Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
SolidGen: An Autoregressive Model for Direct B-rep Synthesis
Authors: Pradeep Kumar Jayaraman, Joseph George Lambourne, Nishkrit Desai, Karl Willis, Aditya Sanghi, Nigel J. W. Morris
TMLR 2023 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate qualitatively, quantitatively, and through perceptual evaluation by human subjects that Solid Gen can produce high quality, realistic CAD models. In this section we perform experiments to qualitatively and quantitatively evaluate our method on unconditional generation, and various conditional generation tasks based on class labels, images, and voxels. |
| Researcher Affiliation | Collaboration | Pradeep Kumar Jayaraman EMAIL Autodesk Research Joseph G. Lambourne EMAIL Autodesk Research Nishkrit Desai EMAIL University of Toronto, Vector Institute Karl D.D. Willis EMAIL Autodesk Research Aditya Sanghi EMAIL Autodesk Research Nigel J.W. Morris EMAIL Autodesk Research |
| Pseudocode | No | The paper describes the architecture and various processes (e.g., "Recovering a Boundary Representation", "Masking Invalid Logits") in detail using prose and mathematical equations but does not present any formal pseudocode or algorithm blocks. |
| Open Source Code | No | Our implementation is in Py Torch (Paszke et al., 2019). The paper states the framework used but does not provide a link or an explicit statement about the availability of their code. |
| Open Datasets | Yes | The Deep CAD Dataset (Wu et al., 2021) contains a subset of the CAD models available in the ABC dataset (Koch et al., 2019) that additionally includes the sequence of sketch and extrude CAD modeling operations. |
| Dataset Splits | Yes | The dataset is split in a 90 (train)/5 (validation)/5 (test) proportion. We are left with 49,759 models that are split in 90/5/5 train/validation/test sets and used for all experiments. |
| Hardware Specification | Yes | We train our models for 1000 epochs with batch size 512 using the Adam W optimizer (Loshchilov & Hutter, 2019) (learning rate: 10 4, weight decay: 0.01) on an Nvidia DGX A100 machine. |
| Software Dependencies | No | Our implementation is in Py Torch (Paszke et al., 2019). All data processing related to building indexed B-reps and reconstructing B-reps uses the Open Cascade/python OCC (Paviot, 2008) solid modeling kernel. Specific version numbers for PyTorch and python OCC are not provided. |
| Experiment Setup | Yes | We train our models for 1000 epochs with batch size 512 using the Adam W optimizer (Loshchilov & Hutter, 2019) (learning rate: 10 4, weight decay: 0.01) on an Nvidia DGX A100 machine. All Transformer modules use 8 layers with an embedding dimension of 256, and fully-connected dimension of 512 and 8 attention heads. The model is trained using teacher-forcing (Williams & Zipser, 1989) to maximize the log-likelihood of the training data with a cross-entropy loss with label smoothing of 0.01. |