siyu-li-dream-follower / dpdm Goto Github PK

DPDM is built using PyTorch 1.11.0 and CUDA 11.3. Please use the following command to install the requirements:

pip install -r requirements.txt 

We provide pre-trained checkpoints for all models presented in the paper here. You can sample from the checkpoint using the following command:

python main.py --mode eval --workdir <new_directory> --config <config_file> -- model.ckpt=<checkpoint_path>

By default, the above command generates 16 samples using DDIM with 50 steps on a single GPU. You can modify any entry of the config file directly through the command line, e.g., you may append sampler.num_steps=20 to the above command to change the number of DDIM steps from 50 (default value) to 20.

For class-conditional sampling, you can set sampler.labels to a single integer between 0 and num_classes - 1 (inclusive) to generate all samples from the same class. By default, the configs sets sampler.labels=num_classes which performs random sampling.

All class-conditional models are trained for classifier-free guidance by dropping out the label 10% of the time. By default, the guidance strength is set to 0 but you may change it by setting sampler.guid_scale.

To reproduce the results in the paper, you may adjust the sampler settings accordingly. For example, to obtain the best FID value for MNIST, eps=1.0, the sampler setting is (Table 13 & Table 14):

python main.py --mode eval --workdir <new_directory> --config configs/mnist_28/sample_eps_1.0.yaml -- model.ckpt=<checkpoint_path> sampler.type=edm sampler.s_churn=100. sampler.s_min=0.05 sampler.s_max=50. sampler.num_steps=1000

First, create the following two folders:

mkdir -p data/raw/
mkdir -p data/processed/

Afterwards, run the following commands to download and prepare the data used for training. MNIST and FashionMNIST are downloaded and processed automatically.

wget -P data/raw/ https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
python dataset_tool.py --source data/raw/cifar-10-python.tar.gz --dest data/processed/cifar10.zip

python dataset_tool.py --source==data/raw/imagenet/ILSVRC/Data/CLS-LOC/train --dest=data/processed/imagenet.zip --resolution=32x32 --transform=center-crop

Before training, you should compute FID statistics of the data:

python precompute_data_mnist_fid_statistics.py
python precompute_data_mnist_fid_statistics.py --test
python precompute_data_mnist_fid_statistics.py --is_fmnist
python precompute_data_mnist_fid_statistics.py --is_fmnist --test
python compute_fid_statistics.py --path data/processed/cifar10.zip --file cifar10.npz 
python compute_fid_statistics.py --path data/processed/imagenet.zip --file imagenet.npz
python compute_fid_statistics.py --path data/processed/celeba.zip --file celeba.npz

We provide configurations to reproduce our models here. You may use the following command for training:

python main.py --mode train --workdir <new_directory> --config <dataset>

Our models are trained on eight NVIDIA A100 (80Gb) GPUs; you can set the number of GPUs per node via setup.n_gpus_per_node. We use large batch sizes which are split into several iterations; to reduce the required GPU memory, you may increase the flag train.n_splits (by a multiple of 2).

If you find the provided code or checkpoints useful for your research, please consider citing our paper:

@article{dockhorn2022differentially,
    title={{Differentially Private Diffusion Models}},
    author={Dockhorn, Tim and Cao, Tianshi and Vahdat, Arash and Kreis, Karsten},
    journal={arXiv:2210.09929},
    year={2022}
}

Copyright © 2023, NVIDIA Corporation. All rights reserved.

The code of this work is made available under the NVIDIA Source Code License. Please see our main LICENSE file.

Pre-trained checkpooints are licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

For any code dependencies related to StyleGAN3 (stylegan3/, torch_utils/, and dnnlib/), the license is the Nvidia Source Code License by NVIDIA Corporation, see StyleGAN3 LICENSE.

Our DPDMs are based on the DDPM++ architecture which is realeased under Apache License 2.0.