[ENH] Add preprocessing (scalers and normalizers) to tslib datamodule by echo-xiao · Pull Request #2368 · sktime/pytorch-forecasting

echo-xiao · 2026-08-03T21:51:44Z

Reference Issues/PRs

What does this implement/fix? Explain your changes.

TslibDataModule accepted scalers / target_normalizer but never used them
(stored and ignored). This makes them functional, following #2302 and reusing
ScalerAdapter:

__init__ wraps them in ScalerAdapter, adds _*_fitted flags and a
per-series _preprocess_cache.
_fit_scalers / _fit_target_normalizer fit on the train split only.
_normalize_features / _normalize_target transform per series (no-op until
fitted; only configured continuous columns).
_preprocess_data applies the transforms and caches per series;
setup("fit") fits on _train_indices.

What should a reviewer concentrate their feedback on?

no deepcopy of scalers
deferring target_scale** to a follow-up
only continuous features are scaled, the models don't consume them yet
target_scale, collate_fn and the models are untouched

Did you add any tests for the change?

Yes — 7 tests in pytorch_forecasting/data/tests/test_tslib_data_module.py:

adapter wrapping + flags;
fit-on-train (StandardScaler → ~0 mean / ~1 std);
target-normalizer fit flag;
feature transform (only configured columns);
target no-op-until-fitted; _preprocess_data scaling + cache identity;
setup("fit") produces scaled samples;
a no-scaler backward-compat guard (byte-identical output, no target_scale).

PR checklist

The PR title starts with [ENH].
Added/modified tests.
Used pre-commit hooks (pre-commit run --files on both changed files, clean).

Add test_no_scalers_leaves_data_untouched to guard backward compatibility: when no scalers are configured, continuous features pass through untouched and target_scale is absent from the preprocessed output.

echo-xiao · 2026-08-03T21:58:43Z

hi, @phoeenniixx please review these and let me know if these decisions are right:

no deepcopy of scalers
deferring target_scale to a follow-up, collate_fn and the models are untouched
only continuous features are scaled, the models don't consume them yet

TODO:

lift the shared preprocessing into the base datamodule
implement target_scale for tslib (sibling of [ENH] Add target_scale support to EncoderDecoderDataModule #2359)

codecov · 2026-08-04T00:05:13Z

Codecov Report

❌ Patch coverage is 94.28571% with 4 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (main@2c58076). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
...forecasting/data/data_module/_tslib_data_module.py	94.28%	4 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2368   +/-   ##
=======================================
  Coverage        ?   90.43%           
=======================================
  Files           ?      203           
  Lines           ?    11090           
  Branches        ?        0           
=======================================
  Hits            ?    10029           
  Misses          ?     1061           
  Partials        ?        0

Flag	Coverage Δ
cpu	`90.43% <94.28%> (?)`
pytest	`90.43% <94.28%> (?)`

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

phoeenniixx · 2026-08-04T11:55:55Z



+def _make_ts(n_series: int = 20, length: int = 40, offset: float = 100.0) -> TimeSeries:
+    """合成数据集:连续特征 ``x`` 远离 0(~offset),便于看出标准化;目标 ``y`` 是正弦。"""


please use english docstrings

phoeenniixx

Nice!
I have a few questions:

Did you look at the thuml implementation? I think we should try to keep it as close to this package as possible - when it comes to the preprocessing logic as this dm is based on this package
- Currently it supports all the scalers and normalizers that EncoderDecoderDataModule supports, but is this true for the tslib implementation as well?

Also, please use english docstrings, to maintain uniformity across the documentation.

…sing

Address review feedback on sktime#2368: the helper docstring and inline comments added for the preprocessing tests were written in Chinese. Rewrite them in English to stay consistent with the rest of the codebase documentation.

