Documentation

pythae

This library implements some of the most common (Variational) Autoencoder models under a unified implementation. In particular, it provides the possibility to perform benchmark experiments and comparisons by training the models with the same autoencoding neural network architecture. The feature make your own autoencoder allows you to train any of these models with your own data and own Encoder and Decoder neural networks. It integrates experiment monitoring tools such wandb, mlflow or comet-ml 🧪 and allows model sharing and loading from the HuggingFace Hub 🤗 in a few lines of code.

News 📢

As of v0.1.0, Pythae now supports distributed training using PyTorch's DDP. You can now train your favorite VAE faster and on larger datasets, still with a few lines of code. See our speed-up benchmark.

Quick access:

• Installation
• Implemented models / Implemented samplers
• Reproducibility statement / Results flavor
• Model training / Data generation / Custom network architectures / Distributed training
• Model sharing with 🤗 Hub / Experiment tracking with wandb / Experiment tracking with mlflow / Experiment tracking with comet_ml
• Tutorials / Documentation
• Contributing 🚀 / Issues 🛠️
• Citing this repository

Installation

To install the latest stable release of this library run the following using pip

$ pip install pythae

To install the latest github version of this library run the following using pip

$ pip install git+https://github.com/clementchadebec/benchmark_VAE.git

or alternatively you can clone the github repo to access to tests, tutorials and scripts.

$ git clone https://github.com/clementchadebec/benchmark_VAE.git

and install the library

$ cd benchmark_VAE
$ pip install -e .

Available Models

Below is the list of the models currently implemented in the library.

Models	Training example	Paper	Official Implementation
Autoencoder (AE)
Variational Autoencoder (VAE)		link
Beta Variational Autoencoder (BetaVAE)		link
VAE with Linear Normalizing Flows (VAE_LinNF)		link
VAE with Inverse Autoregressive Flows (VAE_IAF)		link	link
Disentangled Beta Variational Autoencoder (DisentangledBetaVAE)		link
Disentangling by Factorising (FactorVAE)		link
Beta-TC-VAE (BetaTCVAE)		link	link
Importance Weighted Autoencoder (IWAE)		link	link
Multiply Importance Weighted Autoencoder (MIWAE)		link
Partially Importance Weighted Autoencoder (PIWAE)		link
Combination Importance Weighted Autoencoder (CIWAE)		link
VAE with perceptual metric similarity (MSSSIM_VAE)		link
Wasserstein Autoencoder (WAE)		link	link
Info Variational Autoencoder (INFOVAE_MMD)		link
VAMP Autoencoder (VAMP)		link	link
Hyperspherical VAE (SVAE)		link	link
Poincaré Disk VAE (PoincareVAE)		link	link
Adversarial Autoencoder (Adversarial_AE)		link
Variational Autoencoder GAN (VAEGAN) 🥗		link	link
Vector Quantized VAE (VQVAE)		link	link
Hamiltonian VAE (HVAE)		link	link
Regularized AE with L2 decoder param (RAE_L2)		link	link
Regularized AE with gradient penalty (RAE_GP)		link	link
Riemannian Hamiltonian VAE (RHVAE)		link	link

See reconstruction and generation results for all aforementionned models

Available Samplers

Below is the list of the models currently implemented in the library.

Samplers	Models	Paper	Official Implementation
Normal prior (NormalSampler)	all models	link
Gaussian mixture (GaussianMixtureSampler)	all models	link	link
Two stage VAE sampler (TwoStageVAESampler)	all VAE based models	link	link
Unit sphere uniform sampler (HypersphereUniformSampler)	SVAE	link	link
Poincaré Disk sampler (PoincareDiskSampler)	PoincareVAE	link	link
VAMP prior sampler (VAMPSampler)	VAMP	link	link
Manifold sampler (RHVAESampler)	RHVAE	link	link
Masked Autoregressive Flow Sampler (MAFSampler)	all models	link	link
Inverse Autoregressive Flow Sampler (IAFSampler)	all models	link	link
PixelCNN (PixelCNNSampler)	VQVAE	link

Reproducibility

We validate the implementations by reproducing some results presented in the original publications when the official code has been released or when enough details about the experimental section of the papers were available. See reproducibility for more details.

