llama2thedog/BIM-JEPA-pretrained

BIM-JEPA (Pretrained)

Pretrained checkpoint for BIM-JEPA, a self-supervised Joint-Embedding Predictive Architecture (JEPA) for 3D BIM element representation learning.

This is the pretrained foundation model (no downstream head). For finetuned classifiers, see:

• llama2thedog/BIM-JEPA-finetuned-ifcnetcore — 89.37% OA on IFCNetCore
• llama2thedog/BIM-JEPA-finetuned-bimgeom — 92.43% OA on BIMGEOM

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 (code + configs)
• Project page

Architecture

• Encoder: 12-layer Transformer, 384 hidden dim, 6 attention heads
• Predictor: 6-layer Transformer, 192 hidden dim
• Tokenizer: 64 groups × 32 points per group (iterative-nearest grouping)
• Pretraining objective: Smooth-L1 regression against EMA-teacher token embeddings, with 4 masked target blocks per sample

Training

Datasets	IFC-884K (3 parts), IFCNet (no-test split), BIMGEOM (no-test split)
Input	4096 points per object (centered + unit-sphere normalized)
Augmentations	Scale, rotate (z-axis), translate
Epochs	500
Batch size	512
Optimizer	AdamW (lr=1e-4, weight decay 0.05)
Schedule	Linear warmup (50 epochs) + cosine decay
Precision	bf16-mixed
Hardware	4 × GPU (DDP)

Full training hyperparameters are in hparams.yaml.

Files

File	Size	Description
bim_jepa_pretrained.ckpt	364 MB	PyTorch Lightning checkpoint (encoder + predictor + EMA teacher)
hparams.yaml	2 KB	Training hyperparameters

Usage

Clone the GitHub repo first to get the model code and environment, then:

from huggingface_hub import hf_hub_download
from bimjepa.models.bim_jepa import BimJepa

ckpt_path = hf_hub_download(
    repo_id="llama2thedog/BIM-JEPA-pretrained",
    filename="bim_jepa_pretrained.ckpt",
)
model = BimJepa.load_from_checkpoint(ckpt_path)
model.eval()

To finetune for classification, use the configs at configs/BIM-JEPA/classification/ in the GitHub repo. Place the downloaded checkpoint at BIM-JEPA/pretrained/bim_jepa_pretrained.ckpt (the default path in the configs), then:

python -m bimjepa.tasks.classification fit \
    -c configs/BIM-JEPA/classification/ifcnet_classification.yaml

Citation

in progress

License

MIT