[BUG] Robust indexing & safe dtype handling in tsai’s TSDataLoaders / TfmdLists #947
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
6 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Fixes sktime/sktime#7885 <[ENH] interface to tsai package>
What does this implement / fix?
While wiring tsai into sktime I hit two subtle indexing / dtype issues that can also bite vanilla tsai users:
----NumPy scalar vs. Python int in TfmdLists + Subset
When self._splits is a NumPy array (dtype=int8/int32/…) the value returned by
idx = self._splits[it] is a NumPy scalar.
Indexing self.items[idx] with that object fails, breaking Learner.predict, dev
notebooks, and downstream libs.
Patch: Convert NumPy scalar → .item() and tiny NumPy array → .tolist() before the final lookup.
Zero behavioural change for normal Python ints/lists.
-----Unsafe dtype when casting NumPy arrays to tensors in TfmdLists.init
Inside the in-place branch (inplace=True, tfms=None) we call typ(tl.items).
If tl.items is an integer array, torch.as_tensor produces a LongTensor, which later collides with models expecting FloatTensors.
Patch: After the existing cast, explicitly re-cast any NumPy array to torch.float32.
Dependency impact
None—pure Python changes, no new packages.
Focus for reviewers
Sanity-check the scalar/array coercion logic in getitem.
Confirm that forcing float32 won’t interfere with edge-cases
where integer tensors are explicitly desired (I could not find any).
I’ve run the basic_motions smoke-tests plus a small custom dataset; everything trains and predicts fine on CPU and GPU.
Thanks for taking a look!