Launching a model training

To launch a model training, you only need to call a TrainingPipeline instance.

>>> from pythae.pipelines import TrainingPipeline
>>> from pythae.models import VAE, VAEConfig
>>> from pythae.trainers import BaseTrainerConfig

>>> # Set up the training configuration
>>> my_training_config = BaseTrainerConfig(
...	output_dir='my_model',
...	num_epochs=50,
...	learning_rate=1e-3,
...	per_device_train_batch_size=200,
...	per_device_eval_batch_size=200,
...	train_dataloader_num_workers=2,
...	eval_dataloader_num_workers=2,
...	steps_saving=20,
...	optimizer_cls="AdamW",
...	optimizer_params={"weight_decay": 0.05, "betas": (0.91, 0.995)},
...	scheduler_cls="ReduceLROnPlateau",
...	scheduler_params={"patience": 5, "factor": 0.5}
... )
>>> # Set up the model configuration 
>>> my_vae_config = model_config = VAEConfig(
...	input_dim=(1, 28, 28),
...	latent_dim=10
... )
>>> # Build the model
>>> my_vae_model = VAE(
...	model_config=my_vae_config
... )
>>> # Build the Pipeline
>>> pipeline = TrainingPipeline(
... 	training_config=my_training_config,
... 	model=my_vae_model
... )
>>> # Launch the Pipeline
>>> pipeline(
...	train_data=your_train_data, # must be torch.Tensor, np.array or torch datasets
...	eval_data=your_eval_data # must be torch.Tensor, np.array or torch datasets
... )

At the end of training, the best model weights, model configuration and training configuration are stored in a final_model folder available in my_model/MODEL_NAME_training_YYYY-MM-DD_hh-mm-ss (with my_model being the output_dir argument of the BaseTrainerConfig). If you further set the steps_saving argument to a certain value, folders named checkpoint_epoch_k containing the best model weights, optimizer, scheduler, configuration and training configuration at epoch k will also appear in my_model/MODEL_NAME_training_YYYY-MM-DD_hh-mm-ss.

Lauching a training on benchmark datasets

We also provide a training script example here that can be used to train the models on benchmarks datasets (mnist, cifar10, celeba ...). The script can be launched with the following commandline

python training.py --dataset mnist --model_name ae --model_config 'configs/ae_config.json' --training_config 'configs/base_training_config.json'

See README.md for further details on this script

Launching data generation

Using the GenerationPipeline

The easiest way to launch a data generation from a trained model consists in using the built-in GenerationPipeline provided in Pythae. Say you want to generate 100 samples using a MAFSampler all you have to do is 1) relaod the trained model, 2) define the sampler's configuration and 3) create and launch the GenerationPipeline as follows

>>> from pythae.models import AutoModel
>>> from pythae.samplers import MAFSamplerConfig
>>> from pythae.pipelines import GenerationPipeline
>>> # Retrieve the trained model
>>> my_trained_vae = AutoModel.load_from_folder(
...	'path/to/your/trained/model'
... )
>>> my_sampler_config = MAFSamplerConfig(
...	n_made_blocks=2,
...	n_hidden_in_made=3,
...	hidden_size=128
... )
>>> # Build the pipeline
>>> pipe = GenerationPipeline(
...	model=my_trained_vae,
...	sampler_config=my_sampler_config
... )
>>> # Launch data generation
>>> generated_samples = pipe(
...	num_samples=args.num_samples,
...	return_gen=True, # If false returns nothing
...	train_data=train_data, # Needed to fit the sampler
...	eval_data=eval_data, # Needed to fit the sampler
...	training_config=BaseTrainerConfig(num_epochs=200) # TrainingConfig to use to fit the sampler
... )

Using the Samplers

Alternatively, you can launch the data generation process from a trained model directly with the sampler. For instance, to generate new data with your sampler, run the following.

