You can deploy this model and start earning money today!

mohansshf/
dpss-archcad

Name: dpss-archcad
Availability: InStock

2503.22346

Views

No views yet

DPSS — Dual-Pathway Symbol Spotter for ArchCAD

Model weights for the DPSS (Dual-Pathway Symbol Spotter) trained on the ArchCAD-40K dataset for panoptic symbol spotting in architectural CAD drawings.

Model Architecture

Component	Architecture	Params
Image branch	HRNet-W48 (pretrained on COCO-Stuff)	~77M
Point branch	PointTransformerV2	~30M
Adaptive fusion	Attention-based	~5M
Decoder	Mask2Former-style	~20M
Total		~130M

Files

dpss_weights.pth — Clean inference weights (state_dict only, ~500MB)
best_full.pth — Full training checkpoint with optimizer state (for resume)
upstream_code.tar.gz — Upstream DPSS code bundle (all model + data loader code)
training_metrics.json — sPQ history from training

Usage

python

import torch
from huggingface_hub import hf_hub_download

# Download weights
weights_path = hf_hub_download("mohansshf/dpss-archcad", "dpss_weights.pth")

# Load (requires the upstream code — included in upstream_code.tar.gz)
ckpt = torch.load(weights_path, map_location="cpu")
model.load_state_dict(ckpt["state_dict"])

Training

bash

# Clone the training code
git clone https://github.com/ArchiAI-LAB/ArchCAD _upstream
# Or use the bundled code:
wget https://huggingface.co/mohansshf/dpss-archcad/resolve/main/upstream_code.tar.gz
tar xzf upstream_code.tar.gz -C upstream/

# Install deps
pip install torch torchvision numpy scipy Pillow munch pyyaml tensorboardX tensorboard tqdm huggingface_hub
pip install 'git+https://github.com/facebookresearch/detectron2.git'

# Train (downloads dataset automatically from jackluoluo/ArchCAD)
python train_portable.py --batch_size 8 --epochs 100

Dataset

Trained on ArchCAD-40K — 41,097 annotated chunks from 5,538 architectural CAD drawings with 30 semantic categories.

Citation

bibtex

@article{Luo2025ArchCAD,
  title={ArchCAD-400K: An Open Large-Scale Architectural CAD Dataset and New Baseline for Panoptic Symbol Spotting},
  author={Luo R, Liu Z, Cheng T, et al.},
  journal={arXiv preprint arXiv:2503.22346},
  year={2025}
}