---
license: mit
library_name: pytorch
tags:
  - point-cloud
  - 3d
  - classification
  - bim
  - aec
  - building-information-modeling
datasets:
  - bimgeom
---

# BIM-JEPA (Finetuned on BIMGEOM)

Classification model finetuned from [`llama2thedog/BIM-JEPA-pretrained`](https://huggingface.co/llama2thedog/BIM-JEPA-pretrained) on the **BIMGEOM** dataset.

## Performance (test set)

| Metric | Value |
|---|---|
| **Overall Accuracy** | **92.43%** |
| Mean Class Accuracy | 89.53% |
| Macro Precision / Recall / F1 | 89.05 / 89.53 / 89.20 |
| Weighted Precision / Recall / F1 | 92.57 / 92.43 / 92.44 |

## Paper

**Self-supervised learning for BIM element classification using a joint embedding predictive architecture**  
Jack Wei Lun Shi, Wawan Solihin, Yufeng Weng, Yimin Zhao, Leong Hien Poh, Justin K.W. Yeoh  
*Automation in Construction*

- [GitHub repo](https://github.com/jackswl/bim-jepa) (code + configs)
- [Project page](https://jackswl.github.io/bim-jepa/)
- [Pretrained base](https://huggingface.co/llama2thedog/BIM-JEPA-pretrained)

## Architecture

- **Backbone**: BIM-JEPA encoder (12-layer Transformer, 384 dim, 6 heads) — initialized from the pretrained checkpoint
- **Head**: MLP classifier, 256 hidden dim, mean+max pooling, 0.5 dropout
- **Loss**: Cross-entropy with label smoothing (0.1)
- **Encoder schedule**: Frozen for the first 175 epochs, then unfrozen and finetuned end-to-end

## Training

| | |
|---|---|
| Dataset | BIMGEOM |
| Input | 4096 points per object |
| Augmentations | Scale, rotate (all axes), translate |
| Epochs | 350 |
| Batch size | 32 |
| Optimizer | AdamW (head lr=1e-3, encoder lr=1e-4, weight decay 0.05) |
| Schedule | Linear warmup (10 epochs) + cosine decay |
| Precision | bf16-mixed |

Full training hyperparameters are in [`hparams.yaml`](./hparams.yaml).

## Files

| File | Size | Description |
|---|---|---|
| `bim_jepa_finetuned_bimgeom.ckpt` | 256 MB | PyTorch Lightning checkpoint |
| `hparams.yaml` | 2 KB | Training hyperparameters |

## Usage

Clone the [GitHub repo](https://github.com/jackswl/bim-jepa) first to get the model code, then:

```python
from huggingface_hub import hf_hub_download
from bimjepa.models.classification import BimJepaClassification

ckpt_path = hf_hub_download(
    repo_id="llama2thedog/BIM-JEPA-finetuned-bimgeom",
    filename="bim_jepa_finetuned_bimgeom.ckpt",
)
model = BimJepaClassification.load_from_checkpoint(ckpt_path)
model.eval()
```

## Citation

```
in progress
```

## License

MIT