[BUG] TslibBaseModel crashes when metadata=None (Root cause for Samformer issue #2212) #2217
Description
Describe the bug
Hi team! I was digging into the codebase and found a slight inconsistency in _tslib_base_model_v2.py that causes an AttributeError during initialization if metadata=None is passed.
In the __init__ method, we safely default empty metadata on line 57 using self.metadata = metadata or {}. However, a few lines further down, the code accidentally references the raw metadata parameter instead of the safely defaulted self.metadata attribute:
# Line 57 — Safely handles None (Great!)
self.metadata = metadata or {}
# Lines 69-70 — Uses self.metadata (Correct)
self.context_length = self.metadata.get("context_length", 0)
# Lines 72, 79, 86, 89 — Uses the raw parameter (BUG)
# If metadata is None, these .get() calls will crash
feature_indices = metadata.get("feature_indices", {})
feature_dims = metadata.get("n_features", {})
self.feature_names = metadata.get("feature_names", {})
self.features = metadata.get("features", "MS") To Reproduce
You can trigger this crash by initializing any v2 tslib model without metadata:
from pytorch_forecasting.models.dlinear._dlinear_v2 import DLinear
from pytorch_forecasting.metrics import MAE
# This raises an AttributeError: 'NoneType' object has no attribute 'get'
model = DLinear(loss=MAE(), metadata=None)Expected behavior
The base class should consistently use self.metadata for all .get() calls so it can safely fall back to the empty dictionary we defined at the top of the method.
Additional context
I bring this up because this is actually the root cause of #2212! While #2212 specifically mentions Samformer crashing, this bug lives in the shared TslibBaseModel, meaning it currently affects all v2 subclasses (like DLinear, TimeXer, etc.).
PR #2213 currently attempts to patch this for Samformer specifically, but I think a much cleaner, global fix would be to just update those 4 lines in the base class to use self.metadata.
Versions
Details
pytorch-forecasting: main (latest)
cc: @phoeenniixx @PranavBhatP — would appreciate any pointers if you have context on _tslib_base_model_v2.py