[wenh06]

Follow-up to #561 (v4.3.1). As analyzed in #574 (with a correction below), the module-level import crash is indeed fixed, but get_contained_labels still passes string columns' .values directly to set_index:

• label_map.set_index(label_map["symbol"].values, inplace=True)
• label_map.set_index(label_map["description"].values, inplace=True)

Under pandas' new string dtype (pandas 3.0, or 2.x with future.infer_string), Series.values returns an ArrowStringArrayNumpySemantics, which set_index rejects with TypeError. Passing the column label is equivalent on every pandas version and immune to the dtype change.

Correction to the analysis in #574 (verified by column dtype + a pandas sandbox): the other .values sites listed there (label_store columns at ~728/~1325/~1368/~1518/~3232) use integer columns, whose .values stays a numpy array and passes set_index today — they are not currently broken. This PR fixes the two sites that actually raise; the integer-column sites are left untouched to keep the diff minimal (they could be unified cosmetically later if preferred).

Verified in a pandas sandbox (future.infer_string = True):

• df.set_index(df["str_col"].values) → TypeError (both sites' pattern)
• df.set_index("str_col") → OK
• integer-column variants → OK (unchanged behavior)

Fixes #574.

[cbrnr]

Just two quick notes after testing on pandas 3.1:

• set_index("symbol") uses drop=True, so the symbol (or description) column no longer appears in contained_labels. The old .values version kept it, and drop=False restores that behavior.
• inplace=True is deprecated in pandas 3.1 and raises Pandas4Warning. Reassigning avoids it:

label_map = label_map.set_index("symbol", drop=False)
label_map = label_map.set_index("description", drop=False)

[wenh06]

Follow-up: while validating this fix downstream (pandas 3.1.0, wfdb 4.3.1), I noticed that set_index("symbol") consumes the column as the index (drop defaults to True), so the returned contained_labels loses the symbol/description column compared to the pandas<3 path, which indexed by an anonymous array and kept every column. Pushed 9af7b67 adding drop=False to both sites -- the output schema now matches the pandas<3 behavior exactly (verified: (1, 4) with label_store/symbol/description/n_occurrences).

[wenh06]

Thanks for the review! Both points are now addressed on this branch:

• drop=False: pushed in 9af7b67 (we had independently hit the same schema change while validating the fix downstream against pandas 2.3.3 -- the returned contained_labels dropped the symbol column; verified identical output with drop=False).
• inplace=True: pushed in 4166cfa -- both string-column set_index calls now use reassignment.

The remaining inplace=True calls on the integer label_store columns (lines ~1325 and ~1367) predate this PR; happy to convert those too (and the set_index at line 1472 in set_label_elements) if you prefer the whole methods cleaned up -- just say the word.

[cbrnr]

Yes, please remove all inplace=True occurrences if possible.

[wenh06]

Done in ccdf1fc -- all six set_index(..., inplace=True) call sites converted to reassignment (custom_labels in check_label_definitions and get_contained_labels, contained_labels, label_map in set_label_from_source, label_df in label_triplets_to_df, and the module-level ann_label_table). The remaining inplace occurrences are wfdb's own API parameters, not pandas calls.

Verified on pandas 3.1.0: rdann -> get_contained_labels -> set_label_elements now emits zero Pandas4Warning, and the contained_labels schema is unchanged.

[cbrnr]

There's one more inplace=True here: https://github.com/wenh06/wfdb-python/blob/ccdf1fcf9da20acfc00d5c107e24962a3848283d/wfdb/io/annotation.py#L3062

Also, can you rebase please?

[wenh06]

Rebased onto the restructured main (df9ccd7). Since the branch was last updated, main moved to the src layout and independently converted all set_index(inplace=True) call sites (annotation.py including ann_class_table, and _header.py's drop) to reassignments -- so the diff is now exactly the two-line fix this PR was about:

label_map = label_map.set_index("symbol", drop=False)      # was label_map["symbol"].values
label_map = label_map.set_index("description", drop=False)  # was label_map["description"].values

(drop=False keeps the column in the returned contained_labels, matching the pandas<3 output schema.)

The crash itself is still reproducible on current main under pandas 3.1.0 (get_contained_labels on any annotation with symbol/description raises TypeError: ... ArrowStringArray); with this diff, rdann -> get_contained_labels -> set_label_elements runs warning-free and the schema is unchanged.

[cbrnr]

@tompollard please merge if you're happy - I'm not sure about a changelog entry?

[tompollard]

Thanks, looks good. I checked it with pandas 3.1 and 2.2.3: the crash is fixed, and the output is identical to before on pandas 2.

Yes, I think it's worth a changelog entry, since rdann(...).get_contained_labels() fails on main with pandas 3. I've added one in changelog.d/, along with a small test (nothing called get_contained_labels before, which is why CI didn't catch this). I'll merge once the checks pass.

@cbrnr thanks for your reviews on this!

[cbrnr]

Very nice! Thanks @wenh06 and @tompollard!