[Claude + Human] Add remove_nan_targets option to drop NaN-target recordings

[duanyiqun]

Summary

• Add a remove_nan_targets parameter to EEGDashDataset. When target_name is set and remove_nan_targets=True, recordings whose target value is missing (None/NaN) are dropped at construction, with a warning reporting the count.

Design note: default is False

• Issue Remove NaN at import time for the target variable #22 / @arnodelorme suggested defaulting this on. We deliberately default to False instead, because dropping recordings by default would silently change len(dataset) for existing users and is a backward-incompatible behaviour change. Opt in with remove_nan_targets=True.
• Regardless of this flag, the all-targets-missing case still raises a ValueError (handled by the target_name validation in Check var when creating dataset #21), which matches the "error if all the datasets are NaN" request.

Dependency

• Stacked on Check var when creating dataset #21 (target_name validation, PR [Claude + Human] Validate target_name and raise on a misspelled name #381). This branch contains the Check var when creating dataset #21 commit as well; please merge [Claude + Human] Validate target_name and raise on a misspelled name #381 first, after which this branch rebased onto develop shows only its own diff.

Impact

• Opt-in convenience for target-based training so users don't have to post-filter NaN-target recordings manually. Default behaviour is unchanged.

Validation

• ruff check + ruff format --check clean on changed files
• pytest tests/unit_tests/dataset/test_remove_nan_targets.py tests/unit_tests/dataset/test_target_name_validation.py → 7 passed

Closes #22.

🤖 Generated with Claude Code

[bruAristimunha]

I maybe have the impression that this issue was fix in the pass, can you create an minimal code that reproduce the current code source?

[duanyiqun]

@bruAristimunha sure on my dev branch seems not let me double check with code on that.

[duanyiqun]

@bruAristimunha I double-checked on current develop — it's not handled yet. Minimal repro (offline, no network), three recordings where the target age is present / None / NaN:

import tempfile
from eegdash import EEGDashDataset

tmp = tempfile.mkdtemp()
def rec(sub, age):
    r = {
        "dataset": "dsTest",
        "bids_relpath": f"sub-{sub}/eeg/sub-{sub}_task-rest_eeg.set",
        "bidspath": f"dsTest/sub-{sub}/eeg/sub-{sub}_task-rest_eeg.set",
        "storage": {"backend": "local", "base": tmp},
        "entities_mne": {"subject": sub, "task": "rest"},
        "ntimes": 100, "sampling_frequency": 100,
    }
    if age is not None:
        r["age"] = age
    return r

records = [rec("01", 25), rec("02", None), rec("03", float("nan"))]
ds = EEGDashDataset(cache_dir=tmp, dataset="dsTest", records=records,
                    download=False, target_name="age")

print(len(ds.datasets))                                   # -> 3
print([d.description.get("age") for d in ds.datasets])    # -> [25, nan, nan]

Output on develop:

3
[25, nan, nan]

So the NaN/None-target recordings are silently kept, and there is no switch to drop them — EEGDashDataset(..., remove_nan_targets=True) raises TypeError: RawDataset.__init__() got an unexpected keyword argument 'remove_nan_targets' because the kwarg is forwarded to braindecode.

This PR adds that opt-in switch (default False to stay backward-compatible). The "all targets NaN" case is the hard error, handled by the target_name validation in #381 that this branch is stacked on.

[bruAristimunha]

thanks for the PRs @duanyiqun, if you want talk and structure the efforts more, please ping at linkedin :)