tdseries

Immutable pytree frames of tensors with named, indexed dimensions, where time is a first-class dimension stored in exact int64 ticks (nanoseconds).

Built for bundling media signals with co-recorded telemetry and the static geometry that shares their non-time dimensions — e.g. multichannel audio (mic, time) together with microphone positions (mic, 3), rotor speeds (rotor, time) together with rotor positions (rotor, 3):

import tdseries as td

frame = td.Frame({
    "audio":     td.uniform(audio_8xT, sr=44100, dims=("mic", "time")),
    "mic_pos":   td.wrap(pos_8x3, dims=("mic", None)),
    "rps":       td.events(stamps, rps_4xM, dims=("rotor", "time")),
    "rotor_pos": td.wrap(rpos_4x3, dims=("rotor", None)),
    "vad":       td.spans(starts, ends),
    "recording_id": "FLY124",
})

sub  = frame.slice["mic", 0]     # audio -> (T,), mic_pos -> (3,): same mic
clip = frame.time[1.0:4.5]       # seconds; all temporal leaves cut exactly
raw  = frame.ticks[a:b]          # exact int64 tick window
rps  = frame["rps"]              # self-contained absolute-time Series

The model

A dimension is a name plus a domain; every tensor axis carrying the name has an index mapping positions into that domain. Slices are expressed in domain coordinates and translated per-tensor — which is how tensors of different lengths share one dimension:

Time is exact: all anchors and cut points are int64 tick counts, shift is O(1), and the algebra satisfies, at the sample level and without tolerances,

x.ticks[a:b].concat(x.ticks[b:c]) == x.ticks[a:c]

for any tick cut points — including sub-sample cuts through audio and cuts through the middle of a VAD span (split spans re-merge via identity tags).

Everything is immutable. Extracting an entry from a frame hands back a self-contained series re-anchored to absolute time; the frame is untouched.

See DESIGN.md for the full contract.

Development

nix develop        # or: uv sync
uv run pytest

nix develop provides Python + uv and pre-commit hooks (ruff, ruff-format, pyright). Data leaves may be any Array-API array — numpy, torch, JAX, CuPy, … — dispatched through array-api-compat; torch stays optional (uv sync --extra torch). The time/dimension index layer is numpy-int64 regardless of the payload backend.

Lineage

Extracted from the utils.data module of the harmonic-noise-suppression research project; the tick/phase arithmetic and the slice/concat/shift invariants are ported from there, the named-dimension frame model is new.