max / audiofiles
- Co-Authored-By
- Claude Opus 5 (1M context) <noreply@anthropic.com>
9 files changed,
+566 insertions,
-48 deletions
| @@ -4246,6 +4246,7 @@ | |||
| 4246 | 4246 | [[package]] | |
| 4247 | 4247 | name = "quasi-immediate" | |
| 4248 | 4248 | version = "0.48.0" | |
| 4249 | + | source = "git+https://makenot.work/git/max/quasi.git#b23c0c0f34323a65a611dc5fda74533b0ec5c3c9" | |
| 4249 | 4250 | dependencies = [ | |
| 4250 | 4251 | "docengine", | |
| 4251 | 4252 | "egui", | |
| @@ -4256,6 +4257,7 @@ | |||
| 4256 | 4257 | [[package]] | |
| 4257 | 4258 | name = "quasi-router" | |
| 4258 | 4259 | version = "0.48.0" | |
| 4260 | + | source = "git+https://makenot.work/git/max/quasi.git#b23c0c0f34323a65a611dc5fda74533b0ec5c3c9" | |
| 4259 | 4261 | dependencies = [ | |
| 4260 | 4262 | "makeover-layout", | |
| 4261 | 4263 | ] | |
| @@ -5687,7 +5689,7 @@ | |||
| 5687 | 5689 | ||
| 5688 | 5690 | [[package]] | |
| 5689 | 5691 | name = "tagtree" | |
| 5690 | - | version = "0.4.0" | |
| 5692 | + | version = "0.4.1" | |
| 5691 | 5693 | ||
| 5692 | 5694 | [[package]] | |
| 5693 | 5695 | name = "target-lexicon" | |
| @@ -7545,33 +7547,9 @@ | |||
| 7545 | 7547 | ] | |
| 7546 | 7548 | ||
| 7547 | 7549 | [[patch.unused]] | |
| 7548 | - | name = "quasi-axum" | |
| 7549 | - | version = "0.48.0" | |
| 7550 | - | ||
| 7551 | - | [[patch.unused]] | |
| 7552 | - | name = "quasi-basics" | |
| 7553 | - | version = "0.48.0" | |
| 7554 | - | ||
| 7555 | - | [[patch.unused]] | |
| 7556 | - | name = "quasi-http" | |
| 7557 | - | version = "0.48.0" | |
| 7558 | - | ||
| 7559 | - | [[patch.unused]] | |
| 7560 | - | name = "quasi-notifs" | |
| 7561 | - | version = "0.48.0" | |
| 7562 | - | ||
| 7563 | - | [[patch.unused]] | |
| 7564 | - | name = "quasi-store" | |
| 7550 | + | name = "quasi-type" | |
| 7565 | 7551 | version = "0.1.0" | |
| 7566 | 7552 | ||
| 7567 | - | [[patch.unused]] | |
| 7568 | - | name = "quasi-tauri" | |
| 7569 | - | version = "0.48.0" | |
| 7570 | - | ||
| 7571 | - | [[patch.unused]] | |
| 7572 | - | name = "quasi-webview" | |
| 7573 | - | version = "0.48.0" | |
| 7574 | - | ||
| 7575 | 7553 | [[patch.unused]] | |
| 7576 | 7554 | name = "kberg" | |
| 7577 | 7555 | version = "0.1.0" | |
| @@ -7585,5 +7563,37 @@ | |||
| 7585 | 7563 | version = "0.1.0" | |
| 7586 | 7564 | ||
| 7587 | 7565 | [[patch.unused]] | |
| 7588 | - | name = "quasi-type" | |
| 7566 | + | name = "quasi-axum" | |
| 7567 | + | version = "0.49.0" | |
| 7568 | + | ||
| 7569 | + | [[patch.unused]] | |
| 7570 | + | name = "quasi-basics" | |
| 7571 | + | version = "0.49.0" | |
| 7572 | + | ||
| 7573 | + | [[patch.unused]] | |
| 7574 | + | name = "quasi-http" | |
| 7575 | + | version = "0.49.0" | |
| 7576 | + | ||
| 7577 | + | [[patch.unused]] | |
| 7578 | + | name = "quasi-immediate" | |
| 7579 | + | version = "0.49.0" | |
| 7580 | + | ||
| 7581 | + | [[patch.unused]] | |
| 7582 | + | name = "quasi-notifs" | |
| 7583 | + | version = "0.49.0" | |
| 7584 | + | ||
| 7585 | + | [[patch.unused]] | |
| 7586 | + | name = "quasi-router" | |
| 7587 | + | version = "0.49.0" | |
| 7588 | + | ||
| 7589 | + | [[patch.unused]] | |
| 7590 | + | name = "quasi-store" | |
| 7589 | 7591 | version = "0.1.0" | |
| 7592 | + | ||
| 7593 | + | [[patch.unused]] | |
| 7594 | + | name = "quasi-tauri" | |
| 7595 | + | version = "0.49.0" | |
| 7596 | + | ||
| 7597 | + | [[patch.unused]] | |
| 7598 | + | name = "quasi-webview" | |
| 7599 | + | version = "0.49.0" |
| @@ -68,7 +68,7 @@ | |||
| 68 | 68 | .map(|h| ClassifierJobResult::Trained(h.map(|head| head_info(&head)))), | |
| 69 | 69 | ClassifierJob::AutoApply { k } => auto_apply(db, k).map(ClassifierJobResult::AutoApplied), | |
| 70 | 70 | ClassifierJob::Cluster { k } => { | |
| 71 | - | cluster::cluster_library(db, k, 50).map(|r| ClassifierJobResult::Clustered(r.clusters)) | |
| 71 | + | cluster::cluster_and_persist(db, k, 50).map(ClassifierJobResult::Clustered) | |
| 72 | 72 | } | |
| 73 | 73 | ClassifierJob::Export { path, opts } => { | |
| 74 | 74 | afcl::export_to_path(db, &path, &opts).map(|()| ClassifierJobResult::Exported(path)) |
| @@ -1827,6 +1827,49 @@ | |||
| 1827 | 1827 | END; | |
| 1828 | 1828 | "; | |
| 1829 | 1829 | ||
| 1830 | + | const MIGRATION_039: &str = r" | |
| 1831 | + | -- Clusters become first-class named objects. | |
| 1832 | + | -- | |
| 1833 | + | -- Before this, `cluster_library` produced a grouping and then threw it away: | |
| 1834 | + | -- `apply_cluster_tag` wrote tags and the pile itself stopped existing. The | |
| 1835 | + | -- grouping was the valuable part. A persisted cluster is what answers 'what is | |
| 1836 | + | -- this pile' for the population the filename rules cannot answer at all. | |
| 1837 | + | -- | |
| 1838 | + | -- Membership is derived, on the `sample_neighbours` model in M038: it is | |
