[ENH] Implementing D2 data module layer for tslib models.

[PranavBhatP]

Description

This PR fixes issue #1833, implements a D2 for tslib.

This PR involves the following changes:

• New dataset and datamodule for tslib - link
• Implementation of TimeXer with new data module. - link
• Base class implementation of a TslibBaseModel - link.
• Tests for the tslib data module. - link
• Example notebook - link
• Restructure codebase to include a layers directory for module containing architectural deep learning layer classes.

Checklist

• Linked issues (if existing)
• Amended changelog for large changes (and added myself there as contributor)
• Added/modified tests
• Used pre-commit hooks when committing to ensure that code is compliant with hooks. Install hooks with pre-commit install.
  To run hooks independent of commit, execute pre-commit run --all-files

Make sure to have fun coding!

[fkiraly]

Nice!

Minor structure request: can you kindly make the attention etc modules in layers actual folders, in those private submodules with a single class each?

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[PranavBhatP]

FYI @agobbifbk @fkiraly @phoeenniixx. Rough Pipeline for tslib is complete. I have implemented TimeXer on this pipeline. Tests are pending. There are still some bugs, but training and prediction are working.

[fkiraly]

Great!

• there is an empty file modules.py, can you remove this from the PR?
• can you ensure the new base class has the "experimental" warning for now, similar to the v2 base model
• could you put the base class in another private file, in base?

[PranavBhatP]

Hi @fkiraly @phoeenniixx

I went through the PR #1841, and considering that @phoeenniixx is implementing a v2 version of the test_all_estimators.py, this surely means that I will depend on these changes to implement a newer version test_timexer.py for this PR (currently the v1 version is in #1797)

Is it advisable to stack upon PR #1841? Or should I wait for a merge and work only on the independent tests for the tslib datamodule?

[fkiraly]

(code formatting is failing)

[PranavBhatP]

Did code quality tests change @fkiraly? The pre commit tests are passing on my local repo, but here it is failing.

EDIT: Resolved.

[PranavBhatP]

Hi @agobbifbk , continuing the discussion from the meet about the windowing, this is the code I was talking about
https://github.com/thuml/Time-Series-Library/blob/main/data_provider/data_loader.py

Specifically the implementation in the class Dataset_ETT_hour and Dataset_M4. Here are they are using fixed window sampling for the first one (Dataset_ETT_hour) and random window sampling ( Dataset_M4) for the second, this seems to be a difference but I might be mistaken. The reason they might be using different windowing is because of the forecasting task for which the dataset is used. (long-term or short-tem forecasting etc).

After going through the code deeper, I understand what you were saying, I don't think it matters what windowing we use as long as it matches with model input and makes it convenient to keep single one in the d2. But it would be nice if you could review the current implementation - link and some comments on whether we need separate windowing strategy on top of this.

Also FYI @phoeenniixx , since you were also involved in the discussion.