LU is not same as A when I used torch.solver - pytorch

let A=[[3,2],[1,-3]] and B=[[3],[-10]]
and solve equation AX=B by using torch.solve:
X, LU = torch.solve(B,A)
Then I got X=[[-1],[3]] and LU=[[3,2],[0.333,-3.666]].
According to definition of LU decompose, LU must be same as A, however they aren't same.
Can anyone explain this???
Thank you

The representation you got is a compact way of representing the lower trainagular matrix L and the upper trainagular matrix U. You can use torch.tril and torch.triu to get these matrices explicitly:
L = torch.tril(LU, -1) + torch.eye(LU.shape[-1])
U = torch.triu(LU)
verify:
In [*]: L
Out[*]:
tensor([[1.0000, 0.0000],
[0.3333, 1.0000]])
In [*]: U
Out[*]:
tensor([[ 3.0000, 2.0000],
[ 0.0000, -3.6667]])
And the product is indeed equal to A:
In [*]: torch.dist(L # U , A)
Out[*]: tensor(0.)

Related

SymPy result Filtering

I was recently working on a CodeForce problem
So, I was using SymPy to solve this.
My code is :
from sympy import *
x,y = symbols("x,y", integer = True)
m,n = input().split(" ")
sol = solve([x**2 + y - int(n), y**2 + x - int(m)], [x, y])
print(sol)
What I wanted to do:
Filter only Positive and integer value from SymPy
Ex: If I put 14 28 in the terminal it will give me tons of result, but I just want it to show [(5, 3)]
I don't think that this is the intended way to solve the code force problem (I think you're just supposed to loop over the possible values for one of the variables).
I'll show how to make use of SymPy here anyway though. Your problem is a diophantine system of equations. Although SymPy has a diophantine solver it only works for individual equations rather than systems.
Usually the idea of using a CAS for something like this though is to symbolically find something like a general result that then helps you to write faster concrete numerical code. Here are your equations with m and n as arbitrary symbols:
In [62]: x, y, m, n = symbols('x, y, m, n')
In [63]: eqs = [x**2 + y - n, y**2 + x - m]
Using the polynomial resultant we can eliminate either x or y from this system to obtain a quartic polynomial for the remaining variable:
In [31]: py = resultant(eqs[0], eqs[1], x)
In [32]: py
Out[32]:
2 2 4
m - 2⋅m⋅y - n + y + y
While there is a quartic general formula that SymPy can use (if you use solve or roots here) it is too complicated to be useful for a problem like the one that you are describing. Instead though the rational root theorem tells us that an integer root for y must be a divisor of the constant term:
In [33]: py.coeff(y, 0)
Out[33]:
2
m - n
Therefore the possible values for y are:
In [64]: yvals = divisors(py.coeff(y, 0).subs({m:14, n:28}))
In [65]: yvals
Out[65]: [1, 2, 3, 4, 6, 7, 8, 12, 14, 21, 24, 28, 42, 56, 84, 168]
Since x is m - y**2 the corresponding values for x are:
In [66]: solve(eqs[1], x)
Out[66]:
⎡ 2⎤
⎣m - y ⎦
In [67]: xvals = [14 - yv**2 for yv in yvals]
In [68]: xvals
Out[68]: [13, 10, 5, -2, -22, -35, -50, -130, -182, -427, -562, -770, -1750, -3122, -7042, -28210]
The candidate solutions are then given by:
In [69]: candidates = [(xv, yv) for xv, yv in zip(xvals, yvals) if xv > 0]
In [70]: candidates
Out[70]: [(13, 1), (10, 2), (5, 3)]
From there you can test which values are solutions:
In [74]: eqsmn = [eq.subs({m:14, n:28}) for eq in eqs]
In [75]: [c for c in candidates if all(eq.subs(zip([x,y],c))==0 for eq in eqsmn)]
Out[75]: [(5, 3)]
The algorithmically minded will probably see from the above example how to make a much more efficient way of implementing the solver.
I've figured out the answer to my question ! At first, I was trying to filter the result from solve(). But there is an easy way to do this.
Pseudo code:
solve() gives the intersection point of both Parabolic Equations as a List
I just need to filter() the other types of values. Which in my case is <sympy.core.add.Add>
def rem(_list):
return list(filter(lambda v: type(v) != Add, _list))
Yes, You can also use type(v) == int
Final code:
from sympy import *
# the other values were <sympy.core.add.Add> type. So, I just defined a function to filterOUT these specific types from my list.
def rem(_list):
return list(filter(lambda v: type(v) != Add, _list))
x,y = symbols("x,y", integer = True, negative = False)
output = []
m,n = input().split(' ')
# I need to solve these 2 equations separately. Otherwise, my defined function will not work without loop.
solX = rem(solve((x+(int(n)-x**2)**2 - int(m)), x))
solY = rem(solve((int(m) - y**2)**2 + y - int(n), y))
if len(solX) == 0 or len(solY) == 0:
print(0)
else:
output.extend(solX) # using "Extend" to add multiple values in the list.
output.extend(solY)
print(int((len(output))/2)) # Obviously, result will come in pairs. So, I need to divide the length of the list by 2.
Why I used this way :
I tried to solve it by algorithmic way, but it still had some float numbers. I just wanted to skip the loop thing here again !
As sympy solve() has already found the values. So, I skipped the other way and focused on filtering !
Sadly, code force compiler shows a runtime error! I guess it can't import sympy. However, it works fine in VSCode.

