• Python 3.9
• Cuda 11.8+

Install python package dependencies through pip:

Download data from here and extract them under data folder.

• svg_raw contains the engineering drawings of each CAD model in SVG format, including four views: Front, Top, Right, and FrontTopRight. Each SVG file has been preprocessed through path simplification and deduplication, path reordering, and viewbox normalization. To obtain engineering drawings in PNG format, you can simply convert them using CairoSVG with a single line of code:

  import cairosvg
  
  cairosvg.svg2png(url=svg_path, write_to=png_path, output_width=224, output_height=224, background_color='white')

• svg_vec contains vectorized representations of SVG drawing sequences. Each file stores the stacked drawing sequences for the four views (Front, Top, Right, and FrontTopRight), saved in .npy format to enable fast data loading.

• cad_vec contains our vectorized representation for CAD sequences, saved in .h5 format to enable fast data loading.

[Update] All CAD models used in this project are sourced from DeepCAD. We have also released our pipeline and scripts here for exporting and preprocessing SVG drawings, making them available for anyone interested in converting their own CAD model data.

To train the model in different input options:

Since different input options lead to different models, it is recommended to specify the experiment name using --exp_name for each run. For more configurable parameters and options, please refer to config/config.py.

After training the model, run the model to inference all test data:

After inference, the final results will be saved under proj/your_exp_name/test_results. To evaluate the model inference results and to export and visualize the final CAD models, please refer to the code from DeepCAD.

This repository builds upon the following awesome datasets and projects:

• DeepCAD
• FreeCAD
• CairoSVG

If you find this project useful for your research, please use the following BibTeX entry.