Code for BrepMFR: Enhancing machining feature recognition in B-rep models through deep learning and domain adaptation.
BrepMFR, a novel deep learning network designed for machining feature recognition on B-rep models within the CAD/CAM domain. The original B-rep model is converted into a graph representation for network-friendly input, where graph nodes and edges respectively correspond to faces and edges of the original model. Subsequently, we leverage a graph neural network based on the Transformer architecture and graph attention mechanism to encode both local geometric shape and global topological relationships, achieving high-level semantic extraction and prediction of machining feature categories. Furthermore, to enhance the performance of neural networks on real-world CAD models, we adopt a two-step training strategy within a novel transfer learning framework.
git clone https://github.com/zhangshuming0668/BrepMFR.git
cd BrepMFR
conda env create -f environment.yml
conda activate brep_mfr
Our synthetic CAD datasets have been publicly available on Science Data Bank
For machining feature recognition, the network can be trained using:
python segmentation.py train --dataset_path /path/to/dataset --max_epochs 1000 --batch_size 64
The logs and checkpoints will be stored in a folder called results/BrepMFR based on the experiment name and timestamp, and can be monitored with Tensorboard:
tensorboard --logdir results/<experiment_name>
The best checkpoints based on the smallest validation loss are saved in the results folder. The checkpoints can be used to test the model as follows:
python segmentation.py test --dataset_path /path/to/dataset --checkpoint ./results/BrepMFR/best.ckpt --batch_size 64