Awesome

DREAM-OOD

This is the source code accompanying the paper Dream the Impossible: Outlier Imagination with Diffusion Models by Xuefeng Du, Yiyou Sun, Xiaojin Zhu, and Yixuan Li

The codebase is heavily based on Stable Diffusion.

TODO

Please find the link below for both the outlier and inlier images generated for ImageNet and Cifar100 datasets as follows

(Sorry for the delay in announcing the generated images because I did not find a free drive with a large storage. The link will expire in 30 days due to limitations at my university, thanks! If you have a suggested free online drive with large storage limits, please don't hesitate to forward it to me)

If you find the link expiring, please also email me!!

Check out our

latent-based outlier synthesis papers in ICLR'22 VOS and ICLR'23 NPOS
unknown synthesis for object detection in video datasets CVPR'22 work STUD if you are interested!

Requirements

A suitable conda environment named dreamood can be created and activated with:

conda env create -f environment.yaml
conda activate dreamood

Please also install Xformers.

Dataset Preparation

ImageNet-100

Download the full ImageNet dataset from the official website here.
Preprocess the dataset to get ImageNet-100 by running:

python scripts/generate_in100.py --source_folder xxx --target_folder xxx

where "--source_folder" is the address of the full ImageNet dataset and "--target_folder" specifies the address of the dataset you want to store.

CIFAR-100

The dataloader will download it automatically when first running the programs.

OOD datasets

The OOD datasets with ImageNet-100 as in-distribution are 4 OOD datasets from iNaturalist, SUN, Places, and Textures, which contain the de-duplicated concepts overlapped with ImageNet.
The OOD datasets with CIFAR-100 as in-distribution are 5 OOD datasets, i.e., SVHN, PLACES365, LSUN, ISUN, TEXTURES.
Please refer to Part 1 and 2 of the codebase here.

Datasets for evaluating model generalization

Please download IMAGENET-A and IMAGENET-V2 and process the dataset by running (you need change the address of the datasets on your own):

python scripts/process_imagenetv2_and_a.py

Training

1. Learning the text-conditioned latent space

Please execute the following in the command shell on ImageNet-100:

python scripts/pretrain_in100.py

Please execute the following in the command shell on CIFAR-100:

python scripts/pretrain_cifar100.py

After training, it will generate ID feature embeddings for outlier/inlier embedding sampling.

Pretrained models for ImageNet-100 and Cifar-100.

2. Generate the inlier/outlier embeddings

Please execute the following in the command shell on ImageNet-100:

python scripts/get_embed_in100.py

Please execute the following in the command shell on CIFAR-100:

python scripts/get_embed_cifar100.py

After this step, you will see the generated inlier/outlier embedding in the root directory.

Pretrained embeddings: inliers for ImageNet-100, outlier for ImageNet-100 and outliers for Cifar-100.

3. Synthesizing outliers in the pixel space

First, please download the Stable Diffusion 1.4 model here.

Please execute the following in the command shell on different datasets:

python scripts/dream_ood.py --plms \
--n_iter 50 --n_samples 3 \
--outdir /nobackup-fast/txt2img-samples-in100-demo/ \
--loaded_embedding /nobackup-slow/dataset/my_xfdu/diffusion/outlier_npos_embed.npy\
--ckpt /nobackup-slow/dataset/my_xfdu/diffusion/sd-v1-4.ckpt \
--id_data in100 \
--skip_grid

"--loaded_embedding" means the address of the saved inlier/outlier embeddings obtained by the previous step.
"--outdir" denotes the address you want to save the generated outlier images.
"--n_iter"/"--n_samples" control the number of images you generate in each step and the number of steps for generation.
"--id_data" can be chozen between in100 and cifar100.
For generating 100K images, you can specify the n_iter and n_samples such as n_iter=25,000 and n_samples=4. Consider generating on multiplt GPUs for speed up.
Generated images: outlier image for IN100, inlier image for IN100 and outlier image for cifar100.

4. Training with the generated outliers in the pixel space

Please execute the following in the command shell for OOD detection on ImageNet-100:

python scripts/train_ood_det_in100.py --my_info xxx --load xxx

"--my_info" denotes the name of the folder that contains the generated datasets in the previous step.

Note that in order to save time, we use a pretrained model for initialization, which is trained using the cross-entropy loss.

Please execute the following in the command shell for OOD detection on Cifar-100:

python scripts/train_ood_det_cifar100.py --my_info xxx

Here the model is trained from scratch.

Pretrained models: IN100, CIFAR-100.

5. Training with the generated inliers in the pixel space

Please execute the following in the command shell for generalization on ImageNet-100:

python scripts/train_gene_in100.py

Pretrained models: IN100.

Test-time OOD detection

Please execute the following in the command shell with ImageNet-100 as in-distribution:

python scripts/test_ood_in100.py --load xxx

where "--load" specifies the address of the saved models.

Please execute the following in the command shell with Cifar-100 as in-distribution:

python scripts/test_ood_cifar100.py --load xxx

Test-time ID generalization

Please execute the following in the command shell with ImageNet-100:

python scripts/test_ood_in100_robustness.py --load xxx

Citation

If you found any part of this code is useful in your research, please consider citing our paper:

  @inproceedings{du2023dream,
      title={Dream the Impossible: Outlier Imagination with Diffusion Models}, 
      author={Xuefeng Du and Yiyou Sun and Xiaojin Zhu and Yixuan Li },
      booktitle={Advances in Neural Information Processing Systems},
      year = {2023}
}