ll_stepnet — Overview

ll_stepnet is a neural-network package for processing STEP / B-Rep CAD files, built with a clean separation of concerns. It turns raw STEP text into token IDs, geometric features, and a topology graph, fuses them in a transformer + graph-neural-network encoder, and exposes task-specific heads for classification, property prediction, similarity, captioning, and QA.

The installed top-level package is stepnet (import from stepnet import ...).

Modules

Module	Responsibility
stepnet.tokenizer	Convert STEP text → token IDs
stepnet.features	Extract geometric properties per entity
stepnet.topology	Build entity-reference graphs
stepnet.encoder	Transformer + GNN encoder fusing all representations
stepnet.tasks	Task-specific prediction heads
stepnet.data	Dataset / DataLoader helpers
stepnet.trainer	Training loop

Model architecture

STEP File
  → Tokenizer        → Token IDs → Transformer Encoder → Token Embedding
  → Feature Extractor → Geometric Features
  → Topology Builder  → Graph → Graph Neural Network → Graph Embedding

Token Embedding + Graph Embedding → Fusion Layer → Final Encoding → Task Head

A first taste

from stepnet import STEPTokenizer, STEPFeatureExtractor

tokenizer = STEPTokenizer()
step_text = "#31=CONICAL_SURFACE('',#1837,2.6797,0.7854);"
token_ids = tokenizer.encode(step_text)

extractor = STEPFeatureExtractor()
features = extractor.extract_geometric_features(step_text)
print(features["entity_type"], features["numeric_params"])

Status

First real trained checkpoint + native MLX

ll_stepnet now ships a trained classifier: STEPForClassification trained on real DeepCAD models to predict face-count complexity (≤4 / 5–6 / 7+ faces, a geometric label derived from the reconstructed solid). Validation accuracy 0.976 vs a 0.436 majority-class baseline (per-class 0.991 / 0.978 / 0.933). A native-MLX port (ll_stepnet/mlx/train_classification_mlx.py) loads those exact trained weights and reproduces the PyTorch model on Apple Silicon — verified at 100% argmax agreement, identical 0.976 accuracy (--mode parity). The other task heads (property prediction, similarity, captioning, QA) provide architectures + trainers; train them on your data before relying on their outputs.

stepnet also provides the generative models (STEPVAE, StructuredDiffusion, VQVAEModel, CADGenerationPipeline). The valid-CAD generation path now lives in ll_gen as trained, program-based generators (autoregressive command model + latent diffusion).

Use the sidebar for Installation, Usage, and the API Reference.