icaruswizard / deep-generative-models

Collection of PyTorch implementation for several deep generative models.

Deep Generative Models

A collection of my PyTorch implementation of several deep generative models.
This repertory is in progressing, feel free to raise an issue if you find any bug.

Requirement

• PyTorch >= 1.0 (This code was develop on 1.3.1, but it should also work fine on other version)
• tensorboard (tb-nightly)
• numpy, scipy (ndarry support)
• matplotlib, moviepy (visualizing result)
• tqdm (progress bar)

Setup

Recommend to setup with Anaconda

git clone https://github.com/IcarusWizard/Deep-Generative-Models
cd Deep-Generative-Models
pip install -e .

Models

• Auto-regressive
  • PixelRNN
  • PixelCNN
  • PixelSNAIL
• FLOW
  • NICE
  • RealNVP
  • Glow
• VAE
  • Vanilla VAE
  • VQ-VAE
• GAN
  • Vanilla GAN
  • DCGAN
  • InfoGAN
  • WGAN
  • WGAN-GP
  • SNGAN
  • SAGAN

Datasets

• MNIST
• SVHN
• Cifar10
• CelebA

If you want to add new dataset, you need to define a creator function which returns three torch.utils.data.Dataset for training, validation, testing, and a dict holds the configuration of the dataset (c, h, w). Then you can add your custom loader through:

import degmo
def custom_creater():
  ......
  return training_set, validation_set, testing_set, config

degmo.add_dataset('name', custom_creater)

Code

Structure

logs/          # default tensorboard log folder
checkpoints/   # default checkpoints folder
degmo/         # main folder
  data/          # dataset functions
  config/        # default configurations
  utils.py       # shared utility functions
  modules.py     # shared utility modules
  <method>/
    trainer.py        # training procedure
    run_utils.py      # runtime utility functions
    utils.py          # method's utility functions
    modules.py        # method's utility modules
    <model.py>        # model class
    ......

Train

Run python -m degmo.train_<method> --dataset <dataset> --model <model> to train in default configuration. You can run python -m degmo.check_default_config <method> to find the default configuration we provide, or just look inside degmo/config folder.
If you want to tune some parameters for yourself, pass --custom to the training script, run python -m degmo.train_<method> -h to see all the parameters that you can tune.
Note: All the default configurations are tested on a single RTX 2080Ti GPU with 11G memory, if you cannot run some default configurations (i.e. Glow), please consider reduce the batch size or features in config file or with a custom mode.

During and after training, you can use tensorboard --logdir=logs to monitor progress.

Test

Run python -m degmo.test_<model> -h for help.