implement first running version of Variational Autoencoder

Representation-Learning

Representation Learning of Image Data with VAE. Alexander Piehler, Moritz Wagner

Introduction

This Github Repository is developed by Moritz Wagner and Alexander Piehler in collaboration with Adidas. This project stresses the suitability of Variational Autoencoders for representation learning of image data. The main goal was to establish a codebase that allows for reproducing and tracking experiments via mlflow. The world of Variational Autoencoders is quite large and therefore, we restricted ourselves to the ones we believed that would perform best. To further benchmark, we also included PCA and Autoencoders. The following models were considered:

├── VAE
├── beta-VAE
├── InfoVAE
├── GaussMixVAE
├── DIP-VAE
├── PCA
├── Autoencoder

For each model, we found some valid hyperparameter configurations that can be considered in the folder configs/.

Code Structure

This framework is structured as follows:

├── configs
├── data
├── experiments
├── library
├── playground

`configs`

Contains the config files for each model depending on the chosen data set.

`data`

In this folder, you find all relevant python scripts to download and preprocess the data for running the models.

`experiments`

Contains the scripts for running all relevant experiments. To run them in the right order, follow these steps: To run them first change into the directory accordingly: cd experiments

python seed_running.py
python latent_experiment.py
python epochs_experiment.py
python kld_weight_experiment.py
python latent_experiment.py
python tune_models.py
python run_best_configurations.py
python deep_dive.py

`library`

This is the package for this repository, it stores all functions and (model) classes that are used. The package contains different modules

├── models2
├── architectures.py
├── eval_helpers.py
├── evaluator.py
├── postprocessing.py
├── utils.py
├── visualizer.py
└── viz_helpers.py

`playground`

Contains only some playground files which are absolutely not important to consider.

Packaging

package libraryenables pip installation and symling creation
fork the repository and execute pip install -e . from the parent tree.
packaging allows also for updating the packages in the scripts via importlib.reload()

Setup

You can setup Representation-Learning as follows: Note, all commands must be run from parent level of the repository.

Install miniconda
Create a conda environmnet for python 3.7 (conda create -n <env name> python=3.7)
Clone this repository
Install the required packages via pip install -r requirements.txt
Install the local package library as described above.
Download the data by moving to data folder cd data and executing python get_<dataset>_data.py
Run code from run.py --config/config_file.yaml

Further notes on running experiments.

In case that hyperparameters want to be adjusted, you can do so by respectively adjusting the parameters set in the config files. Note, however, the parameter configurations listed we found worked best for the respective problem. If you want to run your own experiments, it is advisable to additionally specify the run_name and the experiment_name. This is required for mlflow to log adequately. An exemplary command could be as follows: python run.py --configs/MNIST/vae.yaml --run_name mnist_vae --experiment_name mnist_vae

Experiment	Description	Models	Output
Different Seeds	try different seeds to evaluate stochasticity	VanillaVae	BoxPlot, dataframe
Epochs	evaluate the influence of no. of epocs 40, 80, 120	VanilleVae	traversal plots, disentaglement metrics
Latent Dimensions	evaluate influence of number of latent dims 10, 20, 40	VanillaVae	Active Units
ResNet as Encoder	try ResNet as encoder to evaluate the influence of the architecture on the model	VanillaVae	Traversal Plots, disentanglement metrics

Model	Parameters	Budget	Done
GaussianVae	None	None
BetaVae	beta: [low=1, high=20, steps=2]; max_capacity: [low=1, high=50, step=5]	50h	16.08 21.00h
DIPVae	lambda_dig: [low=1, high=20, steps=2]; lambda_offdig: [low=1, high=30, step=2]	50h	20.8 0.00h
GaussMixVae	temperature: [low=1, high=40, steps=10]; anneal_rate: [low=0.1, high=0.5, steps=0.1] cont_weight [1, 2, 3]; cat_weight = [1, 2, 3]	???	tbd
InfoVae	to be discussed	???	tbd

	#try:
	# #pdb.set_trace()
	# #batch_idx = kwargs['batch_idx']
	# batch_idx = batch_idx
	#
	# # Anneal the temperature at regular intervals
	# if self.batch_idx % self.anneal_interval == 0 and self.training:
	# self.temp = np.maximum(self.temp * np.exp(- self.anneal_rate * batch_idx),
	# self.min_temp)
	#except:
	# pass

moritzwag / representation-learning Goto Github PK

representation-learning's Introduction

Representation-Learning

Introduction

Code Structure

configs

data

experiments

library

playground

Packaging

Setup

Further notes on running experiments.

representation-learning's People

Contributors

Stargazers

Watchers

representation-learning's Issues

Recommend Projects

Recommend Topics

Recommend Org

`configs`

`data`

`experiments`

`library`

`playground`