maennyn/personvit-reid-msmt17-vit-s

---

license: apache-2.0 library_name: transformers tags: - person-reid - re-identification - vision-transformer - custom-code pipeline_tag: image-feature-extraction

PersonViT ReID MSMT17 ViT-S

This repository is a simple Hugging Face port of a PersonViT / TransReID person re-identification model.

The port was AI-generated and very lightly packaged for inference use. It is essentially a Codex vibe-coded Hugging Face wrapper around the original checkpoint format.

Source

Original upstream code / paper implementation reference:

• https://github.com/hustvl/PersonViT

Attribution

This repository is a Hugging Face Transformers port of the original PersonViT / TransReID model.

Original work:

• PersonViT: Large-scale Self-supervised Vision Transformer for Person Re-Identification
• Authors: Bin Hu, Xinggang Wang, Wenyu Liu
• Original repository: https://github.com/hustvl/PersonViT
• Paper: https://arxiv.org/abs/2408.05398

The original work is licensed under the Apache License 2.0. This repository preserves that license.

What This Repo Contains

This repo contains:

• a Hugging Face config.json
• preprocessor_config.json
• custom configuration_personvit_reid.py
• custom modeling_personvit_reid.py
• converted model weights in model.safetensors

Intended Use

This model is intended for:

• person crop embedding extraction
• person re-identification / tracking
• loading via:

from transformers import AutoModel, CLIPImageProcessor

model = AutoModel.from_pretrained(
    "your-username/personvit-reid-msmt17-vit-s",
    trust_remote_code=True,
)
processor = CLIPImageProcessor.from_pretrained(
    "your-username/personvit-reid-msmt17-vit-s"
)

## Notes

This is not an official upstream Hugging Face release.
It is a custom inference port for practical integration.

Use trust_remote_code=True when loading.




# PersonViT ReID Hugging Face Scaffold

This directory is an inference-only Hugging Face model-repo scaffold for a
person ReID model intended to be loaded through:

```python
from transformers import AutoModel, CLIPImageProcessor

model = AutoModel.from_pretrained(path_or_repo_id, trust_remote_code=True)
processor = CLIPImageProcessor.from_pretrained(path_or_repo_id)

The forward contract is intentionally server-friendly:

embeddings, _ = model(pixel_values)

What Is Included

• configuration_personvit_reid.py
• modeling_personvit_reid.py
• config.json
• preprocessor_config.json
• export_personvit_reid.py

Intended Target

The default exporter preset is shaped for the planned first target:

• dataset: MSMT17
• backbone class: ViT-S
• crop geometry: 128x256

Current Constraint

The repository does not include the original fine-tuned MSMT17 checkpoint. That means this scaffold can be validated locally with random weights today, but the final production export still requires the real PersonViT/TransReID checkpoint at export time.

Export

Random-weight validation scaffold:

python -m server.model_repos.personvit_reid.export_personvit_reid \
  --output-dir /tmp/personvit-reid-msmt17-vit-s \
  --init-random

Checkpoint-backed export:

python -m server.model_repos.personvit_reid.export_personvit_reid \
  --output-dir /tmp/personvit-reid-msmt17-vit-s \
  --checkpoint /path/to/checkpoint.pth

The exporter writes:

• config.json
• preprocessor_config.json
• configuration_personvit_reid.py
• modeling_personvit_reid.py
• README.md
• model.safetensors

Notes

• This is intentionally a minimal inference package, not the upstream training tree.
• If the upstream checkpoint uses different tensor names, adapt the checkpoint mapping in export_personvit_reid.py.
• The server integration can use config.reid_target_size and config.reid_aspect_ratio to preserve person-style crop geometry.