max / audiofiles

Scope the bench drum-detection metric to drum-expected samples The "Layer 1 accuracy (drum detection)" line counted is_drum_class(predicted) across every result without consulting expected, so a non-drum sample predicted as a drum counted as correct. Its own comment already claimed the narrower thing: "predicted is any drum class when expected is drum". Masked exactly by the corpus: expected_class_from_dir maps only the seven drum classes and is_drum_class covers exactly those seven, so the subset is the whole set and today's 89.4% is unchanged. It stops being wrong the moment a non-drum class enters the corpus, which the correctness thread will do. The printed fraction now carries the same denominator as the percentage rather than total_classified, which would have disagreed once the two diverge.

Author: Max Johnson <me@maxj.phd> · 2026-07-29 18:04 UTC

Signed with PGP, not checked

Commit: 8addfeb0e3a7f18a51dcbd2582c5927da57a7589

Parent: a436e63

1 file changed, +20 insertions, -4 deletions

M crates/audiofiles-bench/src/main.rs +20 -4

			@@ -738,19 +738,35 @@
738	738		.count();
739	739		let strict_acc = strict_correct as f64 / total_classified as f64 * 100.0;
740	740
741		-	// Layer 1 accuracy: predicted is any drum class when expected is drum
742		-	let drum_correct = all_results
	741	+	// Layer 1 accuracy: predicted is any drum class when expected is drum.
	742	+	//
	743	+	// Scoped to the drum-expected subset, which is what the name claims. It
	744	+	// previously counted is_drum_class(predicted) over every result without
	745	+	// consulting expected at all, so a non-drum sample predicted as a drum
	746	+	// scored as correct. The corpus is currently 100% drums, which masked it
	747	+	// exactly; the number is unchanged today and stops being wrong the moment
	748	+	// a non-drum class enters the corpus.
	749	+	let drum_expected: Vec<&ClassifyResult> = all_results
	750	+	.iter()
	751	+	.filter(\|r\| is_drum_class(r.expected))
	752	+	.collect();
	753	+	let drum_correct = drum_expected
743	754		.iter()
744	755		.filter(\|r\| is_drum_class(r.predicted))
745	756		.count();
746		-	let drum_acc = drum_correct as f64 / total_classified as f64 * 100.0;
	757	+	let drum_acc = if drum_expected.is_empty() {
	758	+	0.0
	759	+	} else {
	760	+	drum_correct as f64 / drum_expected.len() as f64 * 100.0
	761	+	};
747	762
748	763		println!(" Overall:");
749	764		println!(
750	765		" Strict accuracy (exact class match): {strict_acc:.1}% ({strict_correct}/{total_classified})"
751	766		);
752	767		println!(
753		-	" Layer 1 accuracy (drum detection): {drum_acc:.1}% ({drum_correct}/{total_classified})"
	768	+	" Layer 1 accuracy (drum detection): {drum_acc:.1}% ({drum_correct}/{})",
	769	+	drum_expected.len()
754	770		);
755	771		println!();
756	772