creating a list of tuples based on successive items of initial list [duplicate]

I sometimes need to iterate a list in Python looking at the "current" element and the "next" element. I have, till now, done so with code like:
for current, next in zip(the_list, the_list[1:]):
# Do something
This works and does what I expect, but is there's a more idiomatic or efficient way to do the same thing?
Some answers to this problem can simplify by addressing the specific case of taking only two elements at a time. For the general case of N elements at a time, see Rolling or sliding window iterator?.
The documentation for 3.8 provides this recipe:
import itertools
def pairwise(iterable):
"s -> (s0, s1), (s1, s2), (s2, s3), ..."
a, b = itertools.tee(iterable)
next(b, None)
return zip(a, b)
For Python 2, use itertools.izip instead of zip to get the same kind of lazy iterator (zip will instead create a list):
import itertools
def pairwise(iterable):
"s -> (s0, s1), (s1, s2), (s2, s3), ..."
a, b = itertools.tee(iterable)
next(b, None)
return itertools.izip(a, b)
How this works:
First, two parallel iterators, a and b are created (the tee() call), both pointing to the first element of the original iterable. The second iterator, b is moved 1 step forward (the next(b, None)) call). At this point a points to s0 and b points to s1. Both a and b can traverse the original iterator independently - the izip function takes the two iterators and makes pairs of the returned elements, advancing both iterators at the same pace.
Since tee() can take an n parameter (the number of iterators to produce), the same technique can be adapted to produce a larger "window". For example:
def threes(iterator):
"s -> (s0, s1, s2), (s1, s2, s3), (s2, s3, 4), ..."
a, b, c = itertools.tee(iterator, 3)
next(b, None)
next(c, None)
next(c, None)
return zip(a, b, c)
Caveat: If one of the iterators produced by tee advances further than the others, then the implementation needs to keep the consumed elements in memory until every iterator has consumed them (it cannot 'rewind' the original iterator). Here it doesn't matter because one iterator is only 1 step ahead of the other, but in general it's easy to use a lot of memory this way.
Roll your own!
def pairwise(iterable):
it = iter(iterable)
a = next(it, None)
for b in it:
yield (a, b)
a = b
Starting in Python 3.10, this is the exact role of the pairwise function:
from itertools import pairwise
list(pairwise([1, 2, 3, 4, 5]))
# [(1, 2), (2, 3), (3, 4), (4, 5)]
or simply pairwise([1, 2, 3, 4, 5]) if you don't need the result as a list.
I’m just putting this out, I’m very surprised no one has thought of enumerate().
for (index, thing) in enumerate(the_list):
if index < len(the_list):
current, next_ = thing, the_list[index + 1]
#do something
Since the_list[1:] actually creates a copy of the whole list (excluding its first element), and zip() creates a list of tuples immediately when called, in total three copies of your list are created. If your list is very large, you might prefer
from itertools import izip, islice
for current_item, next_item in izip(the_list, islice(the_list, 1, None)):
print(current_item, next_item)
which does not copy the list at all.
Iterating by index can do the same thing:
#!/usr/bin/python
the_list = [1, 2, 3, 4]
for i in xrange(len(the_list) - 1):
current_item, next_item = the_list[i], the_list[i + 1]
print(current_item, next_item)
Output:
(1, 2)
(2, 3)
(3, 4)
I am really surprised nobody has mentioned the shorter, simpler and most importantly general solution:
Python 3:
from itertools import islice
def n_wise(iterable, n):
return zip(*(islice(iterable, i, None) for i in range(n)))
Python 2:
from itertools import izip, islice
def n_wise(iterable, n):
return izip(*(islice(iterable, i, None) for i in xrange(n)))
It works for pairwise iteration by passing n=2, but can handle any higher number:
>>> for a, b in n_wise('Hello!', 2):
>>> print(a, b)
H e
e l
l l
l o
o !
>>> for a, b, c, d in n_wise('Hello World!', 4):
>>> print(a, b, c, d)
H e l l
e l l o
l l o
l o W
o W o
W o r
W o r l
o r l d
r l d !
This is now a simple Import As of 16th May 2020
from more_itertools import pairwise
for current, next in pairwise(your_iterable):
print(f'Current = {current}, next = {nxt}')
Docs for more-itertools
Under the hood this code is the same as that in the other answers, but I much prefer imports when available.
If you don't already have it installed then:
pip install more-itertools
Example
For instance if you had the fibbonnacci sequence, you could calculate the ratios of subsequent pairs as:
from more_itertools import pairwise
fib= [1,1,2,3,5,8,13]
for current, nxt in pairwise(fib):
ratio=current/nxt
print(f'Curent = {current}, next = {nxt}, ratio = {ratio} ')
As others have pointed out, itertools.pairwise() is the way to go on recent versions of Python. However, for 3.8+, a fun and somewhat more concise (compared to the other solutions that have been posted) option that does not require an extra import comes via the walrus operator:
def pairwise(iterable):
a = next(iterable)
yield from ((a, a := b) for b in iterable)
A basic solution:
def neighbors( list ):
i = 0
while i + 1 < len( list ):
yield ( list[ i ], list[ i + 1 ] )
i += 1
for ( x, y ) in neighbors( list ):
print( x, y )
Pairs from a list using a list comprehension
the_list = [1, 2, 3, 4]
pairs = [[the_list[i], the_list[i + 1]] for i in range(len(the_list) - 1)]
for [current_item, next_item] in pairs:
print(current_item, next_item)
Output:
(1, 2)
(2, 3)
(3, 4)
code = '0016364ee0942aa7cc04a8189ef3'
# Getting the current and next item
print [code[idx]+code[idx+1] for idx in range(len(code)-1)]
# Getting the pair
print [code[idx*2]+code[idx*2+1] for idx in range(len(code)/2)]

