How blaise143/stochastic-interpolant-diffusion works
This repository trains and samples stochastic-interpolant flow/diffusion models for EMNIST digits. The principal path is configuration-driven training or sampling: training loads and preprocesses digits, learns a conditional vector field, and saves a checkpoint; sampling loads that model and integrates it with an ODE or SDE solver to produce images. The graph preserves the documented solver, scheduler, classifier-free-guidance, attention, distributed-training, and logging capabilities. Relationships involving unsampled configuration and model internals are limited to README or source-index evidence.
Main entry points: main.py Run controller · flow_trainer.py Flow trainer · flow.py Flow model · dataset.py EMNIST loader
Components 20
Application entry
- Run controller
main.py - Run configuration
config/config.yaml
Digit data
- EMNIST loader
data/dataset.py - Image transforms
data/transforms.py - Dataset abstractions
data/base.py
Modeling
- Flow model
models/flow.py - Interpolant base
models/stochastic_interpolant_base.py - U-Net field
models/unet.py - Time and labels
models/embedding.py - Self-attention
models/attention.py - Path schedulers
models/schedulers.py
Training and sampling
- Flow trainer
training/flow_trainer.py - Experiment logger
training/logger.py - Solver loop
sampling/base_solver.py - ODE integrators
sampling/ode_solver.py - SDE integrator
sampling/sde_solver.py - Model checkpoint
checkpoints/flow_model.pth - Generated images
images
Other components
- Research user
- Logging backends
Connections 24
- Research user → Run controller: runs
README.md - Run controller → Run configuration: loads settings
main.py - Run controller → Flow trainer: starts training
main.py - Run controller → Flow model: constructs model
main.py - Run controller → ODE integrators: selects flow solver
README.md - Run controller → SDE integrator: selects diffusion solver
README.md - Run controller → Model checkpoint: loads weights
main.py - Run controller → Generated images: saves images
main.py - Flow trainer → EMNIST loader: loads batches
training/flow_trainer.py - EMNIST loader → Image transforms: applies preprocessing
data/dataset.py - EMNIST loader → Dataset abstractions: extends abstractions
data/dataset.py - Flow trainer → Flow model: trains model
training/flow_trainer.py - Flow trainer → ODE integrators: samples during training
training/flow_trainer.py - Flow trainer → Experiment logger: logs metrics
training/flow_trainer.py - Flow trainer → Model checkpoint: saves weights
training/flow_trainer.py - Flow model → Interpolant base: extends
models/flow.py - Flow model → U-Net field: builds field network
models/flow.py - Flow model → Path schedulers: selects path scheduler
models/flow.py - U-Net field → Time and labels: uses embeddings
models/unet.py - U-Net field → Self-attention: uses optional attention
models/unet.py - Solver loop → Flow model: evaluates field
sampling/base_solver.py - ODE integrators → Solver loop: inherits loop
sampling/ode_solver.py - SDE integrator → Solver loop: inherits loop
sampling/sde_solver.py - Experiment logger → Logging backends: optionally reports
training/logger.py