Lux.jl implementation of Denoising Diffusion Implicit Models (arXiv:2010.02502).
The implementation follows the Keras example.
The model generates images from Gaussian noises by denoising iteratively.
Install Julia and instantiate Project.toml.
Following scripts are tested on a single NVIDIA V100 instance with 32GB of GPU memory. You may need to adjust the batch size and learning rate for your environment.
We use the dataset from 102 Category Flowers Dataset. The user is prompted to download the dataset when running the code for the first time. The dataset is cached for subsequent runs.
julia --startup-file=no \
--project=examples/DDIM \
--threads=auto \
main.jl \
--expt-dir outputThis code runs in about 35 minutes in a single NVIDIA V100 instance with 32GB of GPU memory. We recommend running the code for atleast 80 epochs to get good results.
julia --startup-file=no \
--project=. \
--threads=auto \
main.jl \
--inference-mode \
--saved-model-path output/checkpoints/model_100.jld2 \
--generate-n-images 24 \
--generate-image-seed 12 \
--expt-dir outputusage: main [options] [flags]
Options
--epochs <100::Int>
--image-size <128::Int>
--batchsize <128::Int>
--learning-rate-start <0.001::Float32>
--learning-rate-end <0.001::Float32>
--weight-decay <1.0e-6::Float32>
--checkpoint-interval <25::Int>
--expt-dir <tempnam...>
--diffusion-steps <80::Int>
--generate-image-interval <5::Int>
--channels <[32, 64...::Vector{Int}>
--block-depth <2::Int>
--min-freq <1.0::Float32>
--max-freq <1000.0::Float32>
--embedding-dims <32::Int>
--min-signal-rate <0.02::Float32>
--max-signal-rate <0.95::Float32>
--generate-image-seed <12::Int>
--saved-model-path <nothing>
--generate-n-images <10::Int>
Flags
--inference-mode
-h, --help Print this help message.
--version Print version.