How to define temporal operators as sparse arrays in both numpy (for CPU) and cupy (for GPU)?

[xinychen]

It is possible to define temporal operators as sparse arrays, and doing this is extremely significant in the case that the temporal dimensionality is very large, e.g., T = 100,000.

• We can first check out the original code for defining temporal operators.

import numpy as np

def generate_Psi(T, d, season):
    Psi = []
    for k in range(0, d + 1):
        if k == 0:
            Psi.append(np.append(np.zeros((T - d - season, d)), 
                                 np.append(-1 * np.eye(T - d - season), np.zeros((T - d - season, season)), axis = 1) 
                                 + np.append(np.zeros((T - d - season, season)), np.eye(T - d - season), axis = 1), axis = 1))
        else:
            Psi.append(np.append(np.append(np.zeros((T - d - season, d - k)), 
                                           np.append(-1 * np.eye(T - d - season), np.zeros((T - d - season, season)), axis = 1)
                                           + np.append(np.zeros((T - d - season, season)), np.eye(T - d - season), axis = 1), axis = 1), 
                                 np.zeros((T - d - season, k)), axis = 1))
    return Psi

• For CPU implementation, we can use numpy and scipy.

import numpy as np
from scipy.sparse import coo_matrix

def generate_Psi(T, d, season):
    Psi = []
    for k in range(0, d + 1):
        if k == 0:
            Psi.append(coo_matrix((- np.ones(T - d - season), 
                                   (np.arange(0, T - d - season), np.arange(d, T - season))), 
                                   shape = (T - d - season, T)).tocsr() 
                       + coo_matrix((np.ones(T - d - season), 
                                    (np.arange(0, T - d - season), np.arange(d + season, T))), 
                                    shape = (T - d - season, T)).tocsr())
        else:
            Psi.append(coo_matrix((- np.ones(T - d - season), 
                                   (np.arange(0, T - d - season), np.arange(d - k, T - k - season))), 
                                   shape = (T - d - season, T)).tocsr() 
                       + coo_matrix((np.ones(T - d - season), 
                                     (np.arange(0, T - d - season), np.arange(d - k + season, T - k))), 
                                     shape = (T - d - season, T)).tocsr())
    return Psi

• For GPU implementation, we can use cupy.

import cupy as np
from cupy.sparse import coo_matrix

def generate_Psi(T, d, season):
    Psi = []
    for k in range(0, d + 1):
        if k == 0:
            Psi.append(coo_matrix((- np.ones(T - d - season), 
                                   (np.arange(0, T - d - season), np.arange(d, T - season))), 
                                   shape = (T - d - season, T)).tocsr() 
                       + coo_matrix((np.ones(T - d - season), 
                                    (np.arange(0, T - d - season), np.arange(d + season, T))), 
                                    shape = (T - d - season, T)).tocsr())
        else:
            Psi.append(coo_matrix((- np.ones(T - d - season), 
                                   (np.arange(0, T - d - season), np.arange(d - k, T - k - season))), 
                                   shape = (T - d - season, T)).tocsr() 
                       + coo_matrix((np.ones(T - d - season), 
                                     (np.arange(0, T - d - season), np.arange(d - k + season, T - k))), 
                                     shape = (T - d - season, T)).tocsr())
    return Psi

Note that this is not an issue, but this is a really interesting question for improving your Python implementation.

[xinychen]

In terms of TMF, the temporal operators can also be defined as sparse matrices. Please check out

import numpy as np
from scipy.sparse import coo_matrix

def generate_Psi(T, d):
    Psi = []
    for k in range(0, d + 1):
        if k == 0:
            Psi.append(coo_matrix((np.ones(T - d), 
                                  (np.arange(0, T - d), np.arange(d, T))), 
                                  shape = (T - d, T)).tocsr())
        else:
            Psi.append(coo_matrix((np.ones(T - d), 
                                  (np.arange(0, T - d), np.arange(d - k, T - k))), 
                                  shape = (T - d, T)).tocsr())
    return Psi