| 1839 | + | -- recomputable from `sample_features`, and machine-dependent besides, because | |
| 1840 | + | -- k-means runs over features standardized against this library's own ranges. | |
| 1841 | + | -- So neither table carries sync triggers. The user's name is NOT derived, and | |
| 1842 | + | -- whether it should sync is a separate decision, filed rather than answered | |
| 1843 | + | -- here. | |
| 1844 | + | CREATE TABLE IF NOT EXISTS clusters ( | |
| 1845 | + | id INTEGER PRIMARY KEY, | |
| 1846 | + | -- NULL until the user names it. Naming is the whole point; an unnamed | |
| 1847 | + | -- cluster is a pile still waiting for a word. | |
| 1848 | + | name TEXT, | |
| 1849 | + | -- The member nearest the centroid: a playable representative, and the key a | |
| 1850 | + | -- re-run matches on to carry the name across. Nullable because deleting the | |
| 1851 | + | -- representative sample must not destroy the name the user typed. | |
| 1852 | + | medoid_hash TEXT REFERENCES samples(hash) ON DELETE SET NULL, | |
| 1853 | + | -- The run that produced it: the k asked for, the feature extractor the | |
| 1854 | + | -- vectors came from, and when. A cluster built under a stale | |
| 1855 | + | -- `feature_version` is comparable to nothing built since. | |
| 1856 | + | k INTEGER NOT NULL, | |
| 1857 | + | feature_version INTEGER NOT NULL, | |
| 1858 | + | run_at INTEGER NOT NULL, | |
| 1859 | + | named_at INTEGER | |
| 1860 | + | ); | |
| 1861 | + | CREATE INDEX IF NOT EXISTS idx_clusters_medoid ON clusters(medoid_hash); | |
| 1862 | + | ||
| 1863 | + | CREATE TABLE IF NOT EXISTS cluster_members ( | |
| 1864 | + | cluster_id INTEGER NOT NULL REFERENCES clusters(id) ON DELETE CASCADE, | |
| 1865 | + | sample_hash TEXT NOT NULL REFERENCES samples(hash) ON DELETE CASCADE, | |
| 1866 | + | PRIMARY KEY (cluster_id, sample_hash) | |
| 1867 | + | ); | |
| 1868 | + | -- 'which pile is this sample in' is the read the detail pane makes, and it runs | |
| 1869 | + | -- against the grain of the primary key. | |
| 1870 | + | CREATE INDEX IF NOT EXISTS idx_cluster_members_hash ON cluster_members(sample_hash); | |
| 1871 | + | "; | |
| 1872 | + | ||
| 1830 | 1873 | /// Register `hash_row_id(salt, key) -> TEXT` as a deterministic SQLite | |
| 1831 | 1874 | /// function on the given connection. Used by the M018 sync triggers so the | |
| 1832 | 1875 | /// `sync_changelog.row_id` field never carries cleartext content (tag strings, | |
| @@ -1904,6 +1947,7 @@ | |||
| 1904 | 1947 | MIGRATION_036, | |
| 1905 | 1948 | MIGRATION_037, | |
| 1906 | 1949 | MIGRATION_038, | |
| 1950 | + | MIGRATION_039, | |
| 1907 | 1951 | ]; | |
| 1908 | 1952 | ||
| 1909 | 1953 | /// The schema version this build produces, and the highest one it can read. | |
| @@ -2486,6 +2530,8 @@ | |||
| 2486 | 2530 | "classifier_exemplars", | |
| 2487 | 2531 | "classifier_layer_rules", | |
| 2488 | 2532 | "classifier_layers", | |
| 2533 | + | "cluster_members", | |
| 2534 | + | "clusters", | |
| 2489 | 2535 | "collection_members", | |
| 2490 | 2536 | "collections", | |
| 2491 | 2537 | "config_key_policy", | |
| @@ -2519,7 +2565,7 @@ | |||
| 2519 | 2565 | .conn() | |
| 2520 | 2566 | .query_row("PRAGMA user_version", [], |row| row.get(0)) | |
| 2521 | 2567 | .unwrap(); | |
| 2522 | - | assert_eq!(version, 38); | |
| 2568 | + | assert_eq!(version, SCHEMA_VERSION); | |
| 2523 | 2569 | } | |
| 2524 | 2570 | ||
| 2525 | 2571 | #[test] | |
| @@ -2530,7 +2576,7 @@ | |||
| 2530 | 2576 | .conn() | |
| 2531 | 2577 | .query_row("PRAGMA user_version", [], |row| row.get(0)) | |
| 2532 | 2578 | .unwrap(); | |
| 2533 | - | assert_eq!(version, 38); | |
| 2579 | + | assert_eq!(version, SCHEMA_VERSION); | |
| 2534 | 2580 | } | |
| 2535 | 2581 | ||
| 2536 | 2582 | #[test] | |
| @@ -2714,7 +2760,7 @@ | |||
| 2714 | 2760 | .conn() | |
| 2715 | 2761 | .query_row("PRAGMA user_version", [], |row| row.get(0)) | |
| 2716 | 2762 | .unwrap(); | |
| 2717 | - | assert_eq!(version, 38); | |
| 2763 | + | assert_eq!(version, SCHEMA_VERSION); | |
| 2718 | 2764 | } | |
| 2719 | 2765 | ||
| 2720 | 2766 | /// Simulates the worst-case recovery path: a prior partial migration left | |
| @@ -2758,7 +2804,7 @@ | |||
| 2758 | 2804 | .conn() | |
| 2759 | 2805 | .query_row("PRAGMA user_version", [], |row| row.get(0)) | |
| 2760 | 2806 | .unwrap(); | |
| 2761 | - | assert_eq!(version, 38); | |
| 2807 | + | assert_eq!(version, SCHEMA_VERSION); | |
| 2762 | 2808 | } | |
| 2763 | 2809 | ||
| 2764 | 2810 | /// A vault written by a newer audiofiles is refused, not opened. | |
| @@ -2824,7 +2870,7 @@ | |||
| 2824 | 2870 | /// means adding a migration without meaning to shows up as a failure. | |
| 2825 | 2871 | #[test] | |
| 2826 | 2872 | fn schema_version_matches_the_migration_list() { | |
| 2827 | - | assert_eq!(SCHEMA_VERSION, 38); | |
| 2873 | + | assert_eq!(SCHEMA_VERSION, 39); | |