>>> from pythae.models import AutoModel
>>> from pythae.samplers import NormalSampler
>>> # Retrieve the trained model
>>> my_trained_vae = AutoModel.load_from_folder(
...	'path/to/your/trained/model'
... )
>>> # Define your sampler
>>> my_samper = NormalSampler(
...	model=my_trained_vae
... )
>>> # Generate samples
>>> gen_data = my_samper.sample(
...	num_samples=50,
...	batch_size=10,
...	output_dir=None,
...	return_gen=True
... )

If you set output_dir to a specific path, the generated images will be saved as .png files named 00000000.png, 00000001.png ... The samplers can be used with any model as long as it is suited. For instance, a GaussianMixtureSampler instance can be used to generate from any model but a VAMPSampler will only be usable with a VAMP model. Check here to see which ones apply to your model. Be carefull that some samplers such as the GaussianMixtureSampler for instance may need to be fitted by calling the fit method before using. Below is an example for the GaussianMixtureSampler.

>>> from pythae.models import AutoModel
>>> from pythae.samplers import GaussianMixtureSampler, GaussianMixtureSamplerConfig
>>> # Retrieve the trained model
>>> my_trained_vae = AutoModel.load_from_folder(
...	'path/to/your/trained/model'
... )
>>> # Define your sampler
... gmm_sampler_config = GaussianMixtureSamplerConfig(
...	n_components=10
... )
>>> my_samper = GaussianMixtureSampler(
...	sampler_config=gmm_sampler_config,
...	model=my_trained_vae
... )
>>> # fit the sampler
>>> gmm_sampler.fit(train_dataset)
>>> # Generate samples
>>> gen_data = my_samper.sample(
...	num_samples=50,
...	batch_size=10,
...	output_dir=None,
...	return_gen=True
... )

Define you own Autoencoder architecture

Pythae provides you the possibility to define your own neural networks within the VAE models. For instance, say you want to train a Wassertstein AE with a specific encoder and decoder, you can do the following:

>>> from pythae.models.nn import BaseEncoder, BaseDecoder
>>> from pythae.models.base.base_utils import ModelOutput
>>> class My_Encoder(BaseEncoder):
...	def __init__(self, args=None): # Args is a ModelConfig instance
...		BaseEncoder.__init__(self)
...		self.layers = my_nn_layers()
...		
...	def forward(self, x:torch.Tensor) -> ModelOutput:
...		out = self.layers(x)
...		output = ModelOutput(
...			embedding=out # Set the output from the encoder in a ModelOutput instance 
...		)
...		return output
...
... class My_Decoder(BaseDecoder):
...	def __init__(self, args=None):
...		BaseDecoder.__init__(self)
...		self.layers = my_nn_layers()
...		
...	def forward(self, x:torch.Tensor) -> ModelOutput:
...		out = self.layers(x)
...		output = ModelOutput(
...			reconstruction=out # Set the output from the decoder in a ModelOutput instance
...		)
...		return output
...
>>> my_encoder = My_Encoder()
>>> my_decoder = My_Decoder()

And now build the model

>>> from pythae.models import WAE_MMD, WAE_MMD_Config
>>> # Set up the model configuration 
>>> my_wae_config = model_config = WAE_MMD_Config(
...	input_dim=(1, 28, 28),
...	latent_dim=10
... )
...
>>> # Build the model
>>> my_wae_model = WAE_MMD(
...	model_config=my_wae_config,
...	encoder=my_encoder, # pass your encoder as argument when building the model
...	decoder=my_decoder # pass your decoder as argument when building the model
... )

important note 1: For all AE-based models (AE, WAE, RAE_L2, RAE_GP), both the encoder and decoder must return a ModelOutput instance. For the encoder, the ModelOutput instance must contain the embbeddings under the key embedding. For the decoder, the ModelOutput instance must contain the reconstructions under the key reconstruction.

important note 2: For all VAE-based models (VAE, BetaVAE, IWAE, HVAE, VAMP, RHVAE), both the encoder and decoder must return a ModelOutput instance. For the encoder, the ModelOutput instance must contain the embbeddings and log-covariance matrices (of shape batch_size x latent_space_dim) respectively under the key embedding and log_covariance key. For the decoder, the ModelOutput instance must contain the reconstructions under the key reconstruction.

