SkexGen: Autoregressive Generation of CAD Construction Sequences with Disentangled Codebooks

Xiang Xu, Karl D.D. Willis, Joseph G. Lambourne, Chin-Yi Cheng, Pradeep Kumar Jayaraman, Yasutaka Furukawa

ICML 2022
project | paper | youtube

Installation

Environment

• Linux
• Python >= 3.8
• PyTorch >= 1.10

Dependencies

• Install PyTorch 1.10 with the correct CUDA version.
• Install other dependencies:

  pip install -r requirements.txt
  

• Install pythonocc following the instruction here.

Docker

We also provide the docker image for running SkexGen. You can download it from dockerhub (~10GB).
Note: only tested on CUDA 11.4.

Data

Download the raw json data from DeepCAD. Unzip it into the data folder in the root of this repository. Also download the and train_val_test_split.json and place this in the data folder as well.

Follow these steps to convert DeepCAD data to SkexGen format:

# Under utils folder:

# parse DeepCAD json to a simple obj format 
  python convert.py --data_folder ../data/cad_json --output_folder ../data/cad_obj

# normalize CAD and update the obj file
  python normalize.py --data_folder ../data/cad_obj --out_folder ../data/cad_norm

# parse obj to primitive sequence 
  python parse.py --input ../data/cad_norm --output ../data/cad_data --bit 6

# remove duplicated sketch data
  python deduplicate.py --datapath ../data/cad_data --hash_type s

# remove duplicated extrude data
  python deduplicate.py --datapath ../data/cad_data --hash_type e

# Find all the invalid CAD models
  python invalid.py --datapath ../data/cad_data --bit 6

When running convert.py some files in the DeepCAD dataset fail to generate valid solid models. You may use the the --verbose option to see additional details about the problem files. If the convert.py script hangs during processing it can be safely restarted and will continue from where it left off.

You can download the already pre-processed data

Training

Train sketch branch (topology encoder, geometry encoder, sketch decoder):

  python train_sketch.py --train_data data/cad_data/train_deduplicate_s.pkl \
                         --output proj_log/exp_sketch \
                         --invalid data/cad_data/train_invalid.pkl \
                         --val_data data/cad_data/val.pkl \
                         --bit 6 --maxlen 200 --batchsize 128 --device 0

maxlen: sketch sequence length (default 200)

Train extrude branch (extrude encoder, extrude decoder):

  python train_extrude.py --train_data data/cad_data/train_deduplicate_e.pkl \
                          --val_data data/cad_data/val.pkl \
                          --output proj_log/exp_extrude \
                          --bit 6 --maxlen 5 --batchsize 128 --device 0

maxlen: number of extudes (default 5)

Extract codes:

  python extract_code.py --sketch_weight proj_log/exp_sketch \
                         --ext_weight proj_log/exp_extrude \
                         --device 0 --maxlen 200 --bit 6 \
                         --output proj_log/exp_code \
                         --data data/cad_data/train.pkl \
                         --invalid data/cad_data/train_invalid.pkl 

Train code selector (random generation):

  python train_code.py --input proj_log/exp_code/code.pkl \
                       --output proj_log/exp_code \
                       --batchsize 512 --device 0 \
                       --code 1000 --seqlen 10

seqlen: 4 topology, 2 geometry, 4 extrude, code: max size of codebook is 1000

Download our pretrained models

Evaluation

Random generation:

# sample the codes and autoregressively decode it to sketch and extrude
  python sample.py --sketch_weight proj_log/exp_sketch \
                      --ext_weight proj_log/exp_extrude \
                      --code_weight proj_log/exp_code \
                      --device 1 --bit 6 \
                      --output proj_log/samples 

Visualization:

# Under utils folder:

# convert generated sketch-and-extrude to stl format (timeout prevent occ hanging)
  timeout 180 python visual_obj.py --data_folder ../proj_log/samples 

# render and visualize to images 
  python cad_img.py  --input_dir ../proj_log/samples --output_dir ../proj_log/samples_visual

Evaluate the CAD models (after running visual_obj.py):

# Under utils folder:

# uniformly sample 2000 points 
  python sample_points.py --in_dir ../proj_log/samples --out_dir pcd

# evaluate performance 
  python eval_cad.py --fake ../proj_log/samples \
                     --real ../data/test_eval

Download test_eval and unzip it under the data folder. This contains the point clouds from DeepCAD test set.

Citation

If you find our work useful in your research, please cite our paper SkexGen:

@inproceedings{xu2022skexgen, 
title     = {SkexGen: Autoregressive Generation of CAD Construction Sequences with Disentangled Codebooks},
author    = {Xu, Xiang and Willis, Karl DD and Lambourne, Joseph G and Cheng, Chin-Yi and Jayaraman, Pradeep Kumar and Furukawa, Yasutaka},
booktitle = {International Conference on Machine Learning},
pages={24698--24724},
year={2022},
organization={PMLR}
}

License

Please see the license for further details.