| 2828 | 2874 | assert_eq!(SCHEMA_VERSION as usize, MIGRATIONS.len()); | |
| 2829 | 2875 | } | |
| 2830 | 2876 | ||
| @@ -3160,7 +3206,7 @@ | |||
| 3160 | 3206 | let initial_version: i32 = conn | |
| 3161 | 3207 | .query_row("PRAGMA user_version", [], |row| row.get(0)) | |
| 3162 | 3208 | .unwrap(); | |
| 3163 | - | assert_eq!(initial_version, 38); | |
| 3209 | + | assert_eq!(initial_version, SCHEMA_VERSION); | |
| 3164 | 3210 | ||
| 3165 | 3211 | let batch = format!("BEGIN;\n{bad_sql}\nPRAGMA user_version = 999;\nCOMMIT;"); | |
| 3166 | 3212 | let first_err = conn.execute_batch(&batch).unwrap_err(); | |
| @@ -3225,7 +3271,7 @@ | |||
| 3225 | 3271 | .conn() | |
| 3226 | 3272 | .query_row("PRAGMA user_version", [], |row| row.get(0)) | |
| 3227 | 3273 | .unwrap(); | |
| 3228 | - | assert_eq!(version, 38); | |
| 3274 | + | assert_eq!(version, SCHEMA_VERSION); | |
| 3229 | 3275 | } | |
| 3230 | 3276 | ||
| 3231 | 3277 | #[test] |
| @@ -301,7 +301,7 @@ | |||
| 301 | 301 | pub enum ClassifierJobResult { | |
| 302 | 302 | Trained(Option<TrainedHeadInfo>), | |
| 303 | 303 | AutoApplied(usize), | |
| 304 | - | Clustered(Vec<audiofiles_core::analysis::cluster::Cluster>), | |
| 304 | + | Clustered(Vec<audiofiles_core::analysis::cluster::PersistedCluster>), | |
| 305 | 305 | Exported(std::path::PathBuf), | |
| 306 | 306 | /// Suggestions for the sample identified by `hash`. The hash lets the GUI | |
| 307 | 307 | /// drop the result if the selection changed while the job ran. | |
| @@ -639,11 +639,20 @@ | |||
| 639 | 639 | /// Set a layer's k-NN weight (clamped to `[0, 1]`). | |
| 640 | 640 | fn set_classifier_layer_weight(&self, id: &str, weight: f64) -> BackendResult<()>; | |
| 641 | 641 | ||
| 642 | - | /// Cluster the library's feature vectors into `k` groups (for naming/seeding). | |
| 642 | + | /// Cluster the library's feature vectors into `k` groups and persist them, | |
| 643 | + | /// carrying the names of the previous run across. | |
| 643 | 644 | fn cluster_library( | |
| 644 | 645 | &self, | |
| 645 | 646 | k: usize, | |
| 646 | - | ) -> BackendResult<Vec<audiofiles_core::analysis::cluster::Cluster>>; | |
| 647 | + | ) -> BackendResult<Vec<audiofiles_core::analysis::cluster::PersistedCluster>>; | |
| 648 | + | ||
| 649 | + | /// The stored clusters, so a session opens on the piles the user already named. | |
| 650 | + | fn load_clusters( | |
| 651 | + | &self, | |
| 652 | + | ) -> BackendResult<Vec<audiofiles_core::analysis::cluster::PersistedCluster>>; | |
| 653 | + | ||
| 654 | + | /// Name a stored cluster, or clear the name if `name` is blank. | |
| 655 | + | fn rename_cluster(&self, id: i64, name: &str) -> BackendResult<()>; | |
| 647 | 656 | ||
| 648 | 657 | /// Apply a tag to every member of a named cluster (`source = 'cluster'`). | |
| 649 | 658 | fn apply_cluster_tag(&self, members: &[String], tag: &str) -> BackendResult<usize>; |
| @@ -518,8 +518,12 @@ | |||
| 518 | 518 | ); | |
| 519 | 519 | } | |
| 520 | 520 | R::Clustered(clusters) => { | |
| 521 | - | self.classifier.cluster_names = vec![String::new(); clusters.len()]; | |
| 521 | + | self.classifier.cluster_names = clusters | |
| 522 | + | .iter() | |
| 523 | + | .map(|c| c.name.clone().unwrap_or_default()) | |
| 524 | + | .collect(); | |
| 522 | 525 | self.classifier.clusters = clusters; | |
| 526 | + | self.classifier.clusters_loaded = true; | |
| 523 | 527 | self.status = format!( | |
| 524 | 528 | "Found {} cluster{}.", | |
| 525 | 529 | self.classifier.clusters.len(), | |
| @@ -875,6 +879,44 @@ | |||
| 875 | 879 | ); | |
| 876 | 880 | } | |
| 877 | 881 | ||
| 882 | + | /// Read the stored clusters once per session, so the piles the user already | |
| 883 | + | /// named are there before any re-run. | |
| 884 | + | pub fn ensure_clusters_loaded(&mut self) { | |
| 885 | + | if self.classifier.clusters_loaded { | |
| 886 | + | return; | |
| 887 | + | } | |
| 888 | + | self.classifier.clusters_loaded = true; | |
| 889 | + | let clusters = self.backend.load_clusters().unwrap_or_default(); | |
| 890 | + | self.classifier.cluster_names = clusters | |
| 891 | + | .iter() | |
| 892 | + | .map(|c| c.name.clone().unwrap_or_default()) | |
| 893 | + | .collect(); | |
| 894 | + | self.classifier.clusters = clusters; | |
| 895 | + | } | |
| 896 | + | ||
| 897 | + | /// Persist the edited name of cluster `idx`. Naming is the authored part of a | |
| 898 | + | /// cluster, so it is written through rather than held in the widget buffer. | |
| 899 | + | pub fn commit_cluster_name(&mut self, idx: usize) { | |
| 900 | + | let Some(cluster) = self.classifier.clusters.get(idx) else { | |
| 901 | + | return; | |
| 902 | + | }; | |
| 903 | + | let name = self | |
| 904 | + | .classifier | |
| 905 | + | .cluster_names | |
| 906 | + | .get(idx) | |
| 907 | + | .map(|n| n.trim().to_string()) | |
