want to calculate the distance to every other position in an integer lattice from it's centre and the number of positions at each distance. I'm currently using the following code to calculate this:
x = numpy.arange(-10, 11, 1)
[X, Y] = numpy.meshgrid(x, x)
R = numpy.sqrt(X**2+Y**2)
R2 = numpy.ndarray.flatten(R)
R3 = numpy.unique(R2)
r = R3[1:] # excludes the 0
Nr = numpy.zeros(numpy.size(r))
for i in range(numpy.size(r)):
Nr[i] = numpy.count_nonzero(R2 == r[i]
This tells me that the possible distances are 1, sqrt2, 2, sqrt5 etc.
It also tells me their is 4x1, 4xsqrt2, 4x2, 8xsqrt5 etc.
As this is a common problem in physics I was wondering if there is a function from a library such as numpy or scipy which could return these values more easily.
The lattice is centered at (0,0). So it's symmetric on the four quadrants. So, we can use this restriction to our advantage as we could compute those required unique distances and counts for one quadrant and multiply those counts by 4 to simulate for all four quadrants.
So, let's say we use the first quadrant (upper right quad). We would skip the elements on the (y = 0) line, because otherwise with the multiplication by 4 for simulating on all four quadrants would result in duplicating results. Additionally, this way we won't have to exclude the first element, as done in the original post.
Thus, an implementation would be -
N = 11 # Lattice size
xa, ya = np.ogrid[0:N,1:N] # x's:0:N, y's:1:N
unq_dists, count = np.unique(np.sqrt(xa**2 + ya**2), return_counts=1)
count = count*4
For further performance boost, we could use np.unique on the squared summations and then use np.sqrt on the unique ones. The idea is to perform the slow square-root computation on smaller unique set, like so -
unq_dists, count = np.unique(xa**2 + ya**2, return_counts=1)
unq_dists = np.sqrt(unq_dists)
After you flattened the array, convert it to a Pandas Series and count unique values:
distances = pandas.Series(numpy.ndarray.flatten(R))
distances.value_counts()
# 9.219544 16
# 8.062258 16
# 5.000000 12
#10.000000 12
# 7.071068 12
# 2.236068 8
# ....
Your code can be made more efficient in a few ways. You are first forming two large matrices and then square the terms. It would be better to square first. Also, forming meshgrid only to perform outer addition is unnecessary: there is numpy.add.outer. Lastly, the loop you have is made unnecessary by the return_counts=True option in numpy.unique. (Which, incidentally, flattens the array itself, so you don't have to.) So the code is shortened to three lines.
x = numpy.arange(-10, 11, 1)
R = numpy.sqrt(numpy.add.outer(x**2, x**2))
r, Nr = numpy.unique(R, return_counts=True)
(And if you want to exclude 0 distance, return r[1:] and Nr[1:])
I'm relatively new to Python coding (I'm switching from R mostly due to running time speed) and I'm trying to figure out how to code a proximity graph.
That is suppose i have an array of evenly-spaced points in d-dimensional Euclidean space, these will be my nodes. I want to make these into an undirected graph by connecting two points if and only if they lie within e apart. How can I encode this functionally with parameters:
n: spacing between two points on the same axis
d: dimension of R^d
e: maximum distance allowed for an edge to exist.
The graph-tool library has much of the functionality you need. So you could do something like this, assuming you have numpy and graph-tool:
coords = numpy.meshgrid(*(numpy.linspace(0, (n-1)*delta, n) for i in range(d)))
# coords is a Python list of numpy arrays
coords = [c.flatten() for c in coords]
# now coords is a Python list of 1-d numpy arrays
coords = numpy.array(coords).transpose()
# now coords is a numpy array, one row per point
g = graph_tool.generation.geometric_graph(coords, e*(1+1e-9))
The silly e*(1+1e-9) thing is because your criterion is "distance <= e" and geometric_graph's criterion is "distance < e".
There's a parameter called delta that you didn't mention because I think your description of parameter n is doing duty for two params (spacing between points, and number of points).
This bit of code should work, although it certainly isn't the most efficient. It will go through each node and check its distance to all the other nodes (that haven't already compared to it). If that distance is less than your value e then the corresponding value in the connected matrix is set to one. Zero indicates two nodes are not connected.
In this code I'm assuming that your nodeList is a list of cartesian coordinates of the form nodeList = [[x1,y1,...],[x2,y2,...],...[xN,yN,...]]. I also assume you have some function called calcDistance which returns the euclidean distance between two cartesian coordinates. This is basic enough to implement that I haven't written the code for that, and in any case using a function allows for future generalizing and modability.
numNodes = len(nodeList)
connected = np.zeros([numNodes,numNodes])
for i, n1 in enumerate(nodeList):
for j, n2 in enumerate(nodeList[i:]):
dist = calcDistance(n1, n2)
if dist < e:
connected[i,j] = 1
connected[j,i] = 1
Does anyone know a good method to calculate the empirical/sample covariogram, if possible in Python?
This is a screenshot of a book which contains a good definition of covariagram:
If I understood it correctly, for a given lag/width h, I'm supposed to get all the pair of points that are separated by h (or less than h), multiply its values and for each of these points, calculate its mean, which in this case, are defined as m(x_i). However, according to the definition of m(x_{i}), if I want to compute m(x1), I need to obtain the average of the values located within distance h from x1. This looks like a very intensive computation.
