Sliding window issues

[mibur1]

Hi,

When using any method that requires a sliding window, I noticed that pydfc does not allow the user to specfy the standard deviation of the Gaussian. Rather, this is chosen automatically depending of the size of the window. In sliding_window.py:

window_taper = signal.windows.gaussian(W, std=3*W/22)

This is then used to create a Gaussian that is wider than the chosen window size:

window = signal.convolve(window, window_taper, mode='same') / sum(window_taper)

This is then accounted for by using wider windows:

np.multiply(time_series, window)[:,max(int(l-W/2),0):min(int(l+3*W/2),L)]

Is there a specific reason why this is implemented that way? For connectivity estimates, I would intuitively expect the window size to correspond the total included samples.

However, the main concern I have is that the same extended window is used for rectangular windows. As far as I can see, the resulting zero padding is not accounted for when calculating FC, leading to erroneous estimates.

Possible solutions could be:

• Use the fixed size for both windows (rectangular and tapered) np.multiply(time_series, window)[:,l:l+W]
• Handle the window shapes separately

Generally, I think the Pearson correlation estimation would also benefit from being vectorized instead of using nested loops as this would increase performance drastically.

I'm happy to discuss this further and can also provide a PR if that would be welcome.

[mtorabi59]

Hi @mibur1 ,

Thank you for bringing this up. A sliding window with rectangular window will indeed give erroneous estimates because of the zero padding. In case of tapered window, it is not a problem, but I agree that keeping the window length equal to W might be better.
Also regarding the sd value, it would be nice to have it as a parameter that can be set. The current value is set following Allen et al. paper.

So the changes being proposed are:

• Using the fixed size for both windows (rectangular and tapered) np.multiply(time_series, window)[:,l:l+W]
• Allowing users to set the sd value for tapered window
• vectorizing Pearson correlation estimation to increase the performance

You are welcome to work on any of these yourself and I'll work on the rest! Thanks for contributing to PydFC!

[mibur1]

Thanks for the quick reply!

I already have code for all of these features, so I can send a PR in the next few days if that's okay.

[mtorabi59]

@mibur1 sounds great! Thanks Micha!