| 908 | + | .unwrap_or_default(); | |
| 909 | + | let id = cluster.id; | |
| 910 | + | match self.backend.rename_cluster(id, &name) { | |
| 911 | + | Ok(()) => { | |
| 912 | + | if let Some(cluster) = self.classifier.clusters.get_mut(idx) { | |
| 913 | + | cluster.name = if name.is_empty() { None } else { Some(name) }; | |
| 914 | + | } | |
| 915 | + | } | |
| 916 | + | Err(e) => self.status = format!("Could not name cluster: {e}"), | |
| 917 | + | } | |
| 918 | + | } | |
| 919 | + | ||
| 878 | 920 | /// Apply the named tag to every member of cluster `idx`. | |
| 879 | 921 | pub fn apply_cluster(&mut self, idx: usize) { | |
| 880 | 922 | let Some(cluster) = self.classifier.clusters.get(idx) else { | |
| @@ -892,6 +934,9 @@ | |||
| 892 | 934 | self.status = "Name the cluster before tagging it.".to_string(); | |
| 893 | 935 | return; | |
| 894 | 936 | } | |
| 937 | + | // Tagging a pile is also naming it: the word the user typed is the | |
| 938 | + | // cluster's name whether or not they pressed anything else first. | |
| 939 | + | self.commit_cluster_name(idx); | |
| 895 | 940 | match self.backend.apply_cluster_tag(&members, &tag) { | |
| 896 | 941 | Ok(n) => { | |
| 897 | 942 | self.status = format!( |
| @@ -379,10 +379,13 @@ | |||
| 379 | 379 | // --- Clustering bootstrap --- | |
| 380 | 380 | /// Requested cluster count; 0 = auto (suggest_k). | |
| 381 | 381 | pub cluster_k: u32, | |
| 382 | - | /// Last clustering result. | |
| 383 | - | pub clusters: Vec<audiofiles_core::analysis::cluster::Cluster>, | |
| 384 | - | /// Editable name (tag) per cluster, parallel to `clusters`. | |
| 382 | + | /// The stored clusters, as of the last run or load. | |
| 383 | + | pub clusters: Vec<audiofiles_core::analysis::cluster::PersistedCluster>, | |
| 384 | + | /// Editable name per cluster, parallel to `clusters`. Seeded from the stored | |
| 385 | + | /// name, so a session opens on the words the user already typed. | |
| 385 | 386 | pub cluster_names: Vec<String>, | |
| 387 | + | /// Whether the stored clusters have been read this session. | |
| 388 | + | pub clusters_loaded: bool, | |
| 386 | 389 | ||
| 387 | 390 | // --- Folder-label harvest --- | |
| 388 | 391 | pub folder_labels: Vec<audiofiles_core::harvest::FolderLabel>, |
| @@ -843,6 +843,9 @@ | |||
| 843 | 843 | .color(theme::content_muted()), | |
| 844 | 844 | ); | |
| 845 | 845 | ui.add_space(theme::space::bound()); | |
| 846 | + | // The stored piles come back before anything is re-run, so a session | |
| 847 | + | // opens on the words the user already typed. | |
| 848 | + | state.ensure_clusters_loaded(); | |
| 846 | 849 | let busy = state.classifier.busy.is_some(); | |
| 847 | 850 | ui.horizontal(|ui| { | |
| 848 | 851 | let mut k = state.classifier.cluster_k.max(2); | |
| @@ -884,6 +887,7 @@ | |||
| 884 | 887 | let clusters = std::mem::take(&mut state.classifier.clusters); | |
| 885 | 888 | let mut play: Option<String> = None; | |
| 886 | 889 | let mut apply: Option<usize> = None; | |
| 890 | + | let mut rename: Option<usize> = None; | |
| 887 | 891 | for (i, cluster) in clusters.iter().enumerate() { | |
| 888 | 892 | ui.horizontal(|ui| { | |
| 889 | 893 | ui.label( | |
| @@ -891,20 +895,28 @@ | |||
| 891 | 895 | .small() | |
| 892 | 896 | .color(theme::content_secondary()), | |
| 893 | 897 | ); | |
| 898 | + | // The representative can be gone: deleting a sample must not | |
| 899 | + | // take the pile or its name with it. | |
| 900 | + | let medoid = cluster.medoid_hash.clone(); | |
| 894 | 901 | if ui | |
| 895 | - | .small_button("Play") | |
| 902 | + | .add_enabled(medoid.is_some(), egui::Button::new("Play").small()) | |
| 896 | 903 | .on_hover_text("Preview a representative sample") | |
| 897 | 904 | .clicked() | |
| 898 | 905 | { | |
| 899 | - | play = Some(cluster.medoid_hash.clone()); | |
| 906 | + | play = medoid; | |
| 900 | 907 | } | |
| 901 | 908 | if let Some(name) = state.classifier.cluster_names.get_mut(i) { | |
| 902 | - | widgets::text_field( | |
| 909 | + | let response = widgets::text_field( | |
| 903 | 910 | ui, | |
| 904 | 911 | egui::TextEdit::singleline(name) | |
| 905 | 912 | .desired_width(150.0) | |
| 906 | - | .hint_text("tag for this group"), | |
| 913 | + | .hint_text("name this group"), | |
| 907 | 914 | ); | |
| 915 | + | // Written through on commit, not per keystroke: the name | |
| 916 | + | // outlives the widget and the run that produced the pile. | |
| 917 | + | if response.lost_focus() || response.changed() && !response.has_focus() { | |
| 918 | + | rename = Some(i); | |
| 919 | + | } | |
| 908 | 920 | } | |
| 909 | 921 | if ui.small_button("Tag").clicked() { | |
| 910 | 922 | apply = Some(i); | |
| @@ -915,6 +927,9 @@ | |||
| 915 | 927 | if let Some(hash) = play { | |
| 916 | 928 | state.trigger_preview(&hash); | |
| 917 | 929 | } | |
| 930 | + | if let Some(i) = rename { | |
