mareikethies/ddpm-ood

Official PyTorch code for "Out-of-distribution detection with denoising diffusion models"

Denoising diffusion models for out-of-distribution detection

Perform reconstruction-based out-of-distribution detection with DDPMs.

Intro

This codebase contains the code to perform unsupervised out-of-distribution detection with diffusion models. It supports the use of DDPMs as well as Latent Diffusion Models (LDM) for dealing with higher dimensional 2D or 3D data. It is based on work published in [1] and [2].

[1] Denoising diffusion models for out-of-distribution detection, CVPR VAND Workshop 2023

[2] Unsupervised 3D out-of-distribution detection with latent diffusion models, MICCAI 2023

Setup

Install

Create a fresh virtualenv (this codebase was developed and tested with Python 3.8) and then install the required packages:

pip install -r requirements.txt

You can also build the docker image

cd docker/
bash create_docker_image.sh

Setup paths

Select where you want your data and model outputs stored.

data_root=/root/for/downloaded/dataset
output_root=/root/for/saved/models

Run with DDPM

We'll use the example of FashionMNIST as an in-distribution dataset and [SVHN,CIFAR10, CelebA] as out-of-distribution datasets.

Download and process datasets

python src/data/get_computer_vision_datasets.py --data_root=${data_root}

N.B. If the error "The daily quota of the file img_align_celeba.zip is exceeded and it can't be downloaded" is thrown, you need to download these files manually from the GDrive and place them in ${data_root}/CelebA/raw/, see here. You can then run

python get_datasets.py --data_root=${data_root} --download_celeba=False

To use your own data, you just need to provide separate csvs containing paths for the train/val/test splits.

Train models

Examples here use FashionMNIST as the in-distribution dataset. Commands for other datasets are given in README_additional.md.

python train_ddpm.py \
--output_dir=${output_root} \
--model_name=fashionmnist \
--training_ids=${data_root}/data_splits/FashionMNIST_train.csv \
--validation_ids=${data_root}/data_splits/FashionMNIST_val.csv \
--is_grayscale=1 \
--n_epochs=300 \
--beta_schedule=scaled_linear \
--beta_start=0.0015 \
--beta_end=0.0195

You can track experiments in tensorboard

tensorboard --logdir=${output_root}

The code is DistributedDataParallel (DDP) compatible. To train on e.g. 2 GPUs:

torchrun --nproc_per_node=2 --nnodes=1 --node_rank=0 \
train_ddpm.py \
--output_dir=${output_root} \
--model_name=fashionmnist \
--training_ids=${data_root}/data_splits/FashionMNIST_train.csv \
--validation_ids=${data_root}/data_splits/FashionMNIST_val.csv \
--is_grayscale=1 \
--n_epochs=300 \
--beta_schedule=scaled_linear \
--beta_start=0.0015 \
--beta_end=0.0195

Reconstruct data

python reconstruct.py \
--output_dir=${output_root} \
--model_name=fashionmnist \
--validation_ids=${data_root}/data_splits/FashionMNIST_val.csv \
--in_ids=${data_root}/data_splits/FashionMNIST_test.csv \
--out_ids=${data_root}/data_splits/MNIST_test.csv,${data_root}/data_splits/FashionMNIST_vflip_test.csv,${data_root}/data_splits/FashionMNIST_hflip_test.csv \
--is_grayscale=1 \
--beta_schedule=scaled_linear \
--beta_start=0.0015 \
--beta_end=0.0195 \
--num_inference_steps=100 \
--inference_skip_factor=4 \
--run_val=1 \
--run_in=1 \
--run_out=1

The arg inference_skip_factor controls the amount of t starting points that are skipped during reconstruction. This table shows the relationship between values of inference_skip_factor and the number of reconstructions, as needed to reproduce results in Supplementary Table 4 (for max_t=1000).

inference_skip_factor:	1	2	3	4	5	8	16	32	64
num_reconstructions:	100	50	34	25	20	13	7	4	2

N.B. For a quicker run, you can choose to only reconstruct a subset of the validation set with e.g. --first_n_val=1000 or a subset of the in/out datasets with --first_n=1000

Classify samples as OOD

python ood_detection.py \
--output_dir=${output_root} \
--model_name=fashionmnist

Run with LDM

We'll use the 3D Medical Decathlon Dataset here. In this example we'll use the BraTS dataset as the in-distribution dataset, and the other 9 datasets as out-of-distribution datasets.

