Develop Reference

Principal Component Analysis (PCA) in Python numpy using the Snapshot method

I am trying to implement PCA analysis using numpy to mimic the results from sklearn's decomposition.PCA classifier.
I am using as input vectors of N flattened images of fixed size M = 128x192 (image dimensions) joined horizontally into a single matrix D of dimensions MxN
I am aiming to use the Snapshot method, as other implementations (see here and here) crash my build while computing np.cov, since the size of the covariant matrix would be C = D(D^T) = MxM.
The snapshot method first computes C_acute = (D^T)D, then computes the (acute) eigenvectors and values of this NxN matrix. This gives eigenvectors that are (D^T)v, and eigenvalues that are the same.
To retrieve the eigenvectors v from the (acute) eigenvectors, we simply do v = (1/eigenvalue) * (D(v_acute)).
Here is the reference implementation I am using adapted from this SO post (which is known to work):
class TemplatePCA:
def __init__(self, n_components=None):
self.n_components = n_components
def fit_transform(self, X):
X -= np.mean(X, axis = 0)
R = np.cov(X, rowvar=False)
# calculate eigenvectors & eigenvalues of the covariance matrix
evals, evecs = np.linalg.eig(R)
# sort eigenvalue in decreasing order
idx = np.argsort(evals)[::-1]
evecs = evecs[:,idx]
# sort eigenvectors according to same index
evals = evals[idx]
# select the first n eigenvectors (n is desired dimension
# of rescaled data array, or dims_rescaled_data)
evecs = evecs[:, :self.n_components]
# carry out the transformation on the data using eigenvectors
# and return the re-scaled data
return -1 * np.dot(X, evecs) #
Here is the implementation I have so far.
class MyPCA:
def __init__(self, n_components=None):
self.n_components = n_components
def fit_transform(self, X):
X -= np.mean(X, axis = 0)
D = X.T
M, N = D.shape
D_T = X # D.T == (X.T).T == X
C_acute = np.dot(D_T, D)
eigen_values, eigen_vectors_acute = np.linalg.eig(C_acute)
eigen_vectors = []
for i in range(eigen_vectors_acute.shape[0]): # for each eigenvector
v = np.dot(D, eigen_vectors_acute[i]) / eigen_values[i]
eigen_vectors.append(v)
eigen_vectors = np.array(eigen_vectors)
# sort eigenvalues and eigenvectors in decreasing order
idx = np.argsort(eigen_values)[::-1]
eigen_vectors = eigen_vectors[:,idx]
eigen_values = eigen_values[idx]
# select the first n_components eigenvectors
eigen_vectors = eigen_vectors[:, :self.n_components]
# carry out the transformation on the data using eigenvectors
# return the re-scaled data (projection)
return np.dot(C_acute, eigen_vectors)
The reference text I am using notes that:
The eigenvector is now (D^T)v, so to do face detection we first multiply our test image vector by (D^T) before projecting onto the eigenimages.
I am not sure whether it is possible to retrieve the exact same principal components (i.e. eigenvectors) using this method, and it would seem impossible to even get the same eigenvectors back, since the size of the eigen_vectors_acute is only (4, 6) (meaning there are only 4 vectors), compared to the other method where it is (6, 6) (there are 6).
Running both on an input:
x = np.array([
[0.387,123, 789,256, 4878, 5.42],
[0.723,9.78,1.90,1234, 12104,5.25],
[1,123, 67.98,7.91,12756,5.52],
[1.524,1.34,23.456,1.23,6787,3.94],
])
# These two are the same
print(sklearn.decomposition.PCA(n_components=3).fit_transform(x))
print(TemplatePCA(n_components=3).fit_transform(x))
# This one is different
print(MyPCA(n_components=3).fit_transform(x))
Output:
[[ 4282.20163145 147.84415964 -267.73483211]
[-3025.62452358 683.58580386 67.76941319]
[-3599.15380006 -569.33984612 -148.62757658]
[ 2342.57669218 -262.09011737 348.5929955 ]]
[[-4282.20163145 -147.84415964 267.73483211]
[ 3025.62452358 -683.58580386 -67.76941319]
[ 3599.15380006 569.33984612 148.62757658]
[-2342.57669218 262.09011737 -348.5929955 ]]
[[ 3.35535639e+15, -5.70493660e+17, -8.57482740e+17],
[-2.45510474e+15, 4.17428591e+17, 6.27417685e+17],
[-2.82475918e+15, 4.80278997e+17, 7.21885236e+17],
[ 1.92450753e+15, -3.27213928e+17, -4.91820181e+17]]