| 931 | + | state.commit_cluster_name(i); | |
| 932 | + | } | |
| 918 | 933 | if let Some(i) = apply { | |
| 919 | 934 | state.apply_cluster(i); | |
| 920 | 935 | } |
| @@ -11,6 +11,7 @@ | |||
| 11 | 11 | ||
| 12 | 12 | use crate::db::Database; | |
| 13 | 13 | use crate::error::Result; | |
| 14 | + | use rusqlite::OptionalExtension; | |
| 14 | 15 | use tracing::instrument; | |
| 15 | 16 | ||
| 16 | 17 | use super::features::{FEATURE_VERSION, NUM_FEATURES}; | |
| @@ -227,6 +228,220 @@ | |||
| 227 | 228 | }) | |
| 228 | 229 | } | |
| 229 | 230 | ||
| 231 | + | /// A cluster as it lives in the vault: the grouping, plus the name the user gave it. | |
| 232 | + | /// | |
| 233 | + | /// The unnamed form is not a failure state. A fresh run produces piles with no | |
| 234 | + | /// words on them yet, and naming one is the act this whole table exists for. | |
| 235 | + | #[derive(Debug, Clone, serde::Serialize, serde::Deserialize)] | |
| 236 | + | pub struct PersistedCluster { | |
| 237 | + | pub id: i64, | |
| 238 | + | /// `None` until the user names it. | |
| 239 | + | pub name: Option<String>, | |
| 240 | + | /// `None` only if the representative sample has since been deleted. A name | |
| 241 | + | /// outlives its medoid; the `ON DELETE SET NULL` in M039 is what says so. | |
| 242 | + | pub medoid_hash: Option<String>, | |
| 243 | + | pub member_hashes: Vec<String>, | |
| 244 | + | } | |
| 245 | + | ||
| 246 | + | fn now_secs() -> i64 { | |
| 247 | + | use std::time::{SystemTime, UNIX_EPOCH}; | |
| 248 | + | SystemTime::now() | |
| 249 | + | .duration_since(UNIX_EPOCH) | |
| 250 | + | .map_or(0, |d| d.as_secs() as i64) | |
| 251 | + | } | |
| 252 | + | ||
| 253 | + | /// How strongly a stored cluster claims to be the same pile as a fresh one. | |
| 254 | + | /// | |
| 255 | + | /// Two rungs, and the gap between them is deliberate. An exact medoid match is | |
| 256 | + | /// the same representative choosing the same pile, which is as close to identity | |
| 257 | + | /// as a re-clustering gets. Containment is weaker: the old representative landed | |
| 258 | + | /// in this pile, so this pile inherited it, but the pile may have absorbed or | |
| 259 | + | /// shed a lot around it. | |
| 260 | + | fn claim_strength(old_medoid: &str, fresh: &Cluster) -> u8 { | |
| 261 | + | if fresh.medoid_hash == old_medoid { | |
| 262 | + | 2 | |
| 263 | + | } else { | |
| 264 | + | // 1 for containment, 0 for no claim at all. | |
| 265 | + | u8::from(fresh.member_hashes.iter().any(|h| h == old_medoid)) | |
| 266 | + | } | |
| 267 | + | } | |
| 268 | + | ||
| 269 | + | /// Match fresh clusters to the stored named ones, so a re-run does not throw | |
| 270 | + | /// away the words the user typed. | |
| 271 | + | /// | |
| 272 | + | /// Returns `names[i]` for `fresh[i]`. Greedy over claim strength: every stored | |
| 273 | + | /// name is carried to at most one fresh cluster, and every fresh cluster takes | |
| 274 | + | /// at most one name. Ties resolve by cluster order, which k-means makes | |
| 275 | + | /// deterministic, so the same library re-clustered at the same `k` carries the | |
| 276 | + | /// same names every time. | |
| 277 | + | fn carry_names(fresh: &[Cluster], stored: &[(i64, String, String)]) -> Vec<Option<String>> { | |
| 278 | + | let mut claims: Vec<(u8, usize, usize)> = Vec::new(); | |
| 279 | + | for (si, (_, _, medoid)) in stored.iter().enumerate() { | |
| 280 | + | for (fi, f) in fresh.iter().enumerate() { | |
| 281 | + | let strength = claim_strength(medoid, f); | |
| 282 | + | if strength > 0 { | |
| 283 | + | claims.push((strength, fi, si)); | |
| 284 | + | } | |
| 285 | + | } | |
| 286 | + | } | |
| 287 | + | // Strongest first; then fresh order, then stored order, so the result does | |
| 288 | + | // not depend on how the rows came back from SQLite. | |
| 289 | + | claims.sort_by(|a, b| b.0.cmp(&a.0).then(a.1.cmp(&b.1)).then(a.2.cmp(&b.2))); | |
| 290 | + | ||
| 291 | + | let mut names: Vec<Option<String>> = vec![None; fresh.len()]; | |
| 292 | + | let mut used = vec![false; stored.len()]; | |
| 293 | + | for (_, fi, si) in claims { | |
| 294 | + | if names[fi].is_some() || used[si] { | |
| 295 | + | continue; | |
| 296 | + | } | |
| 297 | + | names[fi] = Some(stored[si].1.clone()); | |
| 298 | + | used[si] = true; | |
| 299 | + | } | |
| 300 | + | names | |
| 301 | + | } | |
| 302 | + | ||
| 303 | + | /// Replace the stored cluster set with `result`, carrying names across. | |
| 304 | + | /// | |
| 305 | + | /// The whole set is replaced rather than merged: a clustering run is a single | |
| 306 | + | /// answer over the whole library, so half of one run beside half of another is | |
| 307 | + | /// not a state the user could interpret. What survives a replacement is the | |
| 308 | + | /// naming, which is the only part the user authored. | |
