[ENH] softs_v2 Model added by Muhammad-Rebaal · Pull Request #2232 · sktime/pytorch-forecasting

Muhammad-Rebaal · 2026-03-25T17:38:19Z

Hi @fkiraly , @phoeenniixx, @PranavBhatP !

I have implemented the SOFTS (Star Aggregate-Dispatch for Time Series Forecasting) model within the PyTorch Forecasting v2 architecture.

Could you please review the PR?

Here is a summary of the changes made:

Isolated Layer Abstraction (pytorch_forecasting/layers/_blocks/_softs_block.py): Implemented the core neural network components, specifically the novel STADModule (Star Aggregate-Dispatch mechanism) and the SoftsEncoderLayer, completely isolated from the base estimating logic to strictly adhere to the project's v2 architectural standards.
Model Construction (pytorch_forecasting/models/softs/): Created the main Softs estimator wrapper which inherits from TslibBaseModel. It effectively handles dynamic feature input alignments (history_cont and history_target), optionally applies RevIN scaling, and routes the forward pass through the STAD-based encoders.
V2 Package Management (_softs_pkg_v2.py): Established the Softs_pkg_v2 class representing model metadata (capability tags, compute requirements) and defined automated testing configurations (get_test_train_params) tightly integrated with the underlying TslibDataModule.
Registry & Exports: Updated the relevant __init__.py files across the layers and models directories to properly expose Softs, Softs_pkg_v2, and the STAD blocks to the broader PyTorch Forecasting ecosystem.

Thank you!

…rch-forecasting into softs_model

codecov · 2026-04-09T16:44:43Z

Codecov Report

❌ Patch coverage is 95.55556% with 6 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (main@7bcf66c). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
pytorch_forecasting/models/softs/_softs_v2.py	92.20%	6 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2232   +/-   ##
=======================================
  Coverage        ?   87.65%           
=======================================
  Files           ?      180           
  Lines           ?    10299           
  Branches        ?        0           
=======================================
  Hits            ?     9028           
  Misses          ?     1271           
  Partials        ?        0

Flag	Coverage Δ
cpu	`87.65% <95.55%> (?)`
pytest	`87.65% <95.55%> (?)`

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

Thanks!
i would suggest adding detailed docstrings with clear descriptions of all the params, so that it is easy to understand the methods and the architecture in general.

Is this code taken from somewhere or is it completely implemented from scratch? I think if it is taken from somewhere, it would be better to add the original authors to the authors tag.

Also, please add the references (like paper, repo etc) to the docstrings as well of the model.

phoeenniixx · 2026-06-01T13:46:39Z

+
+class SoftsEncoderLayer(nn.Module):
+    """
+    Single Encoder layer for SOFTS.


Can you please add more detailed docstrings for all methods and classes?

phoeenniixx · 2026-06-01T13:49:03Z

+        """Get the underlying DataModule class."""
+        from pytorch_forecasting.data._tslib_data_module import TslibDataModule
+
+        return TslibDataModule


Why are we using this datamodule? I think this data module is mainly for tslib models, I have no issue with using this, but pls have a look at EncoderDecoderDataModule as well, maybe that would also be helpful.
I have not looked at the architecture, so I have a question: is the model encoder-decoder based model?

We are using TslibDataModule because SOFTS is a direct-projection MLP model. It only consumes historical input sequences and maps them directly to the forecast window, using EncoderDecoderModule felt unnecessary.

From the perspective of data input we can call it encoder-only model.

Code Ref : Code

I think this can be seen as a encoder-decoder model with decoder being a identity layer?
I think we should use the TslibDataModule only if we are interfacing the model from the tslib package. Otherwise if it fits encoder-decoder model type, we should use EncoderDecoderDataModule

You need to update this data module as well then

phoeenniixx · 2026-06-01T13:49:19Z

+        return TslibDataModule
+
+    @classmethod
+    def _get_test_datamodule_from(cls, trainer_kwargs):


I dont think we need this anymore

phoeenniixx · 2026-06-01T13:49:41Z

+    def get_test_train_params(cls):
+        params = [
+            {},
+            dict(hidden_size=128, n_layers=1, use_revin=True),


Can you try adding some more params to cover more scenarios?

phoeenniixx · 2026-06-01T13:49:56Z

+class Softs(TslibBaseModel):
+    """
+    SOFTS: Efficient Multivariate Time Series Forecasting with Series-Core Fusion.
+    """


Add detailed docstrings

phoeenniixx

I think there are geniune failures, please take a look at that

fkiraly · 2026-06-21T21:42:11Z

@phoeenniixx, I think there are no longer failures. Can you please re-review?

