Runtime crash in Samformer initialization when metadata is None

[lomesh2312]

Description

When initializing the Samformer model, the constructor accepts a metadata parameter with the type hint dict | None = None. However, the __init__ method unconditionally attempts to index into self.metadata (e.g., self.metadata['max_encoder_length']), which immediately throws a TypeError: 'NoneType' object is not subscriptable and crashes initialization if metadata is None.

Location

pytorch_forecasting/models/samformer/_samformer_v2.py, specifically lines 77-79.

Steps to reproduce

import torch.nn as nn
from pytorch_forecasting.models.samformer._samformer_v2 import Samformer

model = Samformer(loss=nn.MSELoss(), hidden_size=512, use_revin=True, metadata=None)

Suggested fix

Add an explicit validation check in the __init__ method, or remove the None default to make metadata a strictly required argument.

if metadata is None:
    raise ValueError("Samformer requires 'metadata' dictionary to be provided for initialization.")

[fnhirwa]

I think we need to revise the docstring of this model as they're too ambiguous to know how the metadata and other arguments are handled. Could you include that in your PR as well?

[lomesh2312]

@fnhirwa
Thanks for the feedback!

Yes, that makes sense. I’ll update the docstring to clearly explain how metadata is expected to be provided and used in the model.

I’ll push the changes shortly.

[lomesh2312]

@fnhirwa
I've pushed the updates including the improved docstring and a small unit test for metadata validation. Please let me know if anything else needs to be adjusted.