| 309 | + | #[instrument(skip_all)] | |
| 310 | + | pub fn persist_clusters( | |
| 311 | + | db: &Database, | |
| 312 | + | result: &ClusterResult, | |
| 313 | + | k: usize, | |
| 314 | + | ) -> Result<Vec<PersistedCluster>> { | |
| 315 | + | let run_at = now_secs(); | |
| 316 | + | db.transaction_core(|_tx| { | |
| 317 | + | let stored: Vec<(i64, String, String)> = { | |
| 318 | + | let mut stmt = db.conn().prepare( | |
| 319 | + | "SELECT id, name, medoid_hash FROM clusters \ | |
| 320 | + | WHERE name IS NOT NULL AND medoid_hash IS NOT NULL ORDER BY id", | |
| 321 | + | )?; | |
| 322 | + | let rows = stmt.query_map([], |row| { | |
| 323 | + | Ok(( | |
| 324 | + | row.get::<_, i64>(0)?, | |
| 325 | + | row.get::<_, String>(1)?, | |
| 326 | + | row.get::<_, String>(2)?, | |
| 327 | + | )) | |
| 328 | + | })?; | |
| 329 | + | rows.collect::<std::result::Result<_, _>>()? | |
| 330 | + | }; | |
| 331 | + | let names = carry_names(&result.clusters, &stored); | |
| 332 | + | ||
| 333 | + | // Members go with the row: M039 cascades cluster_members off clusters. | |
| 334 | + | db.conn().execute("DELETE FROM clusters", [])?; | |
| 335 | + | ||
| 336 | + | let mut out = Vec::with_capacity(result.clusters.len()); | |
| 337 | + | for (fresh, name) in result.clusters.iter().zip(names) { | |
| 338 | + | db.conn().execute( | |
| 339 | + | "INSERT INTO clusters (name, medoid_hash, k, feature_version, run_at, named_at) \ | |
| 340 | + | VALUES (?1, ?2, ?3, ?4, ?5, ?6)", | |
| 341 | + | rusqlite::params![ | |
| 342 | + | name, | |
| 343 | + | fresh.medoid_hash, | |
| 344 | + | k as i64, | |
| 345 | + | result.feature_version, | |
| 346 | + | run_at, | |
| 347 | + | name.as_ref().map(|_| run_at), | |
| 348 | + | ], | |
| 349 | + | )?; | |
| 350 | + | let id = db.conn().last_insert_rowid(); | |
| 351 | + | for hash in &fresh.member_hashes { | |
| 352 | + | db.conn().execute( | |
| 353 | + | "INSERT OR IGNORE INTO cluster_members (cluster_id, sample_hash) VALUES (?1, ?2)", | |
| 354 | + | rusqlite::params![id, hash], | |
| 355 | + | )?; | |
| 356 | + | } | |
| 357 | + | out.push(PersistedCluster { | |
| 358 | + | id, | |
| 359 | + | name, | |
| 360 | + | medoid_hash: Some(fresh.medoid_hash.clone()), | |
| 361 | + | member_hashes: fresh.member_hashes.clone(), | |
| 362 | + | }); | |
| 363 | + | } | |
| 364 | + | Ok(out) | |
| 365 | + | }) | |
| 366 | + | } | |
| 367 | + | ||
| 368 | + | /// Cluster the library and persist the result in one step. | |
| 369 | + | /// | |
| 370 | + | /// The pairing is the point: a run that is not persisted is the behaviour this | |
| 371 | + | /// table replaced. | |
| 372 | + | #[instrument(skip_all)] | |
| 373 | + | pub fn cluster_and_persist( | |
| 374 | + | db: &Database, | |
| 375 | + | k: usize, | |
| 376 | + | max_iters: usize, | |
| 377 | + | ) -> Result<Vec<PersistedCluster>> { | |
| 378 | + | let result = cluster_library(db, k, max_iters)?; | |
| 379 | + | persist_clusters(db, &result, k) | |
| 380 | + | } | |
| 381 | + | ||
| 382 | + | /// The stored clusters, in id order, with their members. | |
| 383 | + | #[instrument(skip_all)] | |
| 384 | + | pub fn load_clusters(db: &Database) -> Result<Vec<PersistedCluster>> { | |
| 385 | + | let mut stmt = db | |
| 386 | + | .conn() | |
| 387 | + | .prepare("SELECT id, name, medoid_hash FROM clusters ORDER BY id")?; | |
| 388 | + | let heads: Vec<(i64, Option<String>, Option<String>)> = stmt | |
| 389 | + | .query_map([], |row| Ok((row.get(0)?, row.get(1)?, row.get(2)?)))? | |
| 390 | + | .collect::<std::result::Result<_, _>>()?; | |
| 391 | + | ||
| 392 | + | let mut members = db.conn().prepare( | |
| 393 | + | "SELECT sample_hash FROM cluster_members WHERE cluster_id = ?1 ORDER BY sample_hash", | |
| 394 | + | )?; | |
| 395 | + | let mut out = Vec::with_capacity(heads.len()); | |
| 396 | + | for (id, name, medoid_hash) in heads { | |
| 397 | + | let member_hashes: Vec<String> = members | |
| 398 | + | .query_map([id], |row| row.get(0))? | |
| 399 | + | .collect::<std::result::Result<_, _>>()?; | |
| 400 | + | out.push(PersistedCluster { | |
| 401 | + | id, | |
| 402 | + | name, | |
| 403 | + | medoid_hash, | |
| 404 | + | member_hashes, | |
| 405 | + | }); | |
| 406 | + | } | |
| 407 | + | Ok(out) | |
| 408 | + | } | |
| 409 | + | ||
| 410 | + | /// Name a stored cluster, or clear the name if `name` is blank. | |
| 411 | + | /// | |
| 412 | + | /// Naming does not tag: [`apply_cluster_tag`] is still what writes tags, and | |
| 413 | + | /// keeping them separate is what lets a user name a pile to think with before | |
| 414 | + | /// deciding the whole pile deserves the word on every member. | |
| 415 | + | #[instrument(skip_all)] | |
| 416 | + | pub fn rename_cluster(db: &Database, id: i64, name: &str) -> Result<()> { | |
| 417 | + | let trimmed = name.trim(); | |
