Xiang Xu, Pradeep Kumar Jayaraman, Joseph G. Lambourne, Karl D.D. Willis, Yasutaka Furukawa
We present a novel generative model for Computer Aided Design (CAD) that 1) represents high-level design concepts of a CAD model as a three-level hierarchical tree of neural codes, from global part arrangement down to local curve geometry; and 2) controls the generation of CAD models by specifying the target design using a code tree. Our method supports diverse and higher-quality generation; novel user controls while specifying design intent; and autocompleting a partial CAD model under construction.
- Linux
- Python 3.8
- CUDA >= 11.4
- GPU with 24 GB ram recommended
- PyTorch >= 1.10
- Install pythonocc following the instruction here (use mamba if conda is too slow).
- Install other dependencies with
pip install -r requirements.txt
We also provide the docker image. Note: only tested on CUDA 11.4.
We use the dataset from DeepCAD for training and evaluation.
The sketch-and-extrude sequences need to be first converted to our obj format following the steps from SkexGen.
Run the following script to download our post-processed DeepCAD data in obj format
python scripts/download.py
After the data is downloaded, run this script to get the solid, profile, loop and CAD model data
sh scripts/process.sh
Run the deduplication script, this will output post-filtered data as train_deduplicate.pkl
sh scripts/deduplicate.sh
Download the ready-to-use post-deduplicate data.
Train the three-level codebook with
sh scripts/codebook.sh
Download our pretrained codebook module from here.
After the codebooks are trained, extract the neural codes corresponding to each training data with
sh scripts/extract_code.sh
Extracted codes from the pretrained model are available here.
Run the following script to train the code-tree generator and model generator for unconditional generation
sh scripts/gen_uncond.sh
Download our pretrained unconditional generation module from here.
For testing, run this script to generate 1000 CAD samples and visualize the results
sh scripts/sample_uncond.sh
For evaluation, uncomment the eval script in sample_uncond.sh
, this would generate > 10,000 samples. Then compute JSD, MMD, and COV scores using eval.sh
. Warning: this step can be very slow.
sh scripts/eval.sh
Please also download the test data and unzip it inside the data
folder. This is required for computing the evaluation metrics.
Train the full model including model encoder for conditional CAD generation
sh scripts/gen_cond.sh
For testing (e.g CAD autocomplete), run this script to generate full CAD model from partial extruded profiles.
sh scripts/sample_cond.sh
This research is partially supported by NSERC Discovery Grants with Accelerator Supplements and DND/NSERC Discovery Grant Supplement, NSERC Alliance Grants, and John R. Evans Leaders Fund (JELF).
If you find our work useful in your research, please cite the following paper
@inproceedings{xu2023hnc,
author = {Xiang Xu, Pradeep Kumar Jayaraman, Joseph G. Lambourne, Karl D.D. Willis, Yasutaka Furukawa},
title = {Hierarchical Neural Coding for Controllable CAD Model Generation},
journal = {ICML},
year = {2023},
}