Using benchmark neural nets

You can also find predefined neural network architectures for the most common data sets (i.e. MNIST, CIFAR, CELEBA ...) that can be loaded as follows

>>> from pythae.models.nn.benchmark.mnist import (
...	Encoder_Conv_AE_MNIST, # For AE based model (only return embeddings)
...	Encoder_Conv_VAE_MNIST, # For VAE based model (return embeddings and log_covariances)
...	Decoder_Conv_AE_MNIST
... )

Replace mnist by cifar or celeba to access to other neural nets.

Distributed Training with Pythae

As of v0.1.0, Pythae now supports distributed training using PyTorch's DDP. It allows you to train your favorite VAE faster and on larger dataset using multi-gpu and/or multi-node training.

To do so, you can build a python script that will then be launched by a launcher (such as srun on a cluster). The only thing that is needed in the script is to specify some elements relative to the distributed environment (such as the number of nodes/gpus) directly in the training configuration as follows

>>> training_config = BaseTrainerConfig(
...     num_epochs=10,
...     learning_rate=1e-3,
...     per_device_train_batch_size=64,
...     per_device_eval_batch_size=64,
...     train_dataloader_num_workers=8,
...     eval_dataloader_num_workers=8,
...     dist_backend="nccl", # distributed backend
...     world_size=8 # number of gpus to use (n_nodes x n_gpus_per_node),
...     rank=5 # process/gpu id,
...     local_rank=1 # node id,
...     master_addr="localhost" # master address,
...     master_port="12345" # master port,
... )

See this example script that defines a multi-gpu VQVAE training on ImageNet dataset. Please note that the way the distributed environnement variables (world_size, rank ...) are recovered may be specific to the cluster and launcher you use.

Benchmark

Below are indicated the training times for a Vector Quantized VAE (VQ-VAE) with Pythae for 100 epochs on MNIST on V100 16GB GPU(s), for 50 epochs on FFHQ (1024x1024 images) and for 20 epochs on ImageNet-1k on V100 32GB GPU(s).

	Train Data	1 GPU	4 GPUs	2x4 GPUs
MNIST (VQ-VAE)	28x28 images (50k)	235.18 s	62.00 s	35.86 s
FFHQ 1024x1024 (VQVAE)	1024x1024 RGB images (60k)	19h 1min	5h 6min	2h 37min
ImageNet-1k 128x128 (VQVAE)	128x128 RGB images (~ 1.2M)	6h 25min	1h 41min	51min 26s

For each dataset, we provide the benchmarking scripts here

Sharing your models with the HuggingFace Hub 🤗

Pythae also allows you to share your models on the HuggingFace Hub. To do so you need:

• a valid HuggingFace account
• the package huggingface_hub installed in your virtual env. If not you can install it with

$ python -m pip install huggingface_hub

• to be logged in to your HuggingFace account using

$ huggingface-cli login

Uploading a model to the Hub

Any pythae model can be easily uploaded using the method push_to_hf_hub

>>> my_vae_model.push_to_hf_hub(hf_hub_path="your_hf_username/your_hf_hub_repo")

Note: If your_hf_hub_repo already exists and is not empty, files will be overridden. In case, the repo your_hf_hub_repo does not exist, a folder having the same name will be created.