First of all, am I understanding this correctly? If so, what is a good way to compute this assuming a two dimensional space? I tried to code this in Python (using numpy and pandas), but it takes a couple of seconds and I'm not even sure it is correct, that is why I will refrain from posting the code here. Here is another attempt of a very naive implementation:
from scipy.spatial.distance import pdist, squareform
distances = squareform(pdist(np.array(coordinates))) # coordinates is a nx2 array
z = np.array(z) # z are the values
cutoff = np.max(distances)/3.0 # somewhat arbitrary cutoff
width = cutoff/15.0
widths = np.arange(0, cutoff + width, width)
Z = []
Cov = []
for w in np.arange(len(widths)-1): # for each width
# for each pairwise distance
for i in np.arange(distances.shape[0]):
for j in np.arange(distances.shape[1]):
if distances[i, j] <= widths[w+1] and distances[i, j] > widths[w]:
m1 = []
m2 = []
# when a distance is within a given width, calculate the means of
# the points involved
for x in np.arange(distances.shape[1]):
if distances[i,x] <= widths[w+1] and distances[i, x] > widths[w]:
m1.append(z[x])
for y in np.arange(distances.shape[1]):
if distances[j,y] <= widths[w+1] and distances[j, y] > widths[w]:
m2.append(z[y])
mean_m1 = np.array(m1).mean()
mean_m2 = np.array(m2).mean()
Z.append(z[i]*z[j] - mean_m1*mean_m2)
Z_mean = np.array(Z).mean() # calculate covariogram for width w
Cov.append(Z_mean) # collect covariances for all widths
However, now I have confirmed that there is an error in my code. I know that because I used the variogram to calculate the covariogram (covariogram(h) = covariogram(0) - variogram(h)) and I get a different plot:
And it is supposed to look like this:
Finally, if you know a Python/R/MATLAB library to calculate empirical covariograms, let me know. At least, that way I can verify what I did.
One could use scipy.cov, but if one does the calculation directly (which is very easy), there are more ways to speed this up.
First, make some fake data that has some spacial correlations. I'll do this by first making the spatial correlations, and then using random data points that are generated using this, where the data is positioned according to the underlying map, and also takes on the values of the underlying map.
Edit 1:
I changed the data point generator so positions are purely random, but z-values are proportional to the spatial map. And, I changed the map so that left and right side were shifted relative to eachother to create negative correlation at large h.
from numpy import *
import random
import matplotlib.pyplot as plt
S = 1000
N = 900
# first, make some fake data, with correlations on two spatial scales
# density map
x = linspace(0, 2*pi, S)
sx = sin(3*x)*sin(10*x)
density = .8* abs(outer(sx, sx))
density[:,:S//2] += .2
# make a point cloud motivated by this density
random.seed(10) # so this can be repeated
points = []
while len(points)<N:
v, ix, iy = random.random(), random.randint(0,S-1), random.randint(0,S-1)
if True: #v<density[ix,iy]:
points.append([ix, iy, density[ix,iy]])
locations = array(points).transpose()
print locations.shape
plt.imshow(density, alpha=.3, origin='lower')
plt.plot(locations[1,:], locations[0,:], '.k')
plt.xlim((0,S))
plt.ylim((0,S))
plt.show()
# build these into the main data: all pairs into distances and z0 z1 values
L = locations
m = array([[math.sqrt((L[0,i]-L[0,j])**2+(L[1,i]-L[1,j])**2), L[2,i], L[2,j]]
for i in range(N) for j in range(N) if i>j])
Which gives:
The above is just the simulated data, and I made no attempt to optimize it's production, etc. I assume this is where the OP starts, with the task below, since the data already exists in a real situation.
Now calculate the "covariogram" (which is much easier than generating the fake data, btw). The idea here is to sort all the pairs and associated values by h, and then index into these using ihvals. That is, summing up to index ihval is the sum over N(h) in the equation, since this includes all pairs with hs below the desired values.
Edit 2:
As suggested in the comments below, N(h) is now only the pairs that are between h-dh and h, rather than all pairs between 0 and h (where dh is the spacing of h-values in ihvals -- ie, S/1000 was used below).
# now do the real calculations for the covariogram
# sort by h and give clear names
i = argsort(m[:,0]) # h sorting
h = m[i,0]
zh = m[i,1]
zsh = m[i,2]
zz = zh*zsh
hvals = linspace(0,S,1000) # the values of h to use (S should be in the units of distance, here I just used ints)
ihvals = searchsorted(h, hvals)
result = []
for i, ihval in enumerate(ihvals[1:]):
start, stop = ihvals[i-1], ihval
N = stop-start
if N>0:
mnh = sum(zh[start:stop])/N
mph = sum(zsh[start:stop])/N
szz = sum(zz[start:stop])/N
C = szz-mnh*mph
result.append([h[ihval], C])
result = array(result)
plt.plot(result[:,0], result[:,1])
plt.grid()
plt.show()
which looks reasonable to me as one can see bumps or troughs at the expected for the h values, but I haven't done a careful check.
The main speedup here over scipy.cov, is that one can precalculate all of the products, zz. Otherwise, one would feed zh and zsh into cov for every new h, and all the products would be recalculated. This calculate could be sped up even more by doing partial sums, ie, from ihvals[n-1] to ihvals[n] at each timestep n, but I doubt that will be necessary.