RyanL22/sapiens-bfloat16

Sapiens Exported Model (Schema 7.3)

This repository provides a re-exported checkpoint of the facebook/sapiens segmentation model using PyTorch 2.5.1, ensuring compatibility with modern torch.export.load() workflows.

---

Background

The original SAPIENS checkpoints were exported in PyTorch 2.1.x and use IR schema version 5.1, which causes torch.export.load() to fail on newer PyTorch versions (e.g., 2.2+), due to a mismatch in how versioning is handled internally.

Many users encounter the following error:

ValueError: invalid literal for int() with base 10: b'5.1'

To address this, we provide a re-exported checkpoint using PyTorch 2.5.1, which uses schema version 7.3, fully compatible with current and future versions of PyTorch.

---

Contents

• ..._bfloat16.pt2: Re-exported IR checkpoint
• Compatible with: torch.export.load() in PyTorch ≥ 2.3.0
• Schema version: 7.3

---

How to Load

from torch.export import load
from huggingface_hub import hf_hub_download

model_path = hf_hub_download("RyanL22/sapiens-bfloat16", "pose/checkpoints/sapiens_1b_goliath_best_goliath_AP_639_bfloat16.pt2")
model = load(model_path).module()

🔧 Make sure you are using PyTorch 2.3.0 or higher to ensure schema 7.x compatibility.

Credits Original model: facebook/sapiens

Re-exported by: @RyanL22