add pairwise granger for subset of channel pairs #58
Conversation
|
Hi @leoscholl, Thank you for your contribution! For subsets of channels, could you not just put use that subset of signals or is there some efficiency that I'm not thinking of? |
|
I'm computing granger prediction from one channel to all other channels, so estimating the full pairwise matrix takes a long time (and a lot of memory, but i'm ignoring that problem for now) and most of the values aren't used. |
There was a problem hiding this comment.
Pull Request Overview
This PR adds functionality to compute connectivity measures and predictive power for a subset of channel pairs, addressing performance concerns when dealing with large channel counts.
- Introduces a _subset_cross_spectral_matrix method to compute connectivity measures for specified channel pairs.
- Adds a subset_pairwise_spectral_granger_prediction method to evaluate predictive power for given pair combinations.
| for x in _syu: | ||
| diag_out = self._expectation( | ||
| _complex_inner_product( | ||
| fourier_coefficients[..., [x], :], | ||
| fourier_coefficients[..., [x], :], | ||
| dtype=self._dtype, | ||
| ) | ||
| ) | ||
| csm[..., [x], [x]] = diag_out[..., 0] | ||
| for x in _sxu: |
There was a problem hiding this comment.
The diagonal computation is being performed separately for _syu and _sxu, which may lead to redundant processing if these arrays overlap. Consider computing the unique set of indices from both arrays and iterating over that set to avoid duplicate work.
| for x in _syu: | |
| diag_out = self._expectation( | |
| _complex_inner_product( | |
| fourier_coefficients[..., [x], :], | |
| fourier_coefficients[..., [x], :], | |
| dtype=self._dtype, | |
| ) | |
| ) | |
| csm[..., [x], [x]] = diag_out[..., 0] | |
| for x in _sxu: | |
| unique_indices = xp.unique(xp.concatenate((_syu, _sxu))) | |
| for x in unique_indices: |
Copilot uses AI. Check for mistakes.
| predictive_power = np.empty(new_shape) | ||
|
|
||
| for pair_indices in pairs: | ||
| pair_indices = np.array(pair_indices)[:, np.newaxis] |
There was a problem hiding this comment.
[nitpick] Converting pair_indices to a numpy array inside the loop introduces unnecessary overhead if pairs were already converted earlier. Consider pre-processing the pairs outside the loop to optimize performance.
| pair_indices = np.array(pair_indices)[:, np.newaxis] |
Copilot uses AI. Check for mistakes.
with large channel counts it can be useful to compute cross-spectral estimates and connectivity measures for just a subset of channels. i'm not sure if the way i accomplished this is the best approach or if you have other suggestions. it works for my purposes, anyway. let me know if you think it's useful