Is there in Python an approach faster than a double "for" cycle to evaluate a function characterized by an internal vector?

I have the following function:
k=np.linspace(0,5,100)
def f(x,y):
m=k
return sum(np.sin(m-x)*np.exp(-y**2))
I would like to obtain a 2D grid of values of f(x,y) evaluated on these two arrays:
x=np.linspace(0,4,30)
y=np.linspace(0,2,70)
Is there a way of calculation faster than a double "for" cycle like this one?
matrix=np.zeros((len(x),len(y)))
for i in range(len(x)):
for j in range(len(y)):
matrix[i,j]=f(x[i],y[j])
z=matrix.T
I tried to use the "numpy meshgrid" function in this way:
xx,yy=np.meshgrid(x, y)
z=f(xx,yy)
however I got the following error message:
ValueError: operands could not be broadcast together with shapes (100,) (70,30).

Here's a numpy approach. If we start with your original array setup,
k = np.linspace(0,5,100)
x = np.linspace(0,4,30)
y = np.linspace(0,2,70)
then
matrix = np.sin(k[:,np.newaxis] - x).sum(axis = 0)[:,np.newaxis]*np.exp(-y**2)
returns the same (30,70) "matrix" calculated with the double "for" cycle.
For reference, https://docs.scipy.org/doc/numpy/user/basics.broadcasting.html outlines the broadcasting rules in numpy and https://www.numpy.org/devdocs/user/theory.broadcasting.html
gives a nice illustration of using these rules.

k=np.linspace(0,5,100)
x=np.linspace(0,4,30)
y=np.linspace(0,2,70)
def f(x,y):
## m=k
return sum(np.sin(k-x)*np.exp(-y**2))
# original
def g():
m = np.zeros((len(x),len(y)))
for i in range(len(x)):
for j in range(len(y)):
m[i,j]=f(x[i],y[j])
return m.T
# k.shape, x.shape, y.shape -> (100,), (30,), (70,)
# sine of each k minus each x
q = np.sin(k[:,None]-x) # q.shape -> (100,30)
# [sine of each k minus each x] times [e to the negative (y squared)]
r = np.exp(-y**2) # r.shape --> (70,)
s = q[...,None] * r # s.shape --> (100,30,70)
t = s.sum(0)
v = t.T # v.shape -> (70,30)
assert np.all(np.isclose(v,g()))
assert np.all(v == g())
Broadcasting

Correct usage of scipy.interpolate.RegularGridInterpolator

I am a little confused by the documentation for scipy.interpolate.RegularGridInterpolator.
Say for instance I have a function f: R^3 => R which is sampled on the vertices of the unit cube. I would like to interpolate so as to find values inside the cube.
import numpy as np
# Grid points / sample locations
X = np.array([[0,0,0], [0,0,1], [0,1,0], [0,1,1], [1,0,0], [1,0,1], [1,1,0], [1,1,1.]])
# Function values at the grid points
F = np.random.rand(8)
Now, RegularGridInterpolator takes a points argument, and a values argument.
points : tuple of ndarray of float, with shapes (m1, ), ..., (mn, )
The points defining the regular grid in n dimensions.
values : array_like, shape (m1, ..., mn, ...)
The data on the regular grid in n dimensions.
I interpret this as being able to call as such:
import scipy.interpolate as irp
rgi = irp.RegularGridInterpolator(X, F)
However, when I do so, I get the following error:
ValueError: There are 8 point arrays, but values has 1 dimensions
What am I misinterpreting in the docs?