Download and process datasets

python src/data/get_decathlon_datasets.py --data_root=${data_root}/Decathlon

Train VQVAE

python train_vqvae.py  \
--output_dir=${output_root} \
--model_name=vqvae_decathlon \
--training_ids=${data_root}/data_splits/Task01_BrainTumour_train.csv \
--validation_ids=${data_root}/data_splits/Task01_BrainTumour_val.csv  \
--is_grayscale=1 \
--n_epochs=300 \
--batch_size=8  \
--eval_freq=10 \
--cache_data=0  \
--vqvae_downsample_parameters=[[2,4,1,1],[2,4,1,1],[2,4,1,1],[2,4,1,1]] \
--vqvae_upsample_parameters=[[2,4,1,1,0],[2,4,1,1,0],[2,4,1,1,0],[2,4,1,1,0]] \
--vqvae_num_channels=[256,256,256,256] \
--vqvae_num_res_channels=[256,256,256,256] \
--vqvae_embedding_dim=128 \
--vqvae_num_embeddings=2048 \
--vqvae_decay=0.9  \
--vqvae_learning_rate=3e-5 \
--spatial_dimension=3 \
--image_roi=[160,160,128] \
--image_size=128

The code is DistributedDataParallel (DDP) compatible. To train on e.g. 2 GPUs run with torchrun --nproc_per_node=2 --nnodes=1 --node_rank=0 train_vqvae.py

Train LDM

python train_ddpm.py \
  --output_dir=${output_root} \
  --model_name=ddpm_decathlon \
  --vqvae_checkpoint=${output_root}/vqvae_decathlon/checkpoint.pth \
  --training_ids=${data_root}/data_splits/Task01_BrainTumour_train.csv \
  --validation_ids=${data_root}/data_splits/Task01_BrainTumour_val.csv  \
  --is_grayscale=1 \
  --n_epochs=12000 \
  --batch_size=6 \
  --eval_freq=25 \
  --checkpoint_every=1000 \
  --cache_data=0  \
  --prediction_type=epsilon \
  --model_type=small \
  --beta_schedule=scaled_linear_beta \
  --beta_start=0.0015 \
  --beta_end=0.0195 \
  --b_scale=1.0 \
  --spatial_dimension=3 \
  --image_roi=[160,160,128] \
  --image_size=128

Reconstruct data

python reconstruct.py \
  --output_dir=${output_root} \
  --model_name=ddpm_decathlon \
  --vqvae_checkpoint=${output_root}/decathlon-vqvae-4layer/checkpoint.pth \
  --validation_ids=${data_root}/data_splits/Task01_BrainTumour_val.csv  \
  --in_ids=${data_root}/data_splits/Task01_BrainTumour_test.csv \
  --out_ids=${data_root}/data_splits/Task02_Heart_test.csv,${data_root}/data_splits/Task03_Liver_test.csv,${data_root}/data_splits/Task04_Hippocampus_test.csv,${data_root}/data_splits/Task05_Prostate_test.csv,${data_root}/data_splits/Task06_Lung_test.csv,${data_root}/data_splits/Task07_Pancreas_test.csv,${data_root}/data_splits/Task08_HepaticVessel_test.csv,${data_root}/data_splits/Task09_Spleen_test.csv\
  --is_grayscale=1 \
  --batch_size=32 \
  --cache_data=0 \
  --prediction_type=epsilon \
  --beta_schedule=scaled_linear_beta \
  --beta_start=0.0015 \
  --beta_end=0.0195 \
  --b_scale=1.0 \
  --spatial_dimension=3 \
  --image_roi=[160,160,128] \
  --image_size=128 \
  --num_inference_steps=100 \
  --inference_skip_factor=2 \
  --run_val=1 \
  --run_in=1 \
  --run_out=1

Classify samples as OOD

python ood_detection.py \
--output_dir=${output_root} \
--model_name=ddpm_decathlon

Acknowledgements

Built with MONAI Generative and MONAI.

Citations

If you use this codebase, please cite

@InProceedings{Graham_2023_CVPR,
    author    = {Graham, Mark S. and Pinaya, Walter H.L. and Tudosiu, Petru-Daniel and Nachev, Parashkev and Ourselin, Sebastien and Cardoso, Jorge},
    title     = {Denoising Diffusion Models for Out-of-Distribution Detection},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2023},
    pages     = {2947-2956}
}
@inproceedings{graham2023unsupervised,
  title={Unsupervised 3D out-of-distribution detection with latent diffusion models},
  author={Graham, Mark S and Pinaya, Walter Hugo Lopez and Wright, Paul and Tudosiu, Petru-Daniel and Mah, Yee H and Teo, James T and J{\"a}ger, H Rolf and Werring, David and Nachev, Parashkev and Ourselin, Sebastien and others},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={446--456},
  year={2023},
  organization={Springer}
}
}