| 418 | + | let name: Option<&str> = if trimmed.is_empty() { | |
| 419 | + | None | |
| 420 | + | } else { | |
| 421 | + | Some(trimmed) | |
| 422 | + | }; | |
| 423 | + | db.conn().execute( | |
| 424 | + | "UPDATE clusters SET name = ?1, named_at = ?2 WHERE id = ?3", | |
| 425 | + | rusqlite::params![name, name.map(|_| now_secs()), id], | |
| 426 | + | )?; | |
| 427 | + | Ok(()) | |
| 428 | + | } | |
| 429 | + | ||
| 430 | + | /// The cluster a sample belongs to, if the library has been clustered. | |
| 431 | + | #[instrument(skip_all)] | |
| 432 | + | pub fn cluster_for_sample(db: &Database, hash: &str) -> Result<Option<PersistedCluster>> { | |
| 433 | + | let id: Option<i64> = db | |
| 434 | + | .conn() | |
| 435 | + | .query_row( | |
| 436 | + | "SELECT cluster_id FROM cluster_members WHERE sample_hash = ?1", | |
| 437 | + | [hash], | |
| 438 | + | |row| row.get(0), | |
| 439 | + | ) | |
| 440 | + | .optional()?; | |
| 441 | + | let Some(id) = id else { return Ok(None) }; | |
| 442 | + | Ok(load_clusters(db)?.into_iter().find(|c| c.id == id)) | |
| 443 | + | } | |
| 444 | + | ||
| 230 | 445 | #[cfg(test)] | |
| 231 | 446 | mod tests { | |
| 232 | 447 | use super::*; | |
| @@ -256,6 +471,166 @@ | |||
| 256 | 471 | [base; NUM_FEATURES] | |
| 257 | 472 | } | |
| 258 | 473 | ||
| 474 | + | /// The library used by every persistence test: two tight groups far apart, so | |
| 475 | + | /// the grouping is not in question and only the bookkeeping is under test. | |
| 476 | + | fn two_group_db() -> Database { | |
| 477 | + | db_with_features(&[ | |
| 478 | + | ("a1", vec_at(0.0)), | |
| 479 | + | ("a2", vec_at(0.1)), | |
| 480 | + | ("a3", vec_at(-0.1)), | |
| 481 | + | ("b1", vec_at(100.0)), | |
| 482 | + | ("b2", vec_at(100.1)), | |
| 483 | + | ("b3", vec_at(99.9)), | |
| 484 | + | ]) | |
| 485 | + | } | |
| 486 | + | ||
| 487 | + | #[test] | |
| 488 | + | fn persisted_clusters_round_trip() { | |
| 489 | + | let db = two_group_db(); | |
| 490 | + | let persisted = cluster_and_persist(&db, 2, 50).unwrap(); | |
| 491 | + | assert_eq!(persisted.len(), 2); | |
| 492 | + | assert!(persisted.iter().all(|c| c.name.is_none())); | |
| 493 | + | ||
| 494 | + | let loaded = load_clusters(&db).unwrap(); | |
| 495 | + | assert_eq!(loaded.len(), 2); | |
| 496 | + | for (a, b) in persisted.iter().zip(&loaded) { | |
| 497 | + | assert_eq!(a.id, b.id); | |
| 498 | + | assert_eq!(a.medoid_hash, b.medoid_hash); | |
| 499 | + | let mut left = a.member_hashes.clone(); | |
| 500 | + | left.sort(); | |
| 501 | + | assert_eq!(left, b.member_hashes); | |
| 502 | + | } | |
| 503 | + | } | |
| 504 | + | ||
| 505 | + | #[test] | |
| 506 | + | fn renaming_persists_and_blank_clears() { | |
| 507 | + | let db = two_group_db(); | |
| 508 | + | let persisted = cluster_and_persist(&db, 2, 50).unwrap(); | |
| 509 | + | let id = persisted[0].id; | |
| 510 | + | ||
| 511 | + | rename_cluster(&db, id, " kicks ").unwrap(); | |
| 512 | + | let named = load_clusters(&db).unwrap(); | |
| 513 | + | assert_eq!(named[0].name.as_deref(), Some("kicks")); | |
| 514 | + | ||
| 515 | + | rename_cluster(&db, id, " ").unwrap(); | |
| 516 | + | assert!(load_clusters(&db).unwrap()[0].name.is_none()); | |
| 517 | + | } | |
| 518 | + | ||
| 519 | + | #[test] | |
| 520 | + | fn a_re_run_carries_the_name_across() { | |
| 521 | + | let db = two_group_db(); | |
| 522 | + | let first = cluster_and_persist(&db, 2, 50).unwrap(); | |
| 523 | + | // Name whichever pile holds a1. | |
| 524 | + | let target = first | |
| 525 | + | .iter() | |
| 526 | + | .find(|c| c.member_hashes.iter().any(|h| h == "a1")) | |
| 527 | + | .unwrap(); | |
| 528 | + | rename_cluster(&db, target.id, "kicks").unwrap(); | |
| 529 | + | ||
| 530 | + | let second = cluster_and_persist(&db, 2, 50).unwrap(); | |
| 531 | + | let carried = second | |
| 532 | + | .iter() | |
| 533 | + | .find(|c| c.member_hashes.iter().any(|h| h == "a1")) | |
| 534 | + | .unwrap(); | |
| 535 | + | assert_eq!(carried.name.as_deref(), Some("kicks")); | |
| 536 | + | // The other pile is untouched: a name is carried, never invented. | |
| 537 | + | assert_eq!(second.iter().filter(|c| c.name.is_some()).count(), 1); | |
| 538 | + | } | |
| 539 | + | ||
| 540 | + | #[test] | |
| 541 | + | fn re_run_replaces_the_set_rather_than_accumulating() { | |
| 542 | + | let db = two_group_db(); | |
| 543 | + | cluster_and_persist(&db, 2, 50).unwrap(); | |
| 544 | + | cluster_and_persist(&db, 2, 50).unwrap(); | |
| 545 | + | assert_eq!(load_clusters(&db).unwrap().len(), 2); | |
| 546 | + | let members: i64 = db | |
| 547 | + | .conn() | |
| 548 | + | .query_row("SELECT COUNT(*) FROM cluster_members", [], |r| r.get(0)) | |
| 549 | + | .unwrap(); | |
| 550 | + | assert_eq!(members, 6); | |
| 551 | + | } | |
| 552 | + | ||
| 553 | + | #[test] | |
| 554 | + | fn containment_carries_a_name_when_the_medoid_moves() { | |
