Skip to content

Releases

TimeNet 0.12.0

Choose a tag to compare

@masquare masquare released this 11 Nov 15:00
f8ea276

TimeNet 0.12.0 is a breaking release. It reads only TimeF format version 2 with control schema version 2, so datasets written by 0.11.0 or by earlier development builds must be rebuilt.

timenet 0.12.0

timenet 0.12.0

Features

  • Breaking: format: Start the readable format range at TimeNet 0.12 (#776)
  • Breaking: registry: Remote registries refuse releases built from locally modified code (#685)
  • Breaking: manifest: Record TimeNet git provenance in each written dataset (#684)
  • Breaking: hints: Give every error one hint for the SDK and one for the CLI (#698)
  • Breaking: cache: Record where each release in the local registry comes from (#674)
  • Breaking: client: List and search every source (#673)
  • Breaking: client: Replace get() with get_manifest() and get_card() (#672)
  • Breaking: build: Build into the local registry, publish with --publish-to (#671)
  • Breaking: Add timenet.build() and publish() (#670)
  • Breaking: cli: Take --local-registry, --remote-registry and --auto-build (#669)
  • Breaking: config: Make the default list of sources from three variables (#668)
  • Breaking: client: Download releases into the local registry (#662)
  • Breaking: client: Take an ordered list of sources (#661)
  • Breaking: client: Read releases through the coordinator (#660)
  • Breaking: coordinator: Plan and run reads through the list of sources (#659)
  • Breaking: coordinator: Plan parents through an ordered list of sources (#658)
  • Breaking: sources: Resolve a release through an ordered list of sources (#657)
  • Breaking: build: Run every build through the coordinator (#655)
  • Breaking: build: Move the build CLI and build() onto the coordinator (#654)
  • Breaking: build: Build one prepared release in the built-in backend (#653)
  • Breaking: coordinator: Plan and run builds through build_release (#652)
  • Breaking: registry: Copy a release between registries (#649)
  • Breaking: registry: Publish a local release with put_release (#648)
  • Breaking: types: Require a label set on classification tasks and remove target_schema (#664)
  • Breaking: format: Give classification tasks a shared label set (#643)
  • Breaking: build: Expose parent preparation in the CLI and builder (#540)
  • Breaking: build: Plan and prepare exact parent releases (#541)
  • Breaking: dataset: Import single Signals into partial copies (#497)
  • Breaking: client: Open datasets as lazy readers and load them whole (#520)
  • Breaking: reader: Read objects by ID with typed get and get_many (#496)
  • Breaking: storage: Share payloads through prepared dataset releases (#494)
  • Breaking: storage: Store irregular time offsets in Parquet payloads (#561)
  • Breaking: storage: Pack large text values in Parquet payloads (#560)
  • Breaking: dataset: Hold signal values behind array handles (#586)
  • Breaking: types: Give text values an explicit read interface (#574)
  • Breaking: Make entity ids int64 and number them per dataset (#523)
  • Breaking: types: Spell the OTHER license value in uppercase (#579)
  • Breaking: tasks: Partition tasks into train, validation, and test splits (#398)
  • Breaking: annotations: Make the owner define time annotation scope (#394)
  • Breaking: timef: Store shared Record origins (#389)
  • Breaking: timef: Add offsets to regular time axes (#393)
  • Breaking: tasks: Infer and store task input modalities (#377)
  • format: Add LazyText handles and pack them whatever their size (#656)
  • dataset: Add Signal.is_selected_by, overlaps_with, is_ordinal and slice (#636)
  • connectors: Add reusable PTB-XL recordings (#499)
  • reader: Stream records in batches with a load plan (#532)
  • registry: Prepare payload sources for SDK reads (#542)
  • build: Open declared parents during composition (#539)
  • dataset: Preserve imports through edits (#530)
  • dataset: Resolve the annotations that imported tasks reference (#529)
  • dataset: Import tasks and their dependencies (#606)
  • dataset: Import parent records and metadata (#605)
  • reader: Load values at read time with load plans (#519)
  • format: Let stored value handles keep loaded values (#602)
  • format: Reference exact parent releases by input alias (#512)
  • types: Add a NONE license (#578)
  • dataset: Add basic validation for irregular signals
  • types: Add industrial domain and define every domain (#567)
  • types: Add Hippocratic 3.0 license support (#516)
  • types: Define shared ObjectKind tags (#517)
  • dataset: Select multiple task splits in one call (#447)
  • timef: Distinguish unknown signal units (#390)
  • tasks: Filter tasks by resolved input modalities (#386)

Bug fixes

  • Breaking: format: Report newer versions with an upgrade hint (#774)
  • Breaking: reader: Close catalog connections before a fork (#597)
  • Breaking: torch: Reject missing values instead of filling zeros (#583)
  • registry: Retry rate-limited presigns during parallel downloads (#747)
  • dataset: Make Signal's new docstrings pass ruff 0.17.3 (#645)
  • dataset: Count the Records behind a task's Signals (#589)
  • values: Read Zarr values into buffers of their own (#599)
  • deps: Require DuckDB 1.5.6 to fix the audit's window race (#604)
  • types: Drop Hippocratic-3.0 and enforce canonical SPDX license ids (#543)
  • format: Store annotation descriptions separately (#458)
  • types: Pickle TimeF models through their wire form (#423)
  • tasks: Align record filters and reuse selected rows (#397)

Performance

  • reader: Read imported records in storage order (#531)
  • reader: Cache common modality and metadata checks (#518)
  • dataset: Retain NumPy buffers until Arrow is requested (#581)
  • storage: Store irregular axes once (#401)

Refactoring

  • Breaking: coordinator: Name the plan step type St...
Read more

TimeNet 0.11.0

Choose a tag to compare

@TimoStoff TimoStoff released this 23 Sep 12:02
8f4be1d

TimeNet 0.11.0 replaces the Parquet control tables of TimeF with a DuckDB control plane and a declarative in-memory model. It is a breaking release: TimeF format version 2 replaces version 1, and version 1 directories are not readable. Every registry dataset must be rebuilt with 0.11.0 connectors.

Data model

  • Records hold a Source hierarchy of Signals. Signals are eager or lazy (Signal.from_loader, Signal.from_irregular), and source.select(...) picks Signals to annotate.
  • Every annotated object (TimeFDataset, Record, Source, Signal, Task) implements the SupportsAnnotate protocol with annotate() and annotations.
  • Tasks are declarative: ordered input Records, mixed ordered targets, references to stored Records and annotations, and linear registration with dataset.add_task(task=...).
  • Annotation values are typed: string, integer, float, boolean, or list of strings.

Format

  • control.duckdb (control schema version 1) stores the hierarchy, annotations, tasks, targets, and value chunk locations. The schema is described once in timenet.format.control_schema; audit_control_database() re-checks every relational invariant.
  • Fixed-shape fields are typed columns, so task scope, targets, and annotation values are filterable in SQL. Only free-form metadata is JSON.
  • The manifest reports timef_format_version: 2.

Reader and writer

  • TimeFReader hydrates records one at a time or in batches, reads tasks in one pass, and exposes task_table(), target_table(), and annotation_table() as Arrow tables.
  • TimeFTorchDataset.from_reader() and __getitems__ batch loading.
  • TimeFWriter streams the control plane in typed Arrow batches inside one transaction and validates as it writes.

Removed

Record.time_series, Signal.time_series_id, the TimeSeries alias, Record.add_annotation (use annotate), the Parquet control-plane writers and readers, and the legacy id codec.

Landed as one squash (#361) of #311, #313, #314, #315, #349, #350, #351, #352, #353, #355, #357, #359, and #360.