EulerCAD-Deep-SIGReg
a fine-tune for finding 3D part projections in CAD drawings
2 min readupdated
This model is a fine-tuned version of NicoloLocatelli's EulerCAD-Deep. Built for CAD drawings, it retrieves different projections (views) of the same 3D object by matching their vector representations. Under the hood is a 36.9M-parameter architecture (14 layers, width 384) fine-tuned with the SIGReg loss function from the LeJEPA method, which pushes embeddings (vector representations of data) within a batch toward an isotropic Gaussian distribution. Compared to the base model, it ranks all remaining views of a single part higher overall (boosting mAP), though it hits the exact top match slightly less often. It's a solid fit for ML researchers and engineers working on automated CAD analysis and indexing.
Key features
- Compact 36.9M-parameter architecture (14 layers, width 384), identical to the base EulerCAD-Deep model.
- Fine-tuned with a combined loss function incorporating SIGReg (weighted at w = 0.018), increasing the effective rank of every layer in the network.
- Better overall ranking accuracy for 3D object projections: layer 7 mAP climbed from 0.250 to 0.268 over the baseline.
- Open-source code for inspecting and reproducing training, released alongside the weights under a permissive MIT license.
What you can do
- Search and match different projections of the same 3D part across CAD drawing databases using EulerCAD's built-in analysis tool.
- Use the weights as a starting point for embedding regularization experiments (LeJEPA / SIGReg) in geometric deep learning.
- Compare layer-by-layer representation behavior against the original EulerCAD-Deep on a validation dataset.
Example
git clone [https://github.com/NicoloLocatelli/EulerCAD.git](https://github.com/NicoloLocatelli/EulerCAD.git)
cd EulerCAD
uv sync
uv run eulercad model download deep-sigreg
uv run eulercad analyze runs/deep-sigreg/checkpoints/model.safetensors --split val
Read more
DetailsProjection retrieval metrics (layer 7, val): mAP — 0.268 (vs. 0.250 for Deep), recall@1 — 0.505 (vs. 0.538), recall@10 — 0.734 (vs. 0.760).
Validation loss (next-row) at step 20,000: 0.7191 (vs. 0.7144 for base Deep).
Training details: 20,000 fine-tuning steps with a batch size of 160 drawings, Muon optimizer, and a peak learning rate of 3e-4.
MIT license for both code and weights; training data sourced from the ABC dataset.










Comments