Feat/patchtst fm

[spolisar]

Add support for the new PatchTST-FM forecasting model.

PatchTST-FM is currently setup to use pytorch's mps backend when available, but there is a chance that will have issues. If it becomes a problem, we can disable that.

For now predicting with levels is disabled.

TODO:

• Determine notebook position in example list

[CLAassistant]

All committers have signed the CLA.

[AzulGarza]

hey @spolisar! thanks so much! looks great so far, i left a couple of comments. can also add a plot to the nb? that way we can make sure everything looks good.

[AzulGarza]

these three methods are also being used by sundial

timecopilot/timecopilot/models/foundation/sundial.py

Lines 94 to 137 in fe710ca

	def _left_pad_and_stack_1D(self, tensors: list[torch.Tensor]) -> torch.Tensor:
	max_len = max(len(c) for c in tensors)
	padded = []
	for c in tensors:
	assert isinstance(c, torch.Tensor)
	assert c.ndim == 1
	padding = torch.full(
	size=(max_len - len(c),),
	fill_value=torch.nan,
	device=c.device,
	dtype=c.dtype,
	)
	padded.append(torch.concat((padding, c), dim=-1))
	return torch.stack(padded)
	def _prepare_and_validate_context(
	self,
	context: list[torch.Tensor] | torch.Tensor,
	) -> torch.Tensor:
	if isinstance(context, list):
	context = self._left_pad_and_stack_1D(context)
	assert isinstance(context, torch.Tensor)
	if context.ndim == 1:
	context = context.unsqueeze(0)
	assert context.ndim == 2
	return context
	def _maybe_impute_missing(self, batch: torch.Tensor) -> torch.Tensor:
	if torch.isnan(batch).any():
	batch = batch.float().numpy()
	imputed_rows = []
	for i in range(batch.shape[0]):
	row = batch[i]
	imputed_row = LastValueImputation()(row)
	imputed_rows.append(imputed_row)
	batch = np.vstack(imputed_rows)
	batch = torch.tensor(
	batch,
	dtype=self.dtype,
	device=self.device,
	)
	return batch

. let's abstract them. what comes to my mind is to create a new class, maybe in forecaster.py, called _Processor or a better name. that way we can reuse the same code across different models.

[spolisar]

I copied them from Flowstate, so that sounds good.
The only attributes of the class that are used are dtype and device, so the options that come to mind are:

• A class with dtype and device attributes
• A class with static methods and dtype and device added as args where needed
• Functions with dtype and device added as args where needed

[AzulGarza]

including them as attributes in the class sounds good.

[spolisar]

@AzulGarza does this seem good?
db20d39

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[Copilot]

Pull request overview

Copilot reviewed 6 out of 7 changed files in this pull request and generated 10 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

You can also share your feedback on Copilot code review. Take the survey.

[AzulGarza]

@spolisar let's also take a look at copilot's comments. they are valuable.