_pickle.UnpicklingError when loading .pth weights on PyTorch 2.6+

[LonleyViper]

---

## 🐛 Issue: `_pickle.UnpicklingError` when loading `.pth` weights on PyTorch 2.6+

### Description
When attempting to train EdgeYOLO using a pretrained `.pth` file (e.g. `edgeyolo_tiny_lrelu_coco.pth`) on **PyTorch 2.6+**, the following error occurs:

_pickle.UnpicklingError: Weights only load failed...
Unsupported global: GLOBAL numpy.core.multiarray.scalar...

This is due to a change in **PyTorch 2.6**, which now defaults `torch.load()` to `weights_only=True` and limits deserialization unless explicitly allowed. The older checkpoint format includes `numpy.core.multiarray.scalar`, which is now restricted.

---

### ✅ Workaround / Fix

#### In `/edgeyolo/edgeyolo/models/__init__.py`, inside the `EdgeYOLO.__init__()` method:

**1. Insert before `torch.load(...)`:**
```python
import numpy.core.multiarray
from torch.serialization import add_safe_globals
add_safe_globals({"scalar": numpy.core.multiarray.scalar})

2. Change this line:

torch.load(weights, map_location="cpu")

to:

torch.load(weights, map_location="cpu", weights_only=False)

This resolves the UnpicklingError by explicitly allowlisting the scalar type and disabling the new default weights_only=True.

---

🧪 Environment

• PyTorch: 2.1.2 and 2.6.0
• CUDA: 11.8
• GPU: RTX 4090
• Platform: Docker container with local runtime
• EdgeYOLO: Main branch (May 2025)

---

💡 Suggestion

To future-proof or support others:

• Add a check like:

if hasattr(torch, 'serialization') and hasattr(torch.serialization, 'add_safe_globals'):
    import numpy.core.multiarray
    from torch.serialization import add_safe_globals
    add_safe_globals({"scalar": numpy.core.multiarray.scalar})

• Or document this in the README under “Common Errors with PyTorch 2.6+”

Thanks for your work — this repo rocks 🚀

[SolSearcher]

This worked for me as well! Thanks for the fix.