Ok I feel silly when I answer my own question, but I found my mistake with help from the documentation of the original regulargrid lib:
https://github.com/JohannesBuchner/regulargrid
points should be a list of arrays that specifies how the points are spaced along each axis.
For example, to take the unit cube as above, I should set:
pts = ( np.array([0,1.]), )*3
or if I had data which was sampled at higher resolution along the last axis, I might set:
pts = ( np.array([0,1.]), np.array([0,1.]), np.array([0,0.5,1.]) )
Finally, values has to be of shape corresponding to the grid laid out implicitly by points. For example,
val_size = map(lambda q: q.shape[0], pts)
vals = np.zeros( val_size )
# make an arbitrary function to test:
func = lambda pt: (pt**2).sum()
# collect func's values at grid pts
for i in range(pts[0].shape[0]):
for j in range(pts[1].shape[0]):
for k in range(pts[2].shape[0]):
vals[i,j,k] = func(np.array([pts[0][i], pts[1][j], pts[2][k]]))
So finally,
rgi = irp.RegularGridInterpolator(points=pts, values=vals)
runs and performs as desired.

Your answer is nicer, and it's perfectly OK for you to accept it. I'm just adding this as an "alternate" way to script it.
import numpy as np
import scipy.interpolate as spint
RGI = spint.RegularGridInterpolator
x = np.linspace(0, 1, 3) # or 0.5*np.arange(3.) works too
# populate the 3D array of values (re-using x because lazy)
X, Y, Z = np.meshgrid(x, x, x, indexing='ij')
vals = np.sin(X) + np.cos(Y) + np.tan(Z)
# make the interpolator, (list of 1D axes, values at all points)
rgi = RGI(points=[x, x, x], values=vals) # can also be [x]*3 or (x,)*3
tst = (0.47, 0.49, 0.53)
print rgi(tst)
print np.sin(tst[0]) + np.cos(tst[1]) + np.tan(tst[2])
returns:
1.93765972087
1.92113615659

How can an almost arbitrary plane in a 3D dataset be plotted by matplotlib?

There is an array containing 3D data of shape e.g. (64,64,64), how do you plot a plane given by a point and a normal (similar to hkl planes in crystallography), through this dataset?
Similar to what can be done in MayaVi by rotating a plane through the data.
The resulting plot will contain non-square planes in most cases.
Can those be done with matplotlib (some sort of non-rectangular patch)?
Edit: I almost solved this myself (see below) but still wonder how non-rectangular patches can be plotted in matplotlib...?
Edit: Due to discussions below I restated the question.

This is funny, a similar question I replied to just today. The way to go is: interpolation. You can use griddata from scipy.interpolate:
Griddata
This page features a very nice example, and the signature of the function is really close to your data.
You still have to somehow define the points on you plane for which you want to interpolate the data. I will have a look at this, my linear algebra lessons where a couple of years ago

I have the penultimate solution for this problem. Partially solved by using the second answer to Plot a plane based on a normal vector and a point in Matlab or matplotlib :
# coding: utf-8
import numpy as np
from matplotlib.pyplot import imshow,show
A=np.empty((64,64,64)) #This is the data array
def f(x,y):
return np.sin(x/(2*np.pi))+np.cos(y/(2*np.pi))
xx,yy= np.meshgrid(range(64), range(64))
for x in range(64):
A[:,:,x]=f(xx,yy)*np.cos(x/np.pi)
N=np.zeros((64,64))
"""This is the plane we cut from A.
It should be larger than 64, due to diagonal planes being larger.
Will be fixed."""
normal=np.array([-1,-1,1]) #Define cut plane here. Normal vector components restricted to integers
point=np.array([0,0,0])
d = -np.sum(point*normal)
def plane(x,y): # Get plane's z values
return (-normal[0]*x-normal[1]*y-d)/normal[2]
def getZZ(x,y): #Get z for all values x,y. If z>64 it's out of range
for i in x:
for j in y:
if plane(i,j)<64:
N[i,j]=A[i,j,plane(i,j)]
getZZ(range(64),range(64))
imshow(N, interpolation="Nearest")
show()
It's not the ultimate solution since the plot is not restricted to points having a z value, planes larger than 64 * 64 are not accounted for and the planes have to be defined at (0,0,0).