…sing

`TslibDataModule` accepted `scalers` and `target_normalizer` but never used them. This makes them functional, fit once on the training split, mirroring thuml's data layer. Scope follows thuml's preprocessing, which hardcodes a single global sklearn StandardScaler fit on the train segment: - Only sklearn `StandardScaler` is accepted. Anything else raises `NotImplementedError` rather than failing deep inside the scaler stack or, in the case of `EncoderNormalizer`, silently producing wrongly scaled data (its `fit_per_sequence` semantics are inverted by a global fit). - `target_normalizer` now defaults to `None` instead of `"auto"`, and `"auto"` resolves to no normalization. The inverse transform (`target_scale`) is not implemented yet (sktime#2359), and all three models on this data module guard on `"target_scale" in x`, so normalizing by default would leave every prediction in normalized space with no way back. Making the fit boundary hold required two further changes: - `_ensure_split` computes the train/val/test split once and caches it, matching what sktime#2302 did for `EncoderDecoderTimeSeriesDataModule`. `setup` runs once per stage, so the unconditional `torch.randperm` handed `trainer.test()` a different split than `trainer.fit()` had used, moving series the scalers were fit on into the test set. - The fit calls moved out of the `stage == "fit"` branch, so a standalone `setup("test")` no longer passes raw values through unscaled. Also drops the per-series preprocessing cache added earlier on this branch: it is unrelated to scaling and belongs with the base class work in sktime#2313. Drive-by: `test_multivariate_target` asserted `y.shape[-1] == 2`, but sktime#1960 changed multi-target `__getitem__` to return one tensor per target. The assertion is updated to match. This file is not collected by CI, which is why the failure went unnoticed.

`testpaths` listed only `tests/` and `pytorch_forecasting/tests/`, so `pytorch_forecasting/data/tests/` was never collected. Its tests have therefore not run in CI at all, which is why the codecov patch report shows 0% for `test_tslib_data_module.py` and why a stale assertion in `test_multivariate_target` went unnoticed for months. Adding the directory collects 34 further tests (1165 -> 1199) and they all pass, so this does not turn CI red.

`test_tslib_data_module.py` lived in `pytorch_forecasting/data/tests/`, which is not listed in `testpaths`, so it had never run in CI. That is why codecov reported 0% patch coverage for it, and why a stale assertion in `test_multivariate_target` (left behind by sktime#1960) went unnoticed. Move it to `tests/test_data/`, where every other data layer test already lives -- `test_data_module.py`, `test_d1.py`, `test_encoders.py`, `test_samplers.py`, `test_timeseries.py` -- and where sktime#2302 put the `EncoderDecoderTimeSeriesDataModule` scaler tests this PR mirrors. This also reverts the `testpaths` entry added in the previous commit, which is no longer needed. The 34 tests are now collected and all pass.

echo-xiao · 2026-09-08T00:29:43Z

Nice! I have a few questions:

Did you look at the thuml implementation? I think we should try to keep it as close to this package as possible - when it comes to the preprocessing logic as this dm is based on this package

Currently it supports all the scalers and normalizers that EncoderDecoderDataModule supports, but is this true for the tslib implementation as well?

Also, please use english docstrings, to maintain uniformity across the documentation.

In thuml's data layer, it's preprocessing is:

one hardcoded sklearn standardscaler. no other way to pass a custom scaler
features and target share the one scaler, there is no seperate target_normalizer
it is fit strictly on the train segment.
inverse_transform is exposed on the datasets. metrics are computed in the original scale.
per-window normalisation exists but lives in the model, not in the data layer. the normalizer in usa.py is per-sample and classification-only.

tslib does not support EncoderDecoderDataModule supports.

list/tuple of normalizers
multivariate target + one normalizer
groupnormalizer
encodernormalizer

what this pr does:

scalers are now really fits on the training split only. the split is computed once and cached, scalers are fit once, and fitting happens for every stage. previously setup() redrew the split on each call, so trainer.test() saw a split the scalers had already been fit on.
only sklearn standardscaler is accepted. anything else raises not implemented error.
target normalization is off by default.
the tests now run in CI

`TslibDataModule` now rejects normalizers other than sklearn's `StandardScaler`, so the tutorial's `target_normalizer=TorchNormalizer()` raises `NotImplementedError` and fails the notebook CI job. Switch the tutorial to `StandardScaler` and drop the now unused `TorchNormalizer` import.

review-notebook-app · 2026-09-08T15:08:27Z

Check out this pull request on

See visual diffs & provide feedback on Jupyter Notebooks.

