Python NumPy 矢量化以减少处理时间

Question

I am currently trying to improve the performance of my python code. To do so I would like to vectorize the following code:

def phiGCC(self, signal: np.array, TDOA_values: np.array, __frequency_bins__, alpha: float, e_old) -> float:

        alle_taus = []

        temp_gcc_results = np.zeros((len(TDOA_values), len(self.__frequency_bins__)))

        for counter, frequency in enumerate(__frequency_bins__):

            x = np.array([signal[0][counter], signal[1][counter]])

            y = np.conj(x)
            y = y.T

            E = np.zeros((2,2), dtype = complex)
            E[0,0] = x[0] * y[0]
            E[0,1] = x[0] * y[1]
            E[1,0] = x[1] * y[0]
            E[1,1] = x[1] * y[1]


            if(self.interation_count == 1):
                corr_matr = E
            else:

                corr_matr = alpha * self.altes_E[counter] + (1 - alpha) * E

            self.altes_E[counter] = corr_matr


            max_val = 0
            max_tau = 0
            Rxx12 = corr_matr[0][1]

            for tau_index, tau in enumerate(TDOA_values):

                temp = (Rxx12/abs(Rxx12) * np.exp(-2j * np.pi * tau * frequency)).real  #plus frequenz
                temp_gcc_results[tau_index, counter] = temp

                if(temp > max_val):
                    max_val = temp
                    max_tau = tau


            alle_taus.append(max_tau)



        tau_func = np.mean(temp_gcc_results, axis=1)
        maximaler_wert = np.argmax(tau_func)
        return alle_taus[maximaler_wert]

I have read about vectorization and was able to vectorize another function which was very easy and not as complex as the code I have quoted, because there was no function call within the for-loop and I am not sure how I could vectorize the quoted function.

My code works but really slow... and I hope that I can get close to real-time performance by vectorizing it. Any help is appreciated, thanks!

Answer 1

I haven't tried to understand your full code, so can only suggests some initial steps.

For start I try to eliminate inner loops. Only then is there a chance of tackling the outer one.

    for tau_index, tau in enumerate(TDOA_values):
        temp = (Rxx12/abs(Rxx12) * np.exp(-2j * np.pi * tau * frequency)).real  #plus frequenz
        temp_gcc_results[tau_index, counter] = temp

        if(temp > max_val):
            max_val = temp
            max_tau = tau

At a glance I think this can be replaced with:

 temp_gcc_results[:,counter] = (Rxx12/abs(Rxx12) * np.exp(-2j * np.pi * TDOA_values * frequency)).real

 i = np.argmax(temp_gcc_results[:,counter])
 max_val = temp_gcc_results[i]
 max_tau = TDOA_values[i]

Since I have run your code, or these changes, there might be errors. But as best I can see there's no need to evaluate temp in the (python) loop, since the value doesn't depend on the previous. The if clause can be handled with argmax .

        E = np.zeros((2,2), dtype = complex)
        E[0,0] = x[0] * y[0]
        E[0,1] = x[0] * y[1]
        E[1,0] = x[1] * y[0]
        E[1,1] = x[1] * y[1]

This can probably written as a outer product of x and y . Though it's small enough that there won't be much of a time savings.

        E = np.outer(x,y)
        E = x[:,None]*y

Not a speedup, but if corr_matr is an array:

Rxx12 = corr_matr[0,1]