PyTorch implementation of the Multi-View Information Bottleneck (MIB) model from "Learning Robust Representations via Multi-View Information Bottleneck".
Please read our paper for a more detailed description of the training procedure.
Slides used for the ICLR 2020 poster session can be found here.
Please use the following bibtex for citations:
@inproceedings{federici2020,
title={Learning Robust Representations via Multi-View Information Bottleneck},
author={Marco Federici and Anjan Dutta and Patrick Forré and Nate Kushman and Zeynep Akata},
booktitle={International Conference on Learning Representations},
year={2020}
}
The conda and pip dependencies required to run our code are specified in the environment.yml
environment file and can be installed by running
conda env create -f environment.yml
Activate the environment
conda activate mib
This project provides code to train and evaluate different architectures in unsupervised/self-supervised settings. Each training procedure is described by a `.yml' file which specifies loss function and the value for the respective hyper-parameters.
In order to train a model starting from a random weight initialization, use
python train_representation.py <EXPERIMENT_FOLDER> --config-file=<CONFIG_FILE>
Where <EXPERIMENT_FOLDER>
refers to the directory in which model checkpoints and training log will be stored, while <CONFIG_FILE>
must refer to a valid training configuration file in which the training procedure is specified together with its hyperparameters.
In this project, the configuration files and code to train the following models have been included:
Variational Autoencoders (VAE) configs/VAE.yml
InfoMax configs/InfoMax.yml
Multi-View InfoMax (MV-InfoMax) configs/MV_InfoMax.yml
Multi-View Information Bottleneck (MIB) configs/MIB.yml
In order to train a MIB model and store checkpoints and logs in the directory experiments/MIB
, the following command can be used:
python train_representation.py experiments/MIB --config-file=configs/MIB.yml
Additional flags can be used to specify other hyper-parameters such as number of training epochs, batch size and device used for training (cuda
or cpu
).
An extensive description of the training options can by accessed by using the --help
flag
python train_representation.py --help
To resume the training from the last available checkpoint use:
python train_representation.py <EXPERIMENT_FOLDER>
Where <EXPERIMENT_FOLDER>
must refer to the same directory containing checkpoint and configuration files created by
first running train_representation.py
using the --config-file
argument.
Logs regarding different components of the optimized objectives and performance of a classifier trained on a subset of the training representations and evaluated on the test set can be visualized using Tensorboard. This can be done by typing:
tensorboard --logdir=<EXPERIMENT_FOLDER>
Where <EXPERIMENT_FOLDER>
refers to the path to a previously start experiment
The results.ipynb
notebook provides some simple code to visualize and evaluate the quality of the different representations.
Examples of encoders trained using the VAE, InfoMax, MV-InfoMax and MIB objective can be found and loaded form the trained_models
folder.