python - Fast(est) exponentiation of numpy 3D matrix - Stack Overflow

admin管理员组

文章数量:1123367

Q is a 3D matrix and could for example have the following shape:

(4000, 25, 25)

I want raise Q to the power n for {0, 1, ..., k} and sum it all. Basically, I want to calculate

\sum_{i=0}^{k-1}Q^n

I have the following function that works as expected:

def sum_of_powers(Q: np.ndarray, k: int) -> np.ndarray:
    Qs = np.sum([
        np.linalg.matrix_power(Q, n) for n in range(k)
    ], axis=0)

    return Qs

Is it possible to speed up my function or is there a faster method to obtain the same output?

Q is a 3D matrix and could for example have the following shape:

(4000, 25, 25)

I want raise Q to the power n for {0, 1, ..., k} and sum it all. Basically, I want to calculate

\sum_{i=0}^{k-1}Q^n

I have the following function that works as expected:

def sum_of_powers(Q: np.ndarray, k: int) -> np.ndarray:
    Qs = np.sum([
        np.linalg.matrix_power(Q, n) for n in range(k)
    ], axis=0)

    return Qs

Is it possible to speed up my function or is there a faster method to obtain the same output?

• python
• numpy

Share Improve this question Follow asked 12 hours ago HJA24HJA24 46822 gold badges1515 silver badges4040 bronze badges 2

• 1 np.float64 and k=300 – HJA24 Commented 12 hours ago
• 1 Watch out for both overflow and underflow, then - 300 is a very large exponent. – user2357112 Commented 3 hours ago

Add a comment  | 

3 Answers 3

Sorted by: Reset to default 3

We can perform this calculation in O(log k) matrix operations.

Let M(k) represent the k'th power of the input, and S(k) represent the sum of those powers from 0 to k. Let I represent an appropriate identity matrix.

Approach 1

If you expand the product, you'll find that (M(1) - I) * S(k) = M(k+1) - I. That means we can compute M(k+1) using a standard matrix power (which takes O(log k) matrix multiplications), and compute S(k) by using numpy.linalg.solve to solve the equation (M(1) - I) * S(k) = M(k+1) - I:

import numpy.linalg

def option1(Q, k):
    identity = numpy.eye(Q.shape[-1])
    A = Q - identity
    B = numpy.linalg.matrix_power(Q, k+1) - identity
    return numpy.linalg.solve(A, B)

Approach 2

The standard exponentation by squaring algorithm computes M(2*k) as M(k)*M(k) and M(2*k+1) as M(2*k)*M(1).

We can alter the algorithm to track both S(k-1) and M(k), by computing S(2*k-1) as S(k-1)*M(k) + S(k-1) and S(2*k) as S(2*k-1) * M(1) + I:

import numpy

def option2(Q, k):
    identity = numpy.eye(Q.shape[-1])

    if k == 0:
        res = numpy.empty_like(Q)
        res[:] = identity
        return res

    power = Q
    sum_of_powers = identity

    # Looping over a string might look dumb, but it's actually the most efficient option,
    # as well as the simplest. (It wouldn't be the bottleneck even if it wasn't efficient.)
    for bit in bin(k+1)[3:]:
        sum_of_powers = (sum_of_powers @ power) + sum_of_powers
        power = power @ power
        if bit == "1":
            sum_of_powers = sum_of_powers @ Q
            sum_of_powers += identity
            power = power @ Q
    return sum_of_powers

Matrix exponentiation by integer powers is the same as chaining matrix multiplications. Since you sum up every power up to a limit, you can simply do regular matrix multiplications.

def sum_of_powers(Q: np.ndarray, k: int) -> np.ndarray:
    def gen_powers():
        if k < 0:
            raise ValueError("negative power not implemented")
        # zeroth power
        yield np.broadcast_to(np.eye(Q.shape[-1], dtype=Q.dtype), Q.shape)
        Qi = Q
        for _ in range(1, k):
            yield Qi
            Qi = Qi @ Q
    return sum(gen_powers())

Performance considerations:

• Using regular sum instead of np.sum saves a ton of temporary memory. Even more effective versions using += are possible (avoiding the temporary allocations that come from repeated array + array)
• You may avoid temporary allocations for the matrix multiplication results by ping-ponging the Qi variable between two buffers and using np.matmul(…, out=…)

Accuracy considerations:

