HG-CAD: Hierarchical Graph Learning for CAD Material Recommendation

Section 1: Usage

• Setting up the environment from scratch
  • Python == 3.8.5
  • CUDA Toolkit ("nvcc --version") == 11.3
  • PyTorch == 1.12.0

##################################
# --- Ubuntu 20.04 LTS (X86) --- #
##################################

# Basics
sudo apt-get update
sudo apt-get upgrade -y
sudo apt-get dist-upgrade -y
sudo apt-get install build-essential

# Install Anaconda for environment control
wget https://repo.anaconda.com/archive/Anaconda3-2023.03-Linux-x86_64.sh
bash Anaconda3-*.sh
source ~/.bashrc

# Create a conda environment
conda create --name hgcad python=3.8.5
conda activate hgcad

# Install CUDA Toolkit (NVCC), version == 11.3
wget https://developer.download.nvidia.com/compute/cuda/11.3.0/local_installers/cuda_11.3.0_465.19.01_linux.run
sudo sh cuda_11.3.0_465.19.01_linux.run

# Install PyTorch

######################
# --- Windows 10 --- #
######################

# Create conda environment "hg_cad" (Python==3.8.5, CUDA==11.3, PyTorch==1.12.0)
conda env create -f environment.yml

# Activate conda environment
conda activate hg_cad

# Install torch_geometric
conda install pyg -c pyg

• Obtaining the data used in paper:
  • Make sure you have "gdown" installed (conda install -c conda-forge gdown)
  • Unzip (unzip <file>.zip) and place inside a folder (e.g., ./dataset/)
  • If "zipfile corrupt" error: try zip -FF Corrupted.zip --out New.zip (PS: install zip with sudo apt-get install zip)

# Full dataset (on paper)
gdown https://drive.google.com/uc?id=1f9jIgzSHRuT3jPMO17Vl0EYvq4v5kTue

# Top 40% dataset (much smaller, suitable for sanity checks and test developments)
gdown https://drive.google.com/uc?id=1h4iwI8tuOicZhjHGr60mgRu53kUDwEEM

# Obtain the train and val spit - TODO

• Training: Training on train data, and automatically performs testing on test data when training is finished.

# Additional tuning knobs included INSIDE python file -- see classification.py
python classification.py train --dataset_path [path/to/dataset] --max_epoch 100 --batch_size 16 --gpus 1

• Testing: Testing on test data based on checkpoint and random seed obtained from previously finished training.

python classification.py test --dataset_path [path/to/dataset] --checkpoint [path/to/best.ckpt] --random_seed [seed integer]

• Inference: on a single sample assembly, and produce material label predictions for all of its bodies.

# Additional tuning knobs included INSIDE python file -- see inference.py
python inference.py single_sample --inference_sample [path/to/inference/sample] --checkpoint [path/to/best.ckpt] --vocab [path/to/vocab.pickle]

• Inference: on multiple sample assemblies, and produce material label predictions for one random body per assembly.

python inference.py multiple_sample --inference_sample [path/to/inference/samples] --checkpoint [path/to/best.ckpt] --vocab [path/to/vocab.pickle]

---

Section 2: Data Processing Tools for HG-CAD

Tool	Description

---

Section 3: Baselines

Model	Data Format	Repository	Data Processing Tools	Results
PointNet *	PointCloud OFF Files	Adapted Version	OBJ to OFF	Link
UV-Net *	DGL Graph BIN Files	Adapted Version	STEP to BIN; Visualization	Link
Human Baseline	OBJ and PNG files	Original Version	Prompt Generation	-