Denoising Diffusion Probabilistic Models

This repository contains the code for the seminar paper "Denoising Diffusion Probabilistic Models" that was implemented as part of the course "Project Machine Learning" in WS23/24. The reports for the three milestones can be found in the docs/reports directory.

Installation

pip install . -r requirements.txt

This will install the dependencies listed in requirements.txt.

Training

python3 train_denoiser.py <dataset_name> [options]

Train the denoising module on one of the provided datasets, i.e. CIFAR-10, MNIST, FGVC-Aircraft, Flowers102 or CelebA. The dataset is split into a training set, which is used to train the denoiser, and a test set, which is used to validate the model at end of each epoch. After the training is complete, the training results (metadata, losses, etc.) are stored together with the model weights in a directory named after the experiment.

Positional Parameters

Parameter	Description	Type	Annotations
dataset	The name of the dataset to use.	"cifar10" | "mnist" | "fgvc-aircraft" | "flowers102" | "celeba"

Options

Parameter	Description	Type	Default	Annotations
--run-name	The name of the experiment. If not provided, the name will be generated automatically.	str	None	Generated names have the pattern $timestamp-$dataset.
--image-size	The (square) sample size of the images.	int	32	If the provided image size is different from the dataset's image size, the images will be up- or downsized accordingly.
--time-steps	The number of time steps for the diffusion process.	int	1000
--schedule	The schedule type to use.	"linear" | "polynomial" | "cosine" | "sigmoid"	"linear"
--schedule-start	The start value for the schedule within [0,1].	float	0.0001
--schedule-end	The end value for the schedule within [0,1].	float	0.2
--schedule-tau	The tau value for the schedule	float | None	None	Only applicable for polynimial, cosine and sigmoid schedules.
--epochs	The number of epochs to train for.	int	1000
--subset-size	The number of samples to use from the dataset. If not provided, the entire dataset is used.	int | None	None
--train-split	The percentage in [0, 1] of the dataset to use for training.	float	0.8
--disable-validation	Disable validation on the test set during training.	bool	False
--batch-size	The batch size for the data loader.	int	16
--learning-rate	The learning rate for the optimizer.	float	2e-4
--dropout-rate	The dropout rate for the model.	float	0.0
--seed	The random seed to use. If not specified, the randomized calculations will be non-deterministic.	int | None	None
--out-dir	The directory to save the results to.	pathlib.Path	./out/train
--device	The device to use.	"cpu" | "cuda" | None	None	If not specified, "cuda" will be used if it is available.
--verbose	Flag for enabling verbose logging .	bool	False

Image Generation

python3 generate_images.py <run_dir> <image_count> [options]

Generate images using a trained denoiser. The routine takes the path of a run created by train_denoiser.py and the number of images to generate. The resulting images are saved to the specified output directory.

Parameters

Parameter	Description	Type	Annotations
run_dir	The directory containing the trained model.	Path
image_count	The number of images to generate.	int

Options

Parameter	Description	Type	Default	Annotations
--batch-size	The batch size for generating images. Must be a divisor of num_images.	int	None
--time-steps	The number of time steps for the diffusion process. If not specified, number of time steps from the training will be used.	int | None	None
--seed	The random seed to use. If not specified, the randomized calculations will be non-deterministic.	int | None	None
--out-dir	The directory to save the results to. If not specified, the images are saved to <run_dir>/samples.	pathlib.Path	None
--device	The device to use.	"cpu" | "cuda" | None	"cuda"	If not specified, "cuda" will be used if it is available.

Image Evaluation

python3 evaluate_images.py <image_dir> [options]

Calculate the Inception Score to evaluate the quality of generated images using the InceptionV3 model. The routine takes the path of a directory containing the generated images and saves their resulting score to inception_score.txt in the same directory.

Positional Parameters

Parameter	Description	Type	Annotations
image_dir	The directory containing the generated images.	Path

Options

Parameter	Description	Type	Default	Annotations
--out-file	The file to save the inception score to. If not specified, the score is saved to <image_dir>/inception_score.txt.	pathlib.Path	None
--device	The device to use.	"cpu" | "cuda" | None	"cuda"	If not specified, "cuda" will be used if it is available.
--verbose	Enable verbose logging.	bool	False

Noise Schedule Visualization

python3 apply_noise.py <image_path> [options]

This routine visualizes the forward diffusion process. It takes an image as parameter (e.g. one of the example images in data/images) and applies gradual noise according to the pre-configured schedule.

Positional Parameters

Options

Parameter	Description	Type	Default	Annotations
--schedule	The schedule to use	linear, polynomial, cosine, sigmoid	linear
--schedule-steps	The number of time steps for the diffusion process	int	10
--schedule-start	The starting value for the schedule	float	0.0
--schedule-end	The ending value for the schedule	float	1.0
--schedule-tau	The tau value for the schedule	float	None	Only applicable for polynimial, cosine and sigmoid schedules
--image-size	The (quadratic) size to scale the image to	int	128
--outdir	The directory to save the results to	pathlib.Path	./out/forward/<schedule>
--device	The device to use	torch.Device	"cpu"	If not specified, "cuda" will be used if it is available.
--verbose	Enable verbose logging	bool	False

Authors

• Saif Daknou - Github
• Patrick Lindemann - Github
• Nguyen Pham - Github