For the reduced requirements, I prepared a simple example
import numpy as np
import pylab as plt
data = np.arange((64**3))
data.resize((64,64,64))
def get_slice(volume, orientation, index):
orientation2slicefunc = {
"x" : lambda ar:ar[index,:,:],
"y" : lambda ar:ar[:,index,:],
"z" : lambda ar:ar[:,:,index]
}
return orientation2slicefunc[orientation](volume)
plt.subplot(221)
plt.imshow(get_slice(data, "x", 10), vmin=0, vmax=64**3)
plt.subplot(222)
plt.imshow(get_slice(data, "x", 39), vmin=0, vmax=64**3)
plt.subplot(223)
plt.imshow(get_slice(data, "y", 15), vmin=0, vmax=64**3)
plt.subplot(224)
plt.imshow(get_slice(data, "z", 25), vmin=0, vmax=64**3)
plt.show()
This leads to the following plot:
The main trick is dictionary mapping orienations to lambda-methods, which saves us from writing annoying if-then-else-blocks. Of course you can decide to give different names,
e.g., numbers, for the orientations.
Maybe this helps you.
Thorsten
P.S.: I didn't care about "IndexOutOfRange", for me it's o.k. to let this exception pop out since it is perfectly understandable in this context.

I had to do something similar for a MRI data enhancement:
Probably the code can be optimized but it works as it is.
My data is 3 dimension numpy array representing an MRI scanner. It has size [128,128,128] but the code can be modified to accept any dimensions. Also when the plane is outside the cube boundary you have to give the default values to the variable fill in the main function, in my case I choose: data_cube[0:5,0:5,0:5].mean()
def create_normal_vector(x, y,z):
normal = np.asarray([x,y,z])
normal = normal/np.sqrt(sum(normal**2))
return normal
def get_plane_equation_parameters(normal,point):
a,b,c = normal
d = np.dot(normal,point)
return a,b,c,d #ax+by+cz=d
def get_point_plane_proximity(plane,point):
#just aproximation
return np.dot(plane[0:-1],point) - plane[-1]
def get_corner_interesections(plane, cube_dim = 128): #to reduce the search space
#dimension is 128,128,128
corners_list = []
only_x = np.zeros(4)
min_prox_x = 9999
min_prox_y = 9999
min_prox_z = 9999
min_prox_yz = 9999
for i in range(cube_dim):
temp_min_prox_x=abs(get_point_plane_proximity(plane,np.asarray([i,0,0])))
# print("pseudo distance x: {0}, point: [{1},0,0]".format(temp_min_prox_x,i))
if temp_min_prox_x < min_prox_x:
min_prox_x = temp_min_prox_x
corner_intersection_x = np.asarray([i,0,0])
only_x[0]= i
temp_min_prox_y=abs(get_point_plane_proximity(plane,np.asarray([i,cube_dim,0])))
# print("pseudo distance y: {0}, point: [{1},{2},0]".format(temp_min_prox_y,i,cube_dim))
if temp_min_prox_y < min_prox_y:
min_prox_y = temp_min_prox_y
corner_intersection_y = np.asarray([i,cube_dim,0])
only_x[1]= i
temp_min_prox_z=abs(get_point_plane_proximity(plane,np.asarray([i,0,cube_dim])))
#print("pseudo distance z: {0}, point: [{1},0,{2}]".format(temp_min_prox_z,i,cube_dim))
if temp_min_prox_z < min_prox_z:
min_prox_z = temp_min_prox_z
corner_intersection_z = np.asarray([i,0,cube_dim])
only_x[2]= i
temp_min_prox_yz=abs(get_point_plane_proximity(plane,np.asarray([i,cube_dim,cube_dim])))
#print("pseudo distance z: {0}, point: [{1},{2},{2}]".format(temp_min_prox_yz,i,cube_dim))
if temp_min_prox_yz < min_prox_yz:
min_prox_yz = temp_min_prox_yz
corner_intersection_yz = np.asarray([i,cube_dim,cube_dim])
only_x[3]= i
corners_list.append(corner_intersection_x)
corners_list.append(corner_intersection_y)
corners_list.append(corner_intersection_z)
corners_list.append(corner_intersection_yz)
corners_list.append(only_x.min())
corners_list.append(only_x.max())
return corners_list
def get_points_intersection(plane,min_x,max_x,data_cube,shape=128):
fill = data_cube[0:5,0:5,0:5].mean() #this can be a parameter
extended_data_cube = np.ones([shape+2,shape,shape])*fill
extended_data_cube[1:shape+1,:,:] = data_cube
diag_image = np.zeros([shape,shape])
min_x_value = 999999
for i in range(shape):
for j in range(shape):
for k in range(int(min_x),int(max_x)+1):
current_value = abs(get_point_plane_proximity(plane,np.asarray([k,i,j])))
#print("current_value:{0}, val: [{1},{2},{3}]".format(current_value,k,i,j))
if current_value < min_x_value:
diag_image[i,j] = extended_data_cube[k,i,j]
min_x_value = current_value
min_x_value = 999999
return diag_image
The way it works is the following:
you create a normal vector:
for example [5,0,3]
normal1=create_normal_vector(5, 0,3) #this is only to normalize
then you create a point:
(my cube data shape is [128,128,128])
point = [64,64,64]
You calculate the plane equation parameters, [a,b,c,d] where ax+by+cz=d
plane1=get_plane_equation_parameters(normal1,point)
then to reduce the search space you can calculate the intersection of the plane with the cube:
corners1 = get_corner_interesections(plane1,128)
where corners1 = [intersection [x,0,0],intersection [x,128,0],intersection [x,0,128],intersection [x,128,128], min intersection [x,y,z], max intersection [x,y,z]]
With all these you can calculate the intersection between the cube and the plane:
image1 = get_points_intersection(plane1,corners1[-2],corners1[-1],data_cube)
Some examples:
normal is [1,0,0] point is [64,64,64]
normal is [5,1,0],[5,1,1],[5,0,1] point is [64,64,64]:
normal is [5,3,0],[5,3,3],[5,0,3] point is [64,64,64]:
normal is [5,-5,0],[5,-5,-5],[5,0,-5] point is [64,64,64]:
Thank you.