Downloading models from the Hub

Equivalently, you can download or reload any Pythae's model directly from the Hub using the method load_from_hf_hub

>>> from pythae.models import AutoModel
>>> my_downloaded_vae = AutoModel.load_from_hf_hub(hf_hub_path="path_to_hf_repo")

Monitoring your experiments with wandb 🧪

Pythae also integrates the experiment tracking tool wandb allowing users to store their configs, monitor their trainings and compare runs through a graphic interface. To be able use this feature you will need:

• a valid wandb account
• the package wandb installed in your virtual env. If not you can install it with

$ pip install wandb

• to be logged in to your wandb account using

$ wandb login

Creating a WandbCallback

Launching an experiment monitoring with wandb in pythae is pretty simple. The only thing a user needs to do is create a WandbCallback instance...

>>> # Create you callback
>>> from pythae.trainers.training_callbacks import WandbCallback
>>> callbacks = [] # the TrainingPipeline expects a list of callbacks
>>> wandb_cb = WandbCallback() # Build the callback 
>>> # SetUp the callback 
>>> wandb_cb.setup(
...	training_config=your_training_config, # training config
...	model_config=your_model_config, # model config
...	project_name="your_wandb_project", # specify your wandb project
...	entity_name="your_wandb_entity", # specify your wandb entity
... )
>>> callbacks.append(wandb_cb) # Add it to the callbacks list

...and then pass it to the TrainingPipeline.

>>> pipeline = TrainingPipeline(
...	training_config=config,
...	model=model
... )
>>> pipeline(
...	train_data=train_dataset,
...	eval_data=eval_dataset,
...	callbacks=callbacks # pass the callbacks to the TrainingPipeline and you are done!
... )
>>> # You can log to https://wandb.ai/your_wandb_entity/your_wandb_project to monitor your training

See the detailed tutorial

Monitoring your experiments with mlflow 🧪

Pythae also integrates the experiment tracking tool mlflow allowing users to store their configs, monitor their trainings and compare runs through a graphic interface. To be able use this feature you will need:

• the package mlfow installed in your virtual env. If not you can install it with

$ pip install mlflow

Creating a MLFlowCallback

Launching an experiment monitoring with mlfow in pythae is pretty simple. The only thing a user needs to do is create a MLFlowCallback instance...

>>> # Create you callback
>>> from pythae.trainers.training_callbacks import MLFlowCallback
>>> callbacks = [] # the TrainingPipeline expects a list of callbacks
>>> mlflow_cb = MLFlowCallback() # Build the callback 
>>> # SetUp the callback 
>>> mlflow_cb.setup(
...	training_config=your_training_config, # training config
...	model_config=your_model_config, # model config
...	run_name="mlflow_cb_example", # specify your mlflow run
... )
>>> callbacks.append(mlflow_cb) # Add it to the callbacks list

...and then pass it to the TrainingPipeline.

>>> pipeline = TrainingPipeline(
...	training_config=config,
...	model=model
... )
>>> pipeline(
...	train_data=train_dataset,
...	eval_data=eval_dataset,
...	callbacks=callbacks # pass the callbacks to the TrainingPipeline and you are done!
... )

you can visualize your metric by running the following in the directory where the ./mlruns

$ mlflow ui 

See the detailed tutorial

Monitoring your experiments with comet_ml 🧪

Pythae also integrates the experiment tracking tool comet_ml allowing users to store their configs, monitor their trainings and compare runs through a graphic interface. To be able use this feature you will need:

• the package comet_ml installed in your virtual env. If not you can install it with

$ pip install comet_ml

Creating a CometCallback

Launching an experiment monitoring with comet_ml in pythae is pretty simple. The only thing a user needs to do is create a CometCallback instance...

>>> # Create you callback
>>> from pythae.trainers.training_callbacks import CometCallback
>>> callbacks = [] # the TrainingPipeline expects a list of callbacks
>>> comet_cb = CometCallback() # Build the callback 
>>> # SetUp the callback 
>>> comet_cb.setup(
...	training_config=training_config, # training config
...	model_config=model_config, # model config
...	api_key="your_comet_api_key", # specify your comet api-key
...	project_name="your_comet_project", # specify your wandb project
...	#offline_run=True, # run in offline mode
...	#offline_directory='my_offline_runs' # set the directory to store the offline runs
... )
>>> callbacks.append(comet_cb) # Add it to the callbacks list

...and then pass it to the TrainingPipeline.

