/Easing-Color-Shifts

Repo for "Easing Color Shifts in Score-Based Diffusion Models" project. Trains a score-based diffusion model on FashionMNIST and demonstrates how simple network architectures can perform well on large images.

Primary LanguageJuliaMIT LicenseMIT

Easing Color Shifts with Score-Based Diffusion Models

Note: CUDA is expected to be available.

Setting up the project

Switch to the examples/ folder.

julia --project -e 'using Pkg; Pkg.develop(path="../")'
julia --project -e 'using Pkg; Pkg.instantiate()'
julia --project -e 'using Pkg; Pkg.build()'
julia --project -e 'using Pkg; Pkg.precompile()'

Training a model

Switch to the examples folder. Decide on which experiment TOML to run from /examples/fashion_mnist/experiments.

julia --project training.jl EXPERIMENT_TOML

Analyzing a model

Switch to the examples folder. Decide on which experiment TOML to run from /examples/fashion_mnist/experiments_100 or /examples/fashion_mnist/experiments_30.

Basic analysis for spot checking

julia --project analysis_smooth.jl EXPERIMENT_TOML

Paper submission analysis

julia --project analysis_paper_smooth.jl EXPERIMENT_TOML
julia --project eval_smooth_loss.jl EXPERIMENT_TOML

Plotting

Plotting scripts to regenerate figures from submission are available in examples/fashion_mnist/plots, but may require setting of custom paths to data. Switch to the examples/fashion_mnist/plots folder for the plotting scripts.