This is the codebase for the ICML 2023 paper Input Perturbation Reduces Exposure Bias in Diffusion Models.
This repository is heavily based on openai/guided-diffusion, with training modification of input perturbation.
Also, feel free to check out our ICLR 2024 paper Elucidating the Exposure Bias in Diffusion Models which introduces a simple training-free solution to exposure bias. Repository: ADM-ES and EDM-ES
Our proposed Input Perturbation is an extremely simple plug-in method for general diffusion models. The implementation of Input Perturbation is just two lines of code.
For instance, based on guided-diffusion, the only code modifications are in the script guided_diffusion/gaussian_diffusion.py, in line 765-766:
new_noise = noise + gamma * th.randn_like(noise) # gamma=0.1
x_t = self.q_sample(x_start, t, noise=new_noise)NOTE THAT: change the parameter GPUS_PER_NODE = 4 in the script dist_util.py according to your GPU cluster configuration.
the installation is the same with guided-diffusion
git clone https://github.com/forever208/DDPM-IP.git
cd DDPM-IP
conda create -n ADM python=3.8
conda activate ADM
pip install -e .
(note that, pytorch 1.10~1.13 is recommended as our experiments in paper were done with pytorch 1.10 and pytorch 2.0 has not been tested by us in this repo)
# install the missing packages
conda install mpi4py
conda install numpy
pip install Pillow
pip install opencv-python
We have released checkpoints for the main models in the paper.
(The baseline checkpoint of ImageNet-32 and CelebA-64 are missing due to unexpected server file deletion. If you have trained the ADM base models, welcome to share the checkpoints)
Here are the download links for model checkpoints:
- CIFAR10 32x32: ADM-IP.pt, ADM-baseline.pt
- ImageNet 32x32: ADM-IP.pt
- LSUN tower 64x64: ADM-IP.pt, ADM-baseline.pt
- CelebA 64x64: ADM-IP.pt
- FFHQ 128x128: ADM-IP.pt, ADM-baseline.pt
- CIFAR10 32x32: DDIM-IP (NOTE THAT we use DDIM official code to do DDIM-IP training and sampling)
To unconditionally sample from these models, you can use the image_sample.py scripts.
Sampling from DDPM-IP has no difference with sampling from openai/guided-diffusion since DDPM-IP does not change the sampling process.
For example, we sample 50k images using 100 steps from CIFAR10 by:
mpirun python scripts/image_sample.py \
--image_size 32 --timestep_respacing 100 \
--model_path PATH_TO_CHECKPOINT \
--num_channels 128 --num_head_channels 32 --num_res_blocks 3 --attention_resolutions 16,8 \
--resblock_updown True --use_new_attention_order True --learn_sigma True --dropout 0.3 \
--diffusion_steps 1000 --noise_schedule cosine --use_scale_shift_norm True --batch_size 256 --num_samples 50000
sample 50k images using 100 steps from LSUN_tower by:
mpirun -n 1 python scripts/image_sample.py \
--image_size 64 --timestep_respacing 100 \
--model_path PATH_TO_CHECKPOINT \
--use_fp16 True --num_channels 192 --num_head_channels 64 --num_res_blocks 3 \
--attention_resolutions 32,16,8 --resblock_updown True --use_new_attention_order True \
--learn_sigma True --dropout 0.1 --diffusion_steps 1000 --noise_schedule cosine --use_scale_shift_norm True \
--rescale_learned_sigmas True --batch_size 256 --num_samples 50000
sample 50k images using 100 steps from FFHQ128 by:
mpirun -n 1 python scripts/image_sample.py \
--image_size 128 --timestep_respacing 100 \
--model_path PATH_TO_CHECKPOINT \
--use_fp16 True --num_channels 256 --num_head_channels 64 --num_res_blocks 3 \
--attention_resolutions 32,16,8 --resblock_updown True --use_new_attention_order True \
--learn_sigma True --dropout 0.1 --diffusion_steps 1000 --noise_schedule cosine --use_scale_shift_norm True \
--rescale_learned_sigmas True --batch_size 128 --num_samples 50000
This table summarizes our input perturbation results based on ADM baselines. Input perturbation shows tremendous training acceleration and much better FID results.
FID computation details:
- All FIDs are computed using 50K generated samples (unconditional sampling).
- For CIFAR10 and ImageNet 32x32, we use the whole training data as the reference batch,
- For LSUN tower 64x64 and CelebA 64x64, we randomly pick up 50k samples from the training set, forming the reference batch
This table summarizes our input perturbation results based on DDIM baselines.
