/pytorch_diffusion

PyTorch reimplementation of Diffusion Models

Primary LanguagePythonMIT LicenseMIT

PyTorch pretrained Diffusion Models

A PyTorch reimplementation of Denoising Diffusion Probabilistic Models with checkpoints converted from the author's TensorFlow implementation.

Quickstart

Running

pip install -e git+https://github.com/pesser/pytorch_diffusion.git#egg=pytorch_diffusion
pytorch_diffusion_demo

will start a Streamlit demo. It is recommended to run the demo with a GPU available.

demo

Usage

Diffusion models with pretrained weights for cifar10, lsun-bedroom, lsun_cat or lsun_church can be loaded as follows:

from pytorch_diffusion import Diffusion

diffusion = Diffusion.from_pretrained("lsun_church")
samples = diffusion.denoise(4)
diffusion.save(samples, "lsun_church_sample_{:02}.png")

Prefix the name with ema_ to load the averaged weights that produce better results. The U-Net model used for denoising is available via diffusion.model and can also be instantiated on its own:

from pytorch_diffusion import Model

model = Model(resolution=32,
              in_channels=3,
              out_ch=3,
              ch=128,
              ch_mult=(1,2,2,2),
              num_res_blocks=2,
              attn_resolutions=(16,),
              dropout=0.1)

This configuration example corresponds to the model used on CIFAR-10.

Producing samples

If you installed directly from github, you can find the cloned repository in <venv path>/src/pytorch_diffusion for virtual environments, and <cwd>/src/pytorch_diffusion for global installs. There, you can run

python pytorch_diffusion/diffusion.py <name> <bs> <nb>

where <name> is one of cifar10, lsun-bedroom, lsun_cat, lsun_church, or one of these names prefixed with ema_, <bs> is the batch size and <nb> the number of batches. This will produce samples from the PyTorch models and save them to results/<name>/.

Results

Evaluating 50k samples with torch-fidelity gives

Dataset EMA Framework Model FID
CIFAR10 Train no PyTorch cifar10 12.13775
TensorFlow tf_cifar10 12.30003
yes PyTorch ema_cifar10 3.21213
TensorFlow tf_ema_cifar10 3.245872
CIFAR10 Validation no PyTorch cifar10 14.30163
TensorFlow tf_cifar10 14.44705
yes PyTorch ema_cifar10 5.274105
TensorFlow tf_ema_cifar10 5.325035

To reproduce, generate 50k samples from the converted PyTorch models provided in this repo with

`python pytorch_diffusion/diffusion.py <Model> 500 100`

and with

python -c "import convert as m; m.sample_tf(500, 100, which=['cifar10', 'ema_cifar10'])"

for the original TensorFlow models.

Running conversions

The converted pytorch checkpoints are provided for download. If you want to convert them on your own, you can follow the steps described here.

Setup

This section assumes your working directory is the root of this repository. Download the pretrained TensorFlow checkpoints. It should follow the original structure,

diffusion_models_release/
  diffusion_cifar10_model/
    model.ckpt-790000.data-00000-of-00001
    model.ckpt-790000.index
    model.ckpt-790000.meta
  diffusion_lsun_bedroom_model/
    ...
  ...

Set the environment variable TFROOT to the directory where you want to store the author's repository, e.g.

export TFROOT=".."

Clone the diffusion repository,

git clone https://github.com/hojonathanho/diffusion.git ${TFROOT}/diffusion

and install their required dependencies (pip install ${TFROOT}/requirements.txt). Then add the following to your PYTHONPATH:

export PYTHONPATH=".:./scripts:${TFROOT}/diffusion:${TFROOT}/diffusion/scripts:${PYTHONPATH}"

Testing operations

To test the pytorch implementations of the required operations against their TensorFlow counterparts under random initialization and random inputs, run

python -c "import convert as m; m.test_ops()"

Converting checkpoints

To load the pretrained TensorFlow models, copy the weights into the pytorch models, check for equality on random inputs and finally save the corresponding pytorch checkpoints, run

python -c "import convert as m; m.transplant_cifar10()"
python -c "import convert as m; m.transplant_cifar10(ema=True)"
python -c "import convert as m; m.transplant_lsun_bedroom()"
python -c "import convert as m; m.transplant_lsun_bedroom(ema=True)"
python -c "import convert as m; m.transplant_lsun_cat()"
python -c "import convert as m; m.transplant_lsun_cat(ema=True)"
python -c "import convert as m; m.transplant_lsun_church()"
python -c "import convert as m; m.transplant_lsun_church(ema=True)"

Pytorch checkpoints will be saved in

diffusion_models_converted/
  diffusion_cifar10_model/
    model-790000.ckpt
  ema_diffusion_cifar10_model/
    model-790000.ckpt
  diffusion_lsun_bedroom_model/
    model-2388000.ckpt
  ema_diffusion_lsun_bedroom_model/
    model-2388000.ckpt
  diffusion_lsun_cat_model/
    model-1761000.ckpt
  ema_diffusion_lsun_cat_model/
    model-1761000.ckpt
  diffusion_lsun_church_model/
    model-4432000.ckpt
  ema_diffusion_lsun_church_model/
    model-4432000.ckpt

Sample TensorFlow models

To produce N samples from each of the pretrained TensorFlow models, run

python -c "import convert as m; m.sample_tf(N)"

Pass a list of model names as keyword argument which to specify which models to sample from. Samples will be saved in results/.