• You will find differences to your straightforward approach since rounding errors are different. You can cache the intermediate results to implement exponentiation by squaring. That doesn't improve performance but it reduces the rounding errors since the n'th power result only requires a chain of O(log(n)) matrix multiplications instead of a chain of n matrix multiplications
• np.sum uses pairwise summing, which also has lower errors but also takes a lot of memory. Since both accuracy improvements need a history of past powers, you can recoup costs using both at the same time if you're smart about the implementation

This more accurate version is implemented here:

def sum_of_powers(Q: np.ndarray, k: int) -> np.ndarray:
    Qis = np.empty((k + 1,) + Q.shape, dtype=Q.dtype)
    Qis[0] = np.eye(Q.shape[-1], dtype=Q.dtype)
    if not k:
        return Qis[0]
    Qis[1] = Q
    for ki in range(2, k):
        # exponentiation by squaring
        # For even powers Q^k = Q^(k//2) @ Q^(k//2)
        # For odd powers Q^k = Q^(k//2) @ Q^(k//2 + 1)
        leftk = ki >> 1
        rightk = leftk + (ki & 1)
        np.matmul(Qis[leftk], Qis[rightk], out=Qis[ki])
    return Qis.sum(axis=0)

There is still a difference in accuracy but I'm not sure my version is actually worse. It might be better since normally matrix powers are implemented via eigen decompositions and those are not free of rounding errors, either. But I have not checked numpy's implementation or done any further theoretical considerations.

TL;DR: Yes, the code can be massively speed up using a more efficient code written in Numba taking care of CPU and memory resources and having a lower complexity.

---

Why the current code is inefficient

The code in the question is very inefficient for multiple reasons.

First of all, the algorithmic complexity of the code is O(k**2) while this can be performed in O(k) with successive matrix multiplications as pointed out by @Homer512.

Moreover, the operation is not cache friendly. Indeed, each matrix multiplication operates on a batch of 4000 times 25x25 matrices. This means 2*8*4000*25*25 ≈ 38 MiB of memory are needed to perform the computation. This does not fit in the LLC cache of most CPUs. As a result, the slow DRAM is used to store data so a significant amount of data need to be fetched and stored back for each iteration (114 MiB/iteration).

Additionally, the operation should be sequential since Numpy parallelize no operation and the BLAS library used to perform the matrix multiplications cannot parallelize the code either because the 25x25 matrix is too small for multiple threads to worth it.

Finally, each matrix multiplication is certainly sub-optimal because of the overhead of Numpy internal iterators and the one of calling BLAS functions for each matrix.

---

Faster implementation

We can write a Numba code iteratively performing the matrix multiplication on each 25x25 matrix and summing this up. The code can be parallelized trivially so each thread operate on a chunk along the last axis (of size 4000). The operation is thus parallel, cache friendly, without any Numpy/CPython overhead and SIMD-friendly thanks to Numba. Here is the resulting code:

import numba as nb

@nb.njit('(float64[:,:,::1], int32)', parallel=True)
def sum_of_powers_numba(Q: np.ndarray, k: int) -> np.ndarray:
    s0, s1, s2 = Q.shape
    assert s1 == s2
    Qs = np.zeros((s0, s1, s1))
    for i in nb.prange(s0):
        tmp = np.eye(s1, s1)
        Qs[i,:,:] = tmp
        for n in range(k-1):
            tmp = tmp @ Q[i]
            Qs[i,:,:] += tmp
    return Qs

The BLAS library seems to be sub-optimal for 25x25 matrices but not so bad either. Writing a faster implementation in Numba is very challenging. It is easier in native languages. I think it does not worth the effort of rewriting a specialized implementation for small matrices. I expect the BLIS library to be a bit faster (close to optimal) than the default BLAS implementation (OpenBLAS), so it is certainly better just to try it (or possibly the highly-optimized Intel MKL).

Besides, if you want a more accurate version, then you should adapt the code so to perform an exponentiation by squaring as shown by @Homer512 while still using Numba.

---

Benchmark (updated)

Here are results on my i5-9600KF CPU (6 cores):

sum_of_powers_naive:      23.6  sec
sum_of_powers_homer512:    7.8  sec
sum_of_powers_numba:       1.5  sec  <-----
user2357112's option2:     0.16 sec
user2357112's option1:     0.13 sec

Interestingly, profiling results show that about ~80% of the time is spent in the BLAS library so the main speed up comes from parallelizing the code with nb.prange in Numba).

In the end, the approach of @user2357112 is the best since it is the fastest and certainly the most numerically stable. Note it can be faster by parallelizing the computation over the first axis.

本文标签： pythonFast(est) exponentiation of numpy 3D matrixStack Overflow