4d (multiple frequencies) reshape issue in DatasetEphy about frites HOT 2 CLOSED

Hello @StanSStanman !

Thanks for pointing me out the issue. Probably the best way to fix it would be that the self._reshape becomes a list of shapes instead of just the shape of the first element. Something like :

In ds_ephy :

self._reshape = None
if all([k.ndim == 4 for k in x]):
    logger.debug(f"    4d reshaping")
    self._reshape = []
    for k in range(len(x)):
        n_e, n_r, n_f, n_t = x[k].shape
        x[k] = x[k].reshape(n_e, n_r, n_f * n_t)
        self._reshape.append((n_f, n_t))

Then, you also need to modify the WfMi with something like :

self._reshape = dataset._reshape
if isinstance(self._reshape, list):
    logger.debug(f"    reshaping before computing statistics")
    for k in range(len(mi)):
        n_f, n_t = self._reshape[k]
        n_p, n_s, _ = mi_p[k].shape
        mi[k] = mi[k].reshape(n_s, n_f, n_t)
        mi_p[k] = mi_p[k].reshape(n_p, n_s, n_f, n_t)

If you want to test your code, there's one example illustrating spatio-temporal clusters (online doc and github code).

I suggest that you modify this example such that each simulated subject has a different number of trials. Then you can start to modify the scripts and submit the PR.

What do you think?

from frites.

Hello @EtienneCmb

Thanks for your answer. As we discussed trough other channels, I'm going to fix the bug, but without transforming shape._reshape in a list of tuples.
Indeed this attribute is in the form of (n_freqs, n_times) and those two dimensions should remains the same across all the subjects.

I will also update the example code, to include the case of computing mi across subjects with a variable number of trials.

See you soon!

from frites.