Speed up

[hjenryin]

Thanks you for developing this tool!

When I was using it on my data, I found multiple times that the process get stuck when calling smECE(pred,y). In the end, I realized that it was due to the inefficient implementation of np.convolve. I suggest that sp.signal.convolve be used, as it uses FFT to speed up calculation. This difference has also been pointed out in https://stackoverflow.com/questions/53550764/python-numpy-is-there-a-faster-way-to-convolve.

I have attached an example I came across on real data: debug_y_33.5.modify_affix_to_npy.txt
debug_pred_33.5.modify_affix_to_npy.txt (Github doesn't allow .npy files to be uploaded, so please change the affix back to .npy)

The results are:

• sp.signal.convolve: Time = 0.065s, smECE = 0.006961
• np.convolve:
  • smECE = 0.006961 as well.
  • It seems that the time spent is not stable. It may vary from one to tens of seconds. But even in the best case, it's still way more time-consuming then the sp.signal implementation.