Undoing tensor dstack and restack column-wise

I have two, tensors, a and b:
import torch
a = torch.tensor(([1,2],
[3,4],
[5,6],
[7,8]))
b = torch.tensor(([0,0],
[1,1],
[2,2],
[3,3]))
Which I can stack both horizontally or depth-wise.
d = torch.dstack([a, b])
h = torch.hstack([a, b])
Now, is there any PyTorch function, preferably in one line, that I can apply to d to get h? It sounds like I want to undo the depth-wise stacking, and re-stack them column-wise. I've tried reshaping, and flattening, but neither work, as they both disrupt the ordering of the values.
in your case use torch.unbind
import torch
a = torch.tensor(([1,2],
[3,4],
[5,6],
[7,8]))
b = torch.tensor(([0,0],
[1,1],
[2,2],
[3,3]))
d = torch.dstack([a, b])
h = torch.hstack(torch.unbind(d,2)) # get h from d

list comprehension in matrix multiplication in python 3.x

I have found a code of Matrix Multiplication in Python 3.x but I am not able to understand how list comprehension is working in the below code.
# Program to multiply two matrices using list comprehension
# 3x3 matrix
X = [[12,7,3],
[4 ,5,6],
[7 ,8,9]]
# 3x4 matrix
Y = [[5,8,1,2],
[6,7,3,0],
[4,5,9,1]]
# result is 3x4
result = [[sum(a*b for a,b in zip(X_row,Y_col)) for Y_col in zip(*Y)] for X_row in X]
for r in result:
print(r)
#Santosh, it's probably easier to understand this List Comprehension from pure loop way, like this:
# 3x3 matrix
X = [[12,7,3],
[4,5,6],
[7,8,9]]
# 3x4 matrix
Y = [[5,8,1,2],
[6,7,3,0],
[4,5,9,1]]
# result is 3x4
result = [[0,0,0,0],
[0,0,0,0],
[0,0,0,0]]
# iterate through rows of X
for r in range(len(X)):
# iterate through columns of Y
for c in range(len(Y[0])):
# iterate through rows of Y
for k in range(len(Y)):
result[r][c] += X[r][k] * Y[k][c]
print(result)
Then you prob. can find the similarity with the List Comprehension version, with little reformatting:
def matrix_mul(X, Y):
zip_b = list(zip(*Y))
return [[sum(a * b for a, b in zip(row_a, col_b))
for col_b in zip_b]
for row_a in X]

Solving matrix equations with some Boundary Conditions

How can I solve a system of linear equations with some Boundary conditions, using Numpy?
Ax=B
Where x is a column vector with, let's say x1=0.
For different iterations BCs are going to be different, so different variables of vector x going to be zero.
[A] and [B] are known.
Here is an example from my FEM course:
{F} Is the column vector of known values
[k] is the stiffness matrix with the known values
{U} is the displacement column vector where U1 and U3 are known to be zero, but U2 and U4 need to be found.
Here is an example:
This would result in these values:
Naturally this would reduce to the 2X2 matrix equation, but I because for different elements the BC would be different, I'm looking for some numpy matrix equation solver where I can let it know that some of the unknowns must be this certain value and nothing else.
Is there something similar to np.linalg.solve() with conditions to it?
Thank you.
the matrix k in your example is invertible. that means there is one and only one solution; you can not choose any of the Us. this is the solution:
import numpy as np
k = np.array(((1000, 0, -1000, 0),
(0, 3000, 0, -3000),
(-100, 0, 3000, -2000),
(0, -3000, -2000, 5000)))
F = np.array((0, 0, 0, 5000))
U = np.linalg.solve(k, F)
print(U)
# # or:
# k_inv = np.linalg.inv(k)
# U = k_inv.dot(F)
# [ 5.55555556 8.05555556 5.55555556 8.05555556]
the same in sage:
k = matrix(((1000, 0, -1000, 0),
(0, 3000, 0, -3000),
(-100, 0, 3000, -2000),
(0, -3000, -2000, 5000)))
F = vector((0, 0, 0, 5000))
U = k.inverse() * F
# (50/9, 145/18, 50/9, 145/18)

Resources