The other answers here do not appear to be very efficient with explicit loops over pixels or using scipy.interpolate.griddata, which is designed for unstructured input data. Here is an efficient (vectorized) and generic solution.
There is a pure numpy implementation (for nearest-neighbor "interpolation") and one for linear interpolation, which delegates the interpolation to scipy.ndimage.map_coordinates. (The latter function probably didn't exist in 2013, when this question was asked.)
import numpy as np
from scipy.ndimage import map_coordinates
def slice_datacube(cube, center, eXY, mXY, fill=np.nan, interp=True):
"""Get a 2D slice from a 3-D array.
Copyright: Han-Kwang Nienhuys, 2020.
License: any of CC-BY-SA, CC-BY, BSD, GPL, LGPL
Reference: https://stackoverflow.com/a/62733930/6228891
Parameters:
- cube: 3D array, assumed shape (nx, ny, nz).
- center: shape (3,) with coordinates of center.
can be float.
- eXY: unit vectors, shape (2, 3) - for X and Y axes of the slice.
(unit vectors must be orthogonal; normalization is optional).
- mXY: size tuple of output array (mX, mY) - int.
- fill: value to use for out-of-range points.
- interp: whether to interpolate (rather than using 'nearest')
Return:
- slice: array, shape (mX, mY).
"""
center = np.array(center, dtype=float)
assert center.shape == (3,)
eXY = np.array(eXY)/np.linalg.norm(eXY, axis=1)[:, np.newaxis]
if not np.isclose(eXY[0] # eXY[1], 0, atol=1e-6):
raise ValueError(f'eX and eY not orthogonal.')
# R: rotation matrix: data_coords = center + R # slice_coords
eZ = np.cross(eXY[0], eXY[1])
R = np.array([eXY[0], eXY[1], eZ], dtype=np.float32).T
# setup slice points P with coordinates (X, Y, 0)
mX, mY = int(mXY[0]), int(mXY[1])
Xs = np.arange(0.5-mX/2, 0.5+mX/2)
Ys = np.arange(0.5-mY/2, 0.5+mY/2)
PP = np.zeros((3, mX, mY), dtype=np.float32)
PP[0, :, :] = Xs.reshape(mX, 1)
PP[1, :, :] = Ys.reshape(1, mY)
# Transform to data coordinates (x, y, z) - idx.shape == (3, mX, mY)
if interp:
idx = np.einsum('il,ljk->ijk', R, PP) + center.reshape(3, 1, 1)
slice = map_coordinates(cube, idx, order=1, mode='constant', cval=fill)
else:
idx = np.einsum('il,ljk->ijk', R, PP) + (0.5 + center.reshape(3, 1, 1))
idx = idx.astype(np.int16)
# Find out which coordinates are out of range - shape (mX, mY)
badpoints = np.any([
idx[0, :, :] < 0,
idx[0, :, :] >= cube.shape[0],
idx[1, :, :] < 0,
idx[1, :, :] >= cube.shape[1],
idx[2, :, :] < 0,
idx[2, :, :] >= cube.shape[2],
], axis=0)
idx[:, badpoints] = 0
slice = cube[idx[0], idx[1], idx[2]]
slice[badpoints] = fill
return slice
# Demonstration
nx, ny, nz = 50, 70, 100
cube = np.full((nx, ny, nz), np.float32(1))
cube[nx//4:nx*3//4, :, :] += 1
cube[:, ny//2:ny*3//4, :] += 3
cube[:, :, nz//4:nz//2] += 7
cube[nx//3-2:nx//3+2, ny//2-2:ny//2+2, :] = 0 # black dot
Rz, Rx = np.pi/6, np.pi/4 # rotation angles around z and x
cz, sz = np.cos(Rz), np.sin(Rz)
cx, sx = np.cos(Rx), np.sin(Rx)
Rmz = np.array([[cz, -sz, 0], [sz, cz, 0], [0, 0, 1]])
Rmx = np.array([[1, 0, 0], [0, cx, -sx], [0, sx, cx]])
eXY = (Rmx # Rmz).T[:2]
slice = slice_datacube(
cube,
center=[nx/3, ny/2, nz*0.7],
eXY=eXY,
mXY=[80, 90],
fill=np.nan,
interp=False
)
import matplotlib.pyplot as plt
plt.close('all')
plt.imshow(slice.T) # imshow expects shape (mY, mX)
plt.colorbar()
Output (for interp=False):
For this test case (50x70x100 datacube, 80x90 slice size) the run time is 376 µs (interp=False) and 550 µs (interp=True) on my laptop.