>>> pipeline = TrainingPipeline(
...	training_config=config,
...	model=model
... )
>>> pipeline(
...	train_data=train_dataset,
...	eval_data=eval_dataset,
...	callbacks=callbacks # pass the callbacks to the TrainingPipeline and you are done!
... )
>>> # You can log to https://comet.com/your_comet_username/your_comet_project to monitor your training

See the detailed tutorial

Getting your hands on the code

To help you to understand the way pythae works and how you can train your models with this library we also provide tutorials:

• making_your_own_autoencoder.ipynb shows you how to pass your own networks to the models implemented in pythae

• custom_dataset.ipynb shows you how to use custom datasets with any of the models implemented in pythae

• hf_hub_models_sharing.ipynb shows you how to upload and download models for the HuggingFace Hub

• wandb_experiment_monitoring.ipynb shows you how to monitor you experiments using wandb

• mlflow_experiment_monitoring.ipynb shows you how to monitor you experiments using mlflow

• comet_experiment_monitoring.ipynb shows you how to monitor you experiments using comet_ml

• models_training folder provides notebooks showing how to train each implemented model and how to sample from it using pythae.samplers.

• scripts folder provides in particular an example of a training script to train the models on benchmark data sets (mnist, cifar10, celeba ...)

Dealing with issues 🛠️

If you are experiencing any issues while running the code or request new features/models to be implemented please open an issue on github.

Contributing 🚀

You want to contribute to this library by adding a model, a sampler or simply fix a bug ? That's awesome! Thank you! Please see CONTRIBUTING.md to follow the main contributing guidelines.

Results

Reconstruction

First let's have a look at the reconstructed samples taken from the evaluation set.

Models	MNIST	CELEBA
Eval data
AE
VAE
Beta-VAE
VAE Lin NF
VAE IAF
Disentangled Beta-VAE
FactorVAE
BetaTCVAE
IWAE
MSSSIM_VAE
WAE
INFO VAE
VAMP
SVAE
Adversarial_AE
VAE_GAN
VQVAE
HVAE
RAE_L2
RAE_GP
Riemannian Hamiltonian VAE (RHVAE)

---

Generation

Here, we show the generated samples using using each model implemented in the library and different samplers.

Models	MNIST	CELEBA
AE + GaussianMixtureSampler
VAE + NormalSampler
VAE + GaussianMixtureSampler
VAE + TwoStageVAESampler
VAE + MAFSampler
Beta-VAE + NormalSampler
VAE Lin NF + NormalSampler
VAE IAF + NormalSampler
Disentangled Beta-VAE + NormalSampler
FactorVAE + NormalSampler
BetaTCVAE + NormalSampler
IWAE + Normal sampler
MSSSIM_VAE + NormalSampler
WAE + NormalSampler
INFO VAE + NormalSampler
SVAE + HypershereUniformSampler
VAMP + VAMPSampler
Adversarial_AE + NormalSampler
VAEGAN + NormalSampler
VQVAE + MAFSampler
HVAE + NormalSampler
RAE_L2 + GaussianMixtureSampler
RAE_GP + GaussianMixtureSampler
Riemannian Hamiltonian VAE (RHVAE) + RHVAE Sampler

Citation

If you find this work useful or use it in your research, please consider citing us

@inproceedings{chadebec2022pythae,
 author = {Chadebec, Cl\'{e}ment and Vincent, Louis and Allassonniere, Stephanie},
 booktitle = {Advances in Neural Information Processing Systems},
 editor = {S. Koyejo and S. Mohamed and A. Agarwal and D. Belgrave and K. Cho and A. Oh},
 pages = {21575--21589},
 publisher = {Curran Associates, Inc.},
 title = {Pythae: Unifying Generative Autoencoders in Python - A Benchmarking Use Case},
 volume = {35},
 year = {2022}
}