| 555 | + | // Old representative "b1" is not the fresh medoid, but is in the pile. | |
| 556 | + | let fresh = vec![Cluster { | |
| 557 | + | id: 0, | |
| 558 | + | medoid_hash: "b2".to_string(), | |
| 559 | + | member_hashes: vec!["b1".to_string(), "b2".to_string()], | |
| 560 | + | }]; | |
| 561 | + | let stored = vec![(1, "hats".to_string(), "b1".to_string())]; | |
| 562 | + | assert_eq!(carry_names(&fresh, &stored), vec![Some("hats".to_string())]); | |
| 563 | + | } | |
| 564 | + | ||
| 565 | + | #[test] | |
| 566 | + | fn one_stored_name_reaches_only_one_fresh_cluster() { | |
| 567 | + | // Both fresh piles can claim "b1": one by exact medoid, one by | |
| 568 | + | // containment. The exact match wins and the weaker claim gets nothing. | |
| 569 | + | let fresh = vec![ | |
| 570 | + | Cluster { | |
| 571 | + | id: 0, | |
| 572 | + | medoid_hash: "x".to_string(), | |
| 573 | + | member_hashes: vec!["x".to_string(), "b1".to_string()], | |
| 574 | + | }, | |
| 575 | + | Cluster { | |
| 576 | + | id: 1, | |
| 577 | + | medoid_hash: "b1".to_string(), | |
| 578 | + | member_hashes: vec!["b1".to_string()], | |
| 579 | + | }, | |
| 580 | + | ]; | |
| 581 | + | let stored = vec![(1, "hats".to_string(), "b1".to_string())]; | |
| 582 | + | assert_eq!( | |
| 583 | + | carry_names(&fresh, &stored), | |
| 584 | + | vec![None, Some("hats".to_string())] | |
| 585 | + | ); | |
| 586 | + | } | |
| 587 | + | ||
| 588 | + | #[test] | |
| 589 | + | fn an_unmatched_pile_carries_no_name() { | |
| 590 | + | let fresh = vec![Cluster { | |
| 591 | + | id: 0, | |
| 592 | + | medoid_hash: "z1".to_string(), | |
| 593 | + | member_hashes: vec!["z1".to_string()], | |
| 594 | + | }]; | |
| 595 | + | let stored = vec![(1, "hats".to_string(), "b1".to_string())]; | |
| 596 | + | assert_eq!(carry_names(&fresh, &stored), vec![None]); | |
| 597 | + | } | |
| 598 | + | ||
| 599 | + | #[test] | |
| 600 | + | fn deleting_a_sample_drops_its_membership_but_not_the_name() { | |
| 601 | + | let db = two_group_db(); | |
| 602 | + | let persisted = cluster_and_persist(&db, 2, 50).unwrap(); | |
| 603 | + | let target = persisted | |
| 604 | + | .iter() | |
| 605 | + | .find(|c| c.medoid_hash.as_deref() == Some("a1")) | |
| 606 | + | .or_else(|| persisted.first()) | |
| 607 | + | .unwrap(); | |
| 608 | + | let medoid = target.medoid_hash.clone().unwrap(); | |
| 609 | + | rename_cluster(&db, target.id, "kicks").unwrap(); | |
| 610 | + | ||
| 611 | + | db.conn() | |
| 612 | + | .execute("DELETE FROM samples WHERE hash = ?1", [&medoid]) | |
| 613 | + | .unwrap(); | |
| 614 | + | ||
| 615 | + | let after = load_clusters(&db) | |
| 616 | + | .unwrap() | |
| 617 | + | .into_iter() | |
| 618 | + | .find(|c| c.id == target.id) | |
| 619 | + | .expect("the cluster row outlives its medoid"); | |
| 620 | + | assert_eq!(after.name.as_deref(), Some("kicks")); | |
| 621 | + | assert_eq!(after.medoid_hash, None); | |
| 622 | + | assert!(!after.member_hashes.contains(&medoid)); | |
| 623 | + | } | |
| 624 | + | ||
| 625 | + | #[test] | |
| 626 | + | fn cluster_for_sample_finds_the_pile() { | |
| 627 | + | let db = two_group_db(); | |
| 628 | + | cluster_and_persist(&db, 2, 50).unwrap(); | |
| 629 | + | let found = cluster_for_sample(&db, "a1").unwrap().unwrap(); | |
| 630 | + | assert!(found.member_hashes.iter().any(|h| h == "a1")); | |
| 631 | + | assert!(cluster_for_sample(&db, "nope").unwrap().is_none()); | |
| 632 | + | } | |
| 633 | + | ||
| 259 | 634 | #[test] | |
| 260 | 635 | fn suggest_k_is_sane() { | |
| 261 | 636 | assert_eq!(suggest_k(0), 0); |
| @@ -176,9 +176,24 @@ | |||
| 176 | 176 | fn cluster_library( | |
| 177 | 177 | &self, | |
| 178 | 178 | k: usize, | |
| 179 | - | ) -> BackendResult<Vec<audiofiles_core::analysis::cluster::Cluster>> { | |
| 179 | + | ) -> BackendResult<Vec<audiofiles_core::analysis::cluster::PersistedCluster>> { | |
| 180 | 180 | let db = self.db.lock(); | |
| 181 | - | Ok(audiofiles_core::analysis::cluster::cluster_library(&db, k, 50)?.clusters) | |
| 181 | + | Ok(audiofiles_core::analysis::cluster::cluster_and_persist( | |
| 182 | + | &db, k, 50, | |
| 183 | + | )?) | |
| 184 | + | } | |
| 185 | + | ||
| 186 | + | fn load_clusters( | |
| 187 | + | &self, | |
| 188 | + | ) -> BackendResult<Vec<audiofiles_core::analysis::cluster::PersistedCluster>> { | |
| 189 | + | let db = self.db.lock(); | |
| 190 | + | Ok(audiofiles_core::analysis::cluster::load_clusters(&db)?) | |
| 191 | + | } | |
| 192 | + | ||
| 193 | + | fn rename_cluster(&self, id: i64, name: &str) -> BackendResult<()> { | |
| 194 | + | let db = self.db.lock(); | |
| 195 | + | audiofiles_core::analysis::cluster::rename_cluster(&db, id, name)?; | |
| 196 | + | Ok(()) | |
| 182 | 197 | } | |
| 183 | 198 | ||
| 184 | 199 | fn apply_cluster_tag(&self, members: &[String], tag: &str) -> BackendResult<usize> { |