How to run a .py module?

I've got zero experience with Python. I have looked around some tutorial materials, but it seems difficult to understand a advanced code. So I came here for a more specific answer.
For me the mission is to redo the code in my computer.
Here is the scenario:
I'm a graduate student studying tensor factorization in relation learning. A paper[1] providing a code to run this algorithm, as follows:
import logging, time
from numpy import dot, zeros, kron, array, eye, argmax
from numpy.linalg import qr, pinv, norm, inv
from scipy.linalg import eigh
from numpy.random import rand
__version__ = "0.1"
__all__ = ['rescal', 'rescal_with_random_restarts']
__DEF_MAXITER = 500
__DEF_INIT = 'nvecs'
__DEF_PROJ = True
__DEF_CONV = 1e-5
__DEF_LMBDA = 0
_log = logging.getLogger('RESCAL')
def rescal_with_random_restarts(X, rank, restarts=10, **kwargs):
"""
Restarts RESCAL multiple time from random starting point and
returns factorization with best fit.
"""
models = []
fits = []
for i in range(restarts):
res = rescal(X, rank, init='random', **kwargs)
models.append(res)
fits.append(res[2])
return models[argmax(fits)]
def rescal(X, rank, **kwargs):
"""
RESCAL
Factors a three-way tensor X such that each frontal slice
X_k = A * R_k * A.T. The frontal slices of a tensor are
N x N matrices that correspond to the adjecency matrices
of the relational graph for a particular relation.
For a full description of the algorithm see:
Maximilian Nickel, Volker Tresp, Hans-Peter-Kriegel,
"A Three-Way Model for Collective Learning on Multi-Relational Data",
ICML 2011, Bellevue, WA, USA
Parameters
----------
X : list
List of frontal slices X_k of the tensor X. The shape of each X_k is ('N', 'N')
rank : int
Rank of the factorization
lmbda : float, optional
Regularization parameter for A and R_k factor matrices. 0 by default
init : string, optional
Initialization method of the factor matrices. 'nvecs' (default)
initializes A based on the eigenvectors of X. 'random' initializes
the factor matrices randomly.
proj : boolean, optional
Whether or not to use the QR decomposition when computing R_k.
True by default
maxIter : int, optional
Maximium number of iterations of the ALS algorithm. 500 by default.
conv : float, optional
Stop when residual of factorization is less than conv. 1e-5 by default
Returns
-------
A : ndarray
array of shape ('N', 'rank') corresponding to the factor matrix A
R : list
list of 'M' arrays of shape ('rank', 'rank') corresponding to the factor matrices R_k
f : float
function value of the factorization
iter : int
number of iterations until convergence
exectimes : ndarray
execution times to compute the updates in each iteration
"""
# init options
ainit = kwargs.pop('init', __DEF_INIT)
proj = kwargs.pop('proj', __DEF_PROJ)
maxIter = kwargs.pop('maxIter', __DEF_MAXITER)
conv = kwargs.pop('conv', __DEF_CONV)
lmbda = kwargs.pop('lmbda', __DEF_LMBDA)
if not len(kwargs) == 0:
raise ValueError( 'Unknown keywords (%s)' % (kwargs.keys()) )
sz = X[0].shape
dtype = X[0].dtype
n = sz[0]
k = len(X)
_log.debug('[Config] rank: %d | maxIter: %d | conv: %7.1e | lmbda: %7.1e' % (rank,
maxIter, conv, lmbda))
_log.debug('[Config] dtype: %s' % dtype)
# precompute norms of X
normX = [norm(M)**2 for M in X]
Xflat = [M.flatten() for M in X]