Powered by ReviewNB

echo-xiao added 7 commits August 3, 2026 13:24

[ENH] tslib dm: wrap scalers/normalizer in ScalerAdapter

7fe9a9c

[ENH] tslib dm: fit scalers and target normalizer on train split

8767aeb

[ENH] tslib dm: add feature/target transform methods

606fc59

[ENH] tslib dm: apply scaling in _preprocess_data with per-series cache

823e082

[ENH] tslib dm: fit scalers on train split in setup

3c2efc5

[ENH] tslib dm: regression test for no-op default path (no scaler)

a87c1a6

Add test_no_scalers_leaves_data_untouched to guard backward compatibility: when no scalers are configured, continuous features pass through untouched and target_scale is absent from the preprocessed output.

[ENH] tslib dm: use English for an inline comment

759f3af

echo-xiao requested review from benHeid, fkiraly, jdb78 and phoeenniixx as code owners August 3, 2026 21:51

phoeenniixx reviewed Aug 4, 2026

View reviewed changes

phoeenniixx requested changes Aug 4, 2026

View reviewed changes

phoeenniixx assigned echo-xiao Aug 4, 2026

phoeenniixx added enhancement New feature or request module:datasets&dataloaders ptf-v2 Related to `pytorch-forecasting` v2 labels Aug 4, 2026

Merge remote-tracking branch 'upstream/main' into enh/tslib-preproces…

753f21e

…sing

echo-xiao mentioned this pull request Aug 18, 2026

[ENH] Add preprocessing to tslib datamodule #2330

Open

echo-xiao added 5 commits September 7, 2026 10:13

Merge remote-tracking branch 'upstream/main' into enh/tslib-preproces…

27eeb05

…sing

Navigation Menu

Uh oh!

[ENH] Add preprocessing (scalers and normalizers) to tslib datamodule - #2368

[ENH] Add preprocessing (scalers and normalizers) to tslib datamodule#2368
echo-xiao wants to merge 14 commits into
sktime:mainfrom
echo-xiao:enh/tslib-preprocessing

echo-xiao commented Aug 3, 2026

Uh oh!

echo-xiao commented Aug 3, 2026 •

edited

Loading

Uh oh!

codecov Bot commented Aug 4, 2026 •

edited

Loading

Uh oh!

phoeenniixx Aug 4, 2026

Uh oh!

phoeenniixx left a comment

Uh oh!

echo-xiao commented Sep 8, 2026

Uh oh!

review-notebook-app Bot commented Sep 8, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants



		def _make_ts(n_series: int = 20, length: int = 40, offset: float = 100.0) -> TimeSeries:
		"""合成数据集:连续特征 ``x`` 远离 0(~offset),便于看出标准化;目标 ``y`` 是正弦。"""

Uh oh!

Conversation

echo-xiao commented Aug 3, 2026

Reference Issues/PRs

What does this implement/fix? Explain your changes.

What should a reviewer concentrate their feedback on?

Did you add any tests for the change?

PR checklist

Uh oh!

echo-xiao commented Aug 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov Bot commented Aug 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

phoeenniixx Aug 4, 2026

Choose a reason for hiding this comment

Uh oh!

phoeenniixx left a comment

Choose a reason for hiding this comment

Uh oh!

echo-xiao commented Sep 8, 2026

Uh oh!

review-notebook-app Bot commented Sep 8, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

echo-xiao commented Aug 3, 2026 •

edited

Loading

codecov Bot commented Aug 4, 2026 •

edited

Loading