dongjin-seo2020/adjointdiffusion

How dongjin-seo2020/adjointdiffusion works

AdjointDiffusion generates fabrication-aware structures with diffusion models, then guides reverse sampling using electromagnetic simulation and adjoint gradients. The documented workflow runs from dataset generation through model training to sampling and optimization; baseline optimizers provide comparison methods. The graph includes the central diffusion, training, simulation, and baseline modules. Some connections are based on the README or SOURCE INDEX because their implementation files were not excerpted; uncertain wiring is omitted.

Main entry points: image_sample.py Sampling entry · image_train.py Training entry · train_util.py Training loop · script_util.py Model factory

Components 24

Design workflow

Dataset generation dataset_generation.py
Fabrication data
Sampling entry image_sample.py
Adjoint simulation guided_diffusion/simulation.py
Simulation module simulation.py

Diffusion model

Model factory guided_diffusion/script_util.py
UNet model guided_diffusion/unet.py
Diffusion process guided_diffusion/gaussian_diffusion.py
Timestep spacing guided_diffusion/respace.py
Diffusion losses guided_diffusion/losses.py
Network utilities guided_diffusion/nn.py

Training runtime

Training entry image_train.py
Image loader guided_diffusion/image_datasets.py
Schedule sampler guided_diffusion/resample.py
Training loop guided_diffusion/train_util.py
Mixed precision guided_diffusion/fp16_util.py
Distributed utilities guided_diffusion/dist_util.py
Experiment logger guided_diffusion/logger.py

Baseline optimization

GA CIS baseline baseline_algorithms/[adjoint_optimization] GA_CIS.py
GA waveguide baseline baseline_algorithms/[adjoint_optimization] GA_waveguide.py
NLopt CIS baseline baseline_algorithms/[adjoint_optimization] nlopt_CIS.py
NLopt waveguide baseline baseline_algorithms/[adjoint_optimization] nlopt_waveguide.py

Other components

Researcher
Weights & Biases

Connections 28

Researcher → Dataset generation: generates README.md
Dataset generation → Fabrication data: writes structures README.md
Researcher → Training entry: starts training README.md
Fabrication data → Image loader: loads images
Training entry → Image loader: loads data image_train.py
Training entry → Model factory: creates model image_train.py
Training entry → Schedule sampler: creates sampler image_train.py
Training entry → Training loop: starts loop image_train.py
Model factory → UNet model: constructs guided_diffusion/script_util.py
Model factory → Diffusion process: constructs guided_diffusion/script_util.py
Model factory → Timestep spacing: uses spacing guided_diffusion/script_util.py
Diffusion process → Adjoint simulation: requests gradients guided_diffusion/gaussian_diffusion.py
Training loop → Diffusion process: computes training loss guided_diffusion/train_util.py
Training loop → Mixed precision: trains parameters guided_diffusion/train_util.py
Training loop → Distributed utilities: synchronizes ranks guided_diffusion/train_util.py
Training loop → Experiment logger: logs progress guided_diffusion/train_util.py
Schedule sampler → Diffusion process: samples timesteps guided_diffusion/resample.py
UNet model → Network utilities: uses utilities guided_diffusion/unet.py
Researcher → Sampling entry: starts sampling README.md
Sampling entry → Model factory: creates model image_sample.py
Sampling entry → Diffusion process: samples designs README.md
Sampling entry → Adjoint simulation: configures simulation
Sampling entry → Experiment logger: logs outputs image_sample.py
Adjoint simulation → Weights & Biases: tracks experiments README.md
Researcher → GA CIS baseline: runs baseline README.md
Researcher → GA waveguide baseline: runs baseline README.md
Researcher → NLopt CIS baseline: runs baseline README.md
Researcher → NLopt waveguide baseline: runs baseline README.md

Drawn by GitDiagram on November 10, 2026 from the repository's file tree, README and source files. Read as Markdown ·