sumNormX = sum(normX)
# initialize A
if ainit == 'random':
A = array(rand(n, rank), dtype=dtype)
elif ainit == 'nvecs':
S = zeros((n, n), dtype=dtype)
T = zeros((n, n), dtype=dtype)
for i in range(k):
T = X[i]
S = S + T + T.T
evals, A = eigh(S,eigvals=(n-rank,n-1))
else :
raise 'Unknown init option ("%s")' % ainit
# initialize R
if proj:
Q, A2 = qr(A)
X2 = __projectSlices(X, Q)
R = __updateR(X2, A2, lmbda)
else :
R = __updateR(X, A, lmbda)
# compute factorization
fit = fitchange = fitold = f = 0
exectimes = []
ARAt = zeros((n,n), dtype=dtype)
for iter in xrange(maxIter):
tic = time.clock()
fitold = fit
A = __updateA(X, A, R, lmbda)
if proj:
Q, A2 = qr(A)
X2 = __projectSlices(X, Q)
R = __updateR(X2, A2, lmbda)
else :
R = __updateR(X, A, lmbda)
# compute fit value
f = lmbda*(norm(A)**2)
for i in range(k):
ARAt = dot(A, dot(R[i], A.T))
f += normX[i] + norm(ARAt)**2 - 2*dot(Xflat[i], ARAt.flatten()) + lmbda*(R[i].flatten()**2).sum()
f *= 0.5
fit = 1 - f / sumNormX
fitchange = abs(fitold - fit)
toc = time.clock()
exectimes.append( toc - tic )
_log.debug('[%3d] fit: %.5f | delta: %7.1e | secs: %.5f' % (iter,
fit, fitchange, exectimes[-1]))
if iter > 1 and fitchange < conv:
break
return A, R, f, iter+1, array(exectimes)
def __updateA(X, A, R, lmbda):
n, rank = A.shape
F = zeros((n, rank), dtype=X[0].dtype)
E = zeros((rank, rank), dtype=X[0].dtype)
AtA = dot(A.T,A)
for i in range(len(X)):
F += dot(X[i], dot(A, R[i].T)) + dot(X[i].T, dot(A, R[i]))
E += dot(R[i], dot(AtA, R[i].T)) + dot(R[i].T, dot(AtA, R[i]))
A = dot(F, inv(lmbda * eye(rank) + E))
return A
def __updateR(X, A, lmbda):
r = A.shape[1]
R = []
At = A.T
if lmbda == 0:
ainv = dot(pinv(dot(At, A)), At)
for i in range(len(X)):
R.append( dot(ainv, dot(X[i], ainv.T)) )
else :
AtA = dot(At, A)
tmp = inv(kron(AtA, AtA) + lmbda * eye(r**2))
for i in range(len(X)):
AtXA = dot(At, dot(X[i], A))
R.append( dot(AtXA.flatten(), tmp).reshape(r, r) )
return R
def __projectSlices(X, Q):
q = Q.shape[1]
X2 = []
for i in range(len(X)):
X2.append( dot(Q.T, dot(X[i], Q)) )
return X2
It's boring to paste such a long code but there is no other way to figure out my problems. I'm sorry about this.
I import this module and pass them arguments according to the author's website:
import pickle, sys
from rescal import rescal
rank = sys.argv[1]
X = pickle.load('us-presidents.pickle')
A, R, f, iter, exectimes = rescal(X, rank, lmbda=1.0)
The dataset us-presidents.rdf can be found here.
My questions are:
According to the code note, the tensor X is a list. I don't quite understand this, how do I relate a list to a tensor in Python? Can I understand tensor = list in Python?
Should I convert RDF format to a triple(subject, predicate, object) format first? I'm not sure of the data structure of X. How do I assignment values to X by hand?
Then, how to run it?
I paste the author's code without his authorization, is it an act of infringement? if so, I am so sorry and I will delete it soon.
The problems may be a little bored, but these are important to me. Any help would be greatly appreciated.
[1] Maximilian Nickel, Volker Tresp, Hans-Peter Kriegel,
A Three-Way Model for Collective Learning on Multi-Relational Data,
in Proceedings of the 28th International Conference on Machine Learning, 2011 , Bellevue, WA, USA

