Lazier data loading and static dataset analysis

[PierreGtch]

Amazing, the loop will soon be complete!

With nemar, we can now get really lazy! For example, we can allow the user to download only a specific session, or even a specific run.
We can also have a function to get the exact structure of the dataset (number of subjects, sessions, runs, trials, etc) without downloading the data.
Wdyt @bruAristimunha ?
Maybe best to keep this for a follow-up pr

Originally posted by @PierreGtch in #1084

[davisethan]

Hi @PierreGtch, I recently downloaded many LeftRightImagery datasets with the latest develop branch giving me NEMAR BIDS:

classes = [BNCI2014_001, BNCI2014_004, ...]
for cls in classes:
    ds = cls()
    ds.download(accept=True)

Then, in a separate process, I built evaluations/pipelines:

...
X, _, _ = paradigm.get_data(
    dataset,
    subjects=dataset.subject_list[:1],
    cache_config=dict(use=True, save_array=True, overwrite_array=False),
)
...

I found that paradigm.get_data() isn't wired to find NEMAR downloads:

paradigm.get_data() →
dataset.get_data() →
_get_selected_subject_data() →
_get_single_subject_data_using_cache() →
_get_single_subject_data() # Builds upstream URL

In many cases the dataset is silently, redundantly redownloaded. You opened this issue last month. I'm curious what headway you've made and if help is still wanted. A _get_single_subject_data_from_nemar() function in BaseDataset already exists. But it isn't being used in at least LeftRightImagery dataset classes.

[I-am-Uchenna]

Update after reproducing the routing gap: a general get_data switch is not safe as a narrow patch. NEMAR and legacy loaders can currently produce different session/task selections and sample counts for some deposits, so automatic local-source detection would make results depend on what happens to be cached. I'm stepping back from this slice rather than proposing a non-deterministic source change; the source-selection and cache-key contract needs maintainer alignment first.

[bruAristimunha]

This is an interesting report @I-am-Uchenna, @davisethan. Many thanks for reporting. I will disable the nemar downloading until we are sure of the perfect alignment. If you can point out the discrepancy, please let me know.

[davisethan]

This summarizes my findings: nemar-report.md

[bruAristimunha]

hey @I-am-Uchenna and @davisethan,

Now, it should be fine, and there is no difference anymore between all the details. We re-host the raw and not bids version too at nemar inside the sourcedata folder. Took me a lot of time because we have too many datasets.