(quick question, do we need to add the model to the docs, or is that automatic now?)

phoeenniixx · 2026-06-22T07:38:09Z

(quick question, do we need to add the model to the docs, or is that automatic now?)

We still need to add the API reference, but the table generation is automatic now!

phoeenniixx

Thanks! I think this one is almost ready. Just few nitpicks.
Please add the API reference of the model here: http://github.com/sktime/pytorch-forecasting/blob/main/docs/source/m_layer_v2.rst#api-reference

phoeenniixx · 2026-06-22T07:40:37Z

+from pytorch_forecasting.models.base._tslib_base_model_v2 import TslibBaseModel
+
+
+class Softs(TslibBaseModel):


Can the BaseModel not work here? I mean we mainly use TslibBaseModel for tslib models only

The model is actually part of the tslib model family that's why I used that. Is there any particular reason to use the base model instead ?

I thought tslib model family was from here: https://github.com/thuml/Time-Series-Library
But this model was taken from here: https://github.com/Secilia-Cxy/SOFTS/
That is why I said it is not a tslib model. Am I missing something here?

Yeah, you're right I just misinterpreted thanks for letting me know. I'll adjust that and use the baseClass.

phoeenniixx · 2026-06-22T07:42:00Z

+                hidden_size=64,
+                n_layers=1,
+                use_revin=False,
+                loss=MAE(),


It would be good if we could add other point prediction losses here as well - just to increase the coverage

Thanks for letting me know I've updated that

phoeenniixx · 2026-06-23T16:42:36Z

+    """
+
+    _tags = {
+        "info:name": "SOFTS",


I think the name should be same as the class here: Softs in place of SOFTS?

I think it should be SOFTS all over the place as the actual model name is SOFTS not Softs and also the convention also follows that across other models.

Then you would have to update the class name. The tag name should be exactly same as the class name

Muhammad-Rebaal · 2026-06-24T15:11:38Z

Hi @phoeenniixx,
I've updated the class_name to SOFTS and use the base_class, also updated the api_reference. Kindly have a look at it.
Thanks

phoeenniixx

Please use EncoderDecoderTimeSeriesDataModule in place of tslib one, if we are using the BaseModel and not the tslib one, then we need to use the corresponding data module as well. Although the tests are passing, but both the data modules represent different ideologies, and with time as we add more feats will diverge from each other, so we shouldn't use them interchangeably

phoeenniixx · 2026-07-01T11:33:40Z

+        return x + dispatch_out
+
+
+class SOFTSEncoderLayer(nn.Module):


should it go to _encoders folder?

I considered this during implementation. Looking at the current _encoders/ contents, both Encoder and EncoderLayer are TimeXer-specific (they take cross, tau, delta params and have global-token logic).
Whereas SOFTSEncoderLayer has a fundamentally different interface it takes a 4D tensor (B, C, L, D) and uses STAD instead of attention, so there's no shared contract between them. _blocks/ currently houses model-specific building blocks like ResidualBlock (for DSIPTs), and SOFTSEncoderLayer + STADModule follow the same pattern self-contained blocks specific to one model. Should I move it to a new _softs/ subfolder under layers/ to make the model association clear.

I see, but we can have multiple implementations of encoder layers, no? and that is why we created an _encoders folder that can host multiple implementations. It is not necessary that this encoder layer is just used by SOFTS, what if we see some new model that is derived from SOFTS, that could also use this layer. THe name - SOFTSEncoderLayer already makes the association pretty clear.

Yes, it can be possible that in the future we'd reuse it. I've adjusted that.

phoeenniixx · 2026-07-01T11:35:59Z

+
+        return {"prediction": out}
+
+    def predict_step(


why do we need a predict_step here? can't the BaseModel's implementation enough?
And if not, why do have not added other step functions like test_step?

Yes, I thoroughly looked into the implementation. The BaseModel's implementation is enough.

phoeenniixx

Nice! I think it is almost ready, just have few doubts (see above)

Muhammad-Rebaal · 2026-07-29T14:35:53Z

Hi @phoeenniixx,
I've made the following changes as mentioned:

Moved the softs_encoder block to _encoder directory for future resuability.
Removed the overriding predict_step method as Base_models implementation is enough.

Kindly have a look

…ecasting into softs_model

phoeenniixx

Nice!