To answer Q2: you need to transform the RDF and save it before you can load it from the file 'us-presidents.pickle'. The author of that code probably did that once because the Python native pickle format loads faster. As the pickle format includes the datatype of the data, it is possible that X is some numpy class instance and you would need either an example pickle file as used by this code, or some code doing the pickle.dump to figure out how to convert from RDF to this particular pickle file as rescal expects it.
So this might answer Q1: the tensor consists of a list of elements. From the code you can see that the X parameter to rescal has a length (k = len(X) ) and can be indexed (T = X[i]). So it elements are used as a list (even if it might be some other datatype, that just behaves as such.
As an aside: If you are not familiar with Python and are just interested in the result of the computation, you might get more help contacting the author of the software.

According to the code note, the tensor X is a list. I don't quite understand this, how do I relate a list to a tensor in Python? Can I
understand tensor = list in Python?
Not necessarily but the author of the code has decided to represent the tensor data as a list data structure. As the comments indicate, the list X contains:
List of frontal slices X_k of the tensor X. The shape of each X_k is ('N', 'N')
That means the tensor is repesented as a list of tuples: [(N, N), ..., (N, N)].
I'm not sure of the data structure of X. How do I assignment values to X by hand?
Now that we now the data structure of X, we can assign values to it using assignment. The following will assign the tuple (1, 3) to the first position in the list X (as the first position is at index 0, the second at position 1, et cetera):
X[0] = (1, 3)
Similarly, the following will assign the tuple (2, 4) to the second position:
X[1] = (2, 4)