Please refer to README.md for the data preparation.
Training diffusion models are described in this repository.
Training ADM-IP only requires one more argument --input perturbation 0.1 (set --input perturbation 0.0 for the baseline).
NOTE THAT: if you have problems with slurm multi-node training, try the following setting. Let's say training by 16 GPUs on 2 nodes:
#SBATCH --nodes=2
#SBATCH --ntasks-per-node=8
#SBATCH --cpus-per-task=6
#SBATCH --gres=gpu:8 # 8 gpus for each node
instead of specifying mpiexec -n 16, you run by mpirun python script/image_train.py. (more discussion can be found here)
We share the complete arguments of training ADM-IP in the four datasets:
CIFAR10
mpiexec -n 2 python scripts/image_train.py --input_pertub 0.15 \
--data_dir PATH_TO_DATASET \
--image_size 32 --use_fp16 True --num_channels 128 --num_head_channels 32 --num_res_blocks 3 \
--attention_resolutions 16,8 --resblock_updown True --use_new_attention_order True \
--learn_sigma True --dropout 0.3 --diffusion_steps 1000 --noise_schedule cosine --use_scale_shift_norm True \
--rescale_learned_sigmas True --schedule_sampler loss-second-moment --lr 1e-4 --batch_size 64
ImageNet 32x32 (you can also choose dropout=0.1)
mpiexec -n 4 python scripts/image_train.py --input_pertub 0.1 \
--data_dir PATH_TO_DATASET \
--image_size 32 --use_fp16 True --num_channels 128 --num_head_channels 32 --num_res_blocks 3 \
--attention_resolutions 16,8 --resblock_updown True --use_new_attention_order True \
--learn_sigma True --dropout 0.3 --diffusion_steps 1000 --noise_schedule cosine \
--rescale_learned_sigmas True --schedule_sampler loss-second-moment --lr 1e-4 --batch_size 128
LSUN tower 64x64
mpiexec -n 16 python scripts/image_train.py --input_pertub 0.1 \
--data_dir PATH_TO_DATASET \
--image_size 64 --use_fp16 True --num_channels 192 --num_head_channels 64 --num_res_blocks 3 \
--attention_resolutions 32,16,8 --resblock_updown True --use_new_attention_order True \
--learn_sigma True --dropout 0.1 --diffusion_steps 1000 --noise_schedule cosine --use_scale_shift_norm True \
--rescale_learned_sigmas True --schedule_sampler loss-second-moment --lr 1e-4 --batch_size 16
CelebA 64x64
mpiexec -n 16 python scripts/image_train.py --input_pertub 0.1 \
--data_dir PATH_TO_DATASET \
--image_size 64 --use_fp16 True --num_channels 192 --num_head_channels 64 --num_res_blocks 3 \
--attention_resolutions 32,16,8 --resblock_updown True --use_new_attention_order True \
--learn_sigma True --dropout 0.1 --diffusion_steps 1000 --noise_schedule cosine --use_scale_shift_norm True \
--rescale_learned_sigmas True --schedule_sampler loss-second-moment --lr 1e-4 --batch_size 16
FFHQ 128x128
mpirun -n 16 python scripts/image_train.py --input_pertub 0.1 \
--data_dir PATH_TO_DATASET \
--image_size 128 --use_fp16 True --num_channels 256 --num_head_channels 64 --num_res_blocks 3 \
--attention_resolutions 32,16,8 --resblock_updown True --use_new_attention_order True \
--learn_sigma True --dropout 0.1 --diffusion_steps 1000 --noise_schedule cosine --use_scale_shift_norm True \
--rescale_learned_sigmas True --schedule_sampler loss-second-moment --lr 1e-4 --batch_size 8
If you find our work useful, please feel free to cite by
@inproceedings{ning2023input,
title={Input Perturbation Reduces Exposure Bias in Diffusion Models},
author={Ning, Mang and Sangineto, Enver and Porrello, Angelo and Calderara, Simone and Cucchiara, Rita},
booktitle={International Conference on Machine Learning},
pages={26245--26265},
year={2023},
organization={PMLR}
}
@inproceedings{ningelucidating,
title={Elucidating the Exposure Bias in Diffusion Models},
author={Ning, Mang and Li, Mingxiao and Su, Jianlin and Salah, Albert Ali and Ertugrul, Itir Onal},
booktitle={The Twelfth International Conference on Learning Representations}
}