[ENH] softs_v2 Model added

cd931ac

Muhammad-Rebaal requested review from PranavBhatP, benHeid, fkiraly, fnhirwa, jdb78, phoeenniixx and yarnabrina as code owners March 25, 2026 17:38

fix: ruff issue fixed

c90e494

Muhammad-Rebaal marked this pull request as draft March 25, 2026 17:49

Muhammad-Rebaal added 2 commits March 25, 2026 22:52

fix: ruff issue fixed

6026e0e

Author and ref added

fcb7cad

Muhammad-Rebaal marked this pull request as ready for review March 25, 2026 18:41

Muhammad-Rebaal marked this pull request as draft March 25, 2026 19:23

Muhammad-Rebaal and others added 4 commits April 8, 2026 13:06

Merge branch 'main' into softs_model

2db264b

fix : Pytest failure

5026e82

Merge branch 'softs_model' of https://github.com/Muhammad-Rebaal/pyto…

4d19fa8

…rch-forecasting into softs_model

fix : Code Quality

a44150c

Merge branch 'main' into softs_model

927c50e

Muhammad-Rebaal marked this pull request as ready for review April 29, 2026 19:19

Merge branch 'main' into softs_model

b2aa1f6

phoeenniixx requested changes Jun 1, 2026

View reviewed changes

phoeenniixx and others added 3 commits June 1, 2026 15:52

Merge branch 'main' into softs_model

a18a499

Merge branch 'main' into softs_model

0087cb8

Added Docstring, removed v1 method, add params

8fb4616

Muhammad-Rebaal requested a review from phoeenniixx June 4, 2026 11:41

Merge branch 'main' into softs_model

0afece8

phoeenniixx requested changes Jun 16, 2026

View reviewed changes

Merge branch 'main' into softs_model

b3f75a9

fkiraly added the ptf-v2 Related to `pytorch-forecasting` v2 label Jun 19, 2026

fkiraly assigned Muhammad-Rebaal Jun 19, 2026

Merge branch 'main' into softs_model

b460ae8

Merge branch 'main' into softs_model

d7b34b9

phoeenniixx requested changes Jun 22, 2026

View reviewed changes

Updated the api ref & added loss func

dae93e8

phoeenniixx reviewed Jun 23, 2026

View reviewed changes

Added reference, code refactor and use base_model class

c0f6f17

Muhammad-Rebaal and others added 2 commits June 24, 2026 20:23

Added pytest fix

c8ca76f

Merge branch 'main' into softs_model

f1a2614

phoeenniixx requested changes Jun 25, 2026

View reviewed changes

Muhammad-Rebaal and others added 2 commits June 26, 2026 20:16

ENH: Migrated from TslibDataModule to EncoderDecoderTimeSeriesDataModule

d35353c

Merge branch 'main' into softs_model

5c75c71

phoeenniixx reviewed Jul 1, 2026

View reviewed changes

phoeenniixx requested changes Jul 1, 2026

View reviewed changes

fix : removed predict_step method

9a71c1d

Muhammad-Rebaal requested a review from phoeenniixx July 13, 2026 08:56

Muhammad-Rebaal and others added 3 commits July 14, 2026 21:52

Merge branch 'main' into softs_model

f35af9a

Merge branch 'main' into softs_model

c8528ed

Fix: Moved the SoftsEncoder to _encoder directory

844d82b

Muhammad-Rebaal and others added 3 commits August 3, 2026 17:21

Merge branch 'main' of https://github.com/Muhammad-Rebaal/pytorch-for…

371fca3

…ecasting into softs_model

Fix: Code Quality

aa653fa

Merge branch 'main' into softs_model

dd94e8b

phoeenniixx approved these changes Aug 8, 2026

View reviewed changes

Merge branch 'main' into softs_model

55a1bc5

		from pytorch_forecasting.models.base._tslib_base_model_v2 import TslibBaseModel


		class Softs(TslibBaseModel):

Uh oh!

Conversation

Muhammad-Rebaal commented Mar 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov Bot commented Apr 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

phoeenniixx left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

phoeenniixx left a comment

Choose a reason for hiding this comment

Uh oh!

fkiraly commented Jun 21, 2026

Uh oh!

phoeenniixx commented Jun 22, 2026

Uh oh!

phoeenniixx left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Muhammad-Rebaal Jun 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Muhammad-Rebaal Jun 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Muhammad-Rebaal commented Jun 24, 2026

Uh oh!

phoeenniixx left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

phoeenniixx left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Muhammad-Rebaal commented Mar 25, 2026 •

edited

Loading

codecov Bot commented Apr 9, 2026 •

edited

Loading

Muhammad-Rebaal Jun 23, 2026 •

edited

Loading

Muhammad-Rebaal Jun 23, 2026 •

edited

Loading

phoeenniixx left a comment •

edited

Loading