Skip to content

Implementation of various DDPM papers to understand how they work

License

Notifications You must be signed in to change notification settings

Michedev/DDPMs-Pytorch

Repository files navigation

PyPI - Downloads PyPI PyPI - Python Version

DDPM Pytorch

Pytorch implementation of "Improved Denoising Diffusion Probabilistic Models", "Denoising Diffusion Probabilistic Models" and "Classifier-free Diffusion Guidance"

How to use

There are two ways to use this repository:

  1. Install pip package containing the pytorch lightning model, which includes also the training step

    pip install ddpm
    
  2. Clone the repository to have the full control of the training

     git clone https://github.com/Michedev/DDPMs-Pytorch
    

How to train

  1. Install the project environment via hatch (pip install hatch). There are two environments: default has torch with cuda support, cpu without it.

    hatch env create
    

or hatch env create cpu

  1. Train the model

    hatch run train 
    

    or for the cpu environment

    hatch run cpu:train
    

    Note that this is valid for any hatch run [env:]{command} command

    By default, the version of trained DDPM is from "Improved Denoising Diffusion Probabilistic Models" paper on MNIST dataset. You can switch to the original DDPM by disabling the variational lower bound with the following command:

    hatch run train model.vlb=False
    

    You can also train the DDPM with the Classifier-free Diffusion Guidance by changing the model:

    hatch run train model=unet_class_conditioned
    

    or via the shortcut

    hatch run train-class-conditioned
    

    Finally, under saved_models/{train-datetime} you can find the trained model, the tensorboard logs, the training config

How to generate

  1. Train a model (See previous section)

  2. Generate a new batch of images

    hatch run generate -r RUN
    

    The other options are: [--seed SEED] [--device DEVICE] [--batch-size BATCH_SIZE] [-w W] [--scheduler {linear,cosine,tan}] [-T T]

Configure the training

Under config there are several yaml files containing the training parameters such as model class and paramters, noise steps, scheduler and so on. Note that the hyperparameters in such files are taken from the papers "Improved Denoising Diffusion Probabilistic Models" and "Denoising Diffusion Probabilistic Models". Down below the explaination of the config file for train the model:

defaults:
  - model: unet_paper  # take the model config from model/unet_paper.yaml
  - scheduler: cosine  # use the cosine scheduler from scheduler/cosine.yaml
  - dataset: mnist
  - optional model_dataset: ${model}-${dataset}  # set particular hyper parameters for specific couples (model, dataset)
  - optional model_scheduler: ${model}-${scheduler} # set particular hyper parameters for specific couples (model, scheduler)

batch_size: 128 # train batch size
noise_steps: 4_000  # noising steps; the T in "Improved Denoising Diffusion Probabilistic Models" and "Denoising Diffusion Probabilistic Models"
accelerator: null  # training hardware; for more details see pytorch lightning
devices: null  # training devices to use; for more details see pytorch lightning
gradient_clip_val: 0.0  # 0.0 means gradient clip disabled
gradient_clip_algorithm: norm  # gradient clip has two values: 'norm' or 'value
ema: true  # use Exponential Moving Average implemented in ema.py
ema_decay: 0.99  # decay factor of EMA

hydra:
  run:
    dir: saved_models/${now:%Y_%m_%d_%H_%M_%S}

Project structure

  .
  ├── callbacks  # Pytorch Lightning callbacks for training
  │   ├── ema.py  # exponential moving average callback
  ├── config  # config files for training for hydra
  │   ├── dataset  # dataset config files
  │   ├── model  # model config files
  │   ├── model_dataset  # specific (model, dataset) config
  │   ├── model_scheduler  # specific (model, scheduler) config
  │   ├── scheduler  # scheduler config files
  │   └── train.yaml  # training config file
  ├── generate.py  # script for generating images
  ├── model  # model files
  │   ├── classifier_free_ddpm.py  # Classifier-free Diffusion Guidance
  │   ├── ddpm.py  # Denoising Diffusion Probabilistic Models
  │   ├── distributions.py  # distributions functions for diffusion
  │   ├── unet_class.py  # UNet model for Classifier-free Diffusion Guidance
  │   └── unet.py  # UNet model for Denoising Diffusion Probabilistic Models
  ├── pyproject.toml  # setuptool file to publish model/ to pypi and to manage the envs
  ├── readme.md   # this file
  ├── readme_pip.md  # readme for pypi
  ├── train.py  # script for training
  ├── utils  # utility functions
  └── variance_scheduler  # variance scheduler files
      ├── cosine.py  # cosine variance scheduler
      └── linear.py  # linear variance scheduler

Add custom dataset

To add a custom dataset, you need to create a new class that inherits from torch.utils.data.Dataset and implement the len and getitem methods. Then, you need to add the config file to the config/dataset folder with a similar structure of mnist.yaml

width: 28  # meta info about the dataset
height: 28
channels: 1   # number of image channels
num_classes: 10  # number of classes
files_location: ~/.cache/torchvision_dataset  # location where to store the dataset, in case to be downloaded
train:  #dataset.train is instantiated with this config
  _target_: torchvision.datasets.MNIST  # Dataset class. Following arguments are passed to the dataset class constructor
  root: ${dataset.files_location}
  train: true
  download: true
  transform:
    _target_: torchvision.transforms.ToTensor
val:  #dataset.val is instantiated with this config
  _target_: torchvision.datasets.MNIST # Same dataset of train, but the validation split
  root: ${dataset.files_location}
  train: false
  download: true
  transform:
    _target_: torchvision.transforms.ToTensor

Examples of custom training

Disable the variational lower bound, use Linear scheduler, use 1000 noise steps, train in GPU

hatch run train scheduler=linear accelerator='gpu' model.vlb=False noise_steps=1000

Classifier-free Guidance

Use the labels for Diffusion Guidance, as in "Classifier-free Diffusion Guidance" with the following command

hatch run train model=unet_class_conditioned noise_steps=1000

Add your scheduler

  1. Add a new class (preferabily under variance_scheduler/) which subclasses Scheduler class or just copy the same methods syntax of Scheduler
  2. Define a new config under config/scheduler with the name my-scheduler.yaml containing the following fields
 _target_: {your scheduler import path} (e.g. variance_scheduler.Linear)
... // your scheduler additional parameters

Finally train with the following command

hatch run train scheduler=my-scheduler

Add your dataset

  1. Add a new class which subclasses torch.utils.data.Dataset

  2. Define a new config under config/dataset with the name my-dataset.yaml containing the following fields

width: ???
height: ???
channels: ???
train:
  _target_: {your dataset import path} (e.g. torchvision.datasets.MNIST)
  // your dataset additional parameters
val:
  _target_: {your dataset import path} (e.g. torchvision.datasets.MNIST)
  // your dataset additional parameters

Finally train with the following command

hatch run train dataset=my-dataset