Few-Shot Diffusion Models (FSDM)

• preprint: Few-Shot Diffusion Models - https://arxiv.org/abs/2205.15463

Denoising diffusion probabilistic models (DDPM) are powerful hierarchical latent variable models with remarkable sample generation quality and training stability. These properties can be attributed to parameter sharing in the generative hierarchy, as well as a parameter-free diffusion-based inference procedure. In this paper, we present Few-Shot Diffusion Models (FSDM), a framework for few-shot generation leveraging conditional DDPMs. FSDMs are trained to adapt the generative process conditioned on a small set of images from a given class by aggregating image patch information using a set-based Vision Transformer (ViT). At test time, the model is able to generate samples from previously unseen classes conditioned on as few as 5 samples from that class. We empirically show that FSDM can perform few-shot generation and transfer to new datasets taking full advantage of the conditional DDPM.

Set the env

conda create -n fsdm python=3.6

git clone https://github.com/georgosgeorgos/few-shot-diffusion-models

cd few-shot-diffusion-models

pip install -r requirements.txt

Datasets

We train the models on small sets of dimension 2-20. Train/val/test sets use disjoint classes by default.

Binary:

• Omniglot (back_eval) - (1 x 28 x 28) - 964/97/659

RGB:

• CIFAR100 - (3 x 32 x 32) - 60/20/20
• CIFAR100mix - (3 x 32 x 32) - 60/20/20
• MinImageNet - (3 x 32 x 32) - 64/16/20
• CelebA - (3 x 64 x 64) - 4444/635/1270

Training

Train a DDPM on CIFAR100

sh script/run.sh gpu_num ddpm_cifar100 

Train a FSDM model on CIFAR100 dataset with ViT encoder, FiLM conditioning and MEAN aggregation

sh script/run.sh gpu_num vfsddpm_cifar100_vit_film_mean

Train a MODEL on DATASET with ENCODER, CONDITIONING and AGGREGATION

sh script/run.sh gpu_num {dddpm, cddpm, sddpm, addpm, vfsddpm}_{omniglot, cifar100, cifar100mix, minimagenet, cub, celeba}_{vit, unet}_{mean, lag, cls, sum_patch_mean}

Sampling

Sample a FSDM model on CIFAR100 for new classes after 100K iterations 1000 samples

sh script/sample_conditional.sh gpu_num vfsddpm_cifar100_vit_film_mean_outdistro {date} 100000 1000

Metrics

Compute FID, IS, Precision, Recall for a FSDM model on CIFAR100 new classes

Acknoledgments

A lot of code and ideas borrowed from:

• https://github.com/openai/guided-diffusion
• https://github.com/lucidrains/vit-pytorch
• https://github.com/CompVis/latent-diffusion