WenjieDu / PyPOTS / 38386584692

81%

main: 81%

Build Type

push

github

Committed by web-flow

Commit Message

Feat/tsicl imputation (#884)

* build(deps-dev): update pypots requirement from >=1.4 to >=1.5

Updates the requirements on [pypots](https://github.com/WenjieDu/PyPOTS) to permit the latest version.
- [Release notes](https://github.com/WenjieDu/PyPOTS/releases)
- [Commits](https://github.com/WenjieDu/PyPOTS/compare/v1.4...v1.5)

---
updated-dependencies:
- dependency-name: pypots
  dependency-version: '1.5'
  dependency-type: direct:development
...

Signed-off-by: dependabot[bot] <support@github.com>

* feat: add TS-ICL foundation model for time-series imputation

Vendors the TS-ICL architecture (Perceiver encoder + in-context-learning
transformer head) and its zero-shot imputation inference logic from
https://github.com/EDF-Lab/ts-icl into pypots/nn/modules/tsicl and
pypots/imputation/tsicl, wrapping it as pypots.imputation.TSICL. The
pretrained checkpoint is downloaded from the Hugging Face Hub on first use,
the same way PyPOTS' other pretrained backbones (GPT4TS, MOMENT) fetch
theirs.

TS-ICL has no fine-tuning procedure in the original implementation, so
fit() is a no-op (warns) and predict()/impute() run the pretrained model
directly, feature by feature (channel independence).

Unlike the rest of PyPOTS (BSD-3-Clause), the vendored TS-ICL architecture
and pretrained checkpoints are licensed by EDF SA under the TS-ICL
Non-Commercial License v1.0 (research/evaluation/benchmarking only, not
commercial or production use). This is called out explicitly in the model's
docstring and in a NOTICE file (shipped in the wheel) in both vendored
directories, as required by that license's distribution terms.

* feat: add TS-ICL forecasting support

Adds pypots.forecasting.TSICL, reusing the TS-ICL architecture already
vendored for imputation. Zero-shot, no fine-tuning (fit() warns), with the
autoregressive rollout the original pipeline uses when the requested horizon
exceeds the model's per-pass prediction cap (max_target_len).

The checkpoint-loading and conte... (continued)

Coverage Stats

911 of 1251 new or added lines in 29 files covered. (72.82%)

4 existing lines in 3 files now uncovered.

17821 of 21998 relevant lines covered (81.01%)

0.81 hits per line