How torchopc/torchlitho works
Project type and purpose
• This is an open-source research framework for differentiable computational lithography built in Python.
• It exposes configurable imaging models (Abbe, Hopkins, TCC), analytical resist modeling, and vector lithography routines.
• It leverages PyTorch (+ Lightning) for GPU acceleration and Hydra for flexible experiment management.
File structure
• Top-level:
• configs/ – hundreds of YAMLs for callbacks, data, trainer settings, loggers, experiments
• src/ – core code: data modules, model definitions, training & eval scripts, utilities
• scripts/ – wrapper scripts (e.g., schedule.sh)
• notebooks/, misc/ – Jupyter demos, slide decks, figures
• tests/ – pytest suites for configs, data, training, evaluation
• environment.yaml, requirements.txt, setup.py/pyproject.toml – dependency/build spec
• src/data: PyTorch Lightning DataModule(s), data components
• src/models:
• components/simple_dense_net.py (ML baseline)
• litho/: GDSII parsing, layer/material definitions, numerical functions (Calculate*Image, TCC matrices)
• mnist_module.py, gdsii, litho imaging classes (FilmStack, ImagingModel, Layer, Mask, Resist, Source, etc.)
• src/train.py & src/eval.py – entrypoints wired via Hydra
• src/utils & configs/hydra: configuration instantiation, logging, CLI integration
README
• Confirms heavy use of PyTorch Lightning, Hydra 1.3, multiple logger backends (MLflow, WandB, Comet, Aim).
• Usage patterns: “python src/train.py trainer.max_epochs=20 data.batch_size=64”
• Emphasis on first‐of‐its‐kind differentiable lithography imaging & resist modeling.
1. Configuration Layer (Hydra)
2. Data Layer
– MNIST DataModule
– GDSII parsing & dataset components
3. Model Layer
– SimpleDenseNet (baseline)
– Lithography Imaging Model
• Optics: FilmStack, ProjectionObjective, Source
• Resist: Resist, Receipe, Numerics
• GDSII handling: record ↔ element ↔ structure
• Imaging functions (Calculate1DAerialImage, …)
4. Training Orchestrator
– PyTorch Lightning Trainer
– Callbacks (early stopping, checkpoints, progress bars)
– Logger plugins (TensorBoard, MLflow, WandB, Comet, Aim)
5. Evaluation & Visualization
– src/eval.py
– Jupyter notebooks for demos (“Imaging.ipynb”)
6. CLI & Scripts
– train.py / eval.py entrypoints
– schedule.sh for job orchestration
• Hydra config → injects into DataModule, Model, Trainer
• DataModule → yields batched tensors to Model
• Model.forward → returns predictions → loss → Trainer
• Trainer → invokes Callbacks & Logger → writes checkpoints & metrics
• Eval script or notebook → loads checkpoint → runs forward → generates visualizations
c. Architectural patterns & design principles:
– Modular layering: separation of data, model, training orchestration, evaluation
– Plugin architecture via Hydra (config‐over‐code)
– Abstraction of training loop via Lightning Trainer & callbacks
– Extensible logger interface (multiple backends)
– Functional decomposition of imaging routines
d. Key technologies, frameworks, libraries:
– Python 3.8–3.10, PyTorch ≥2.0, PyTorch Lightning ≥2.0
– Hydra 1.3 for config management
– pytest for testing
– CI via pre-commit, black style
– Logging: TensorBoard, MLflow, WandB, Comet, Aim
Final advice
– Be exhaustive: split “Model” into its subcomponents (Optics, Resist, GDSII).
– Capture both control-plane (Hydra/Cli) and data-plane (tensors/parameters) flows.
– Review the “configs/experiment” folder to show how different experiments slot into this architecture seamlessly.