Sample visualization

disentanglement_lib is an open-source library for research on learning disentangled representation. It supports a variety of different models, metrics and data sets:

disentanglement_lib was created by Olivier Bachem and Francesco Locatello at Google Brain Zurich for the large-scale empirical study

Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations. Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Rätsch, Sylvain Gelly, Bernhard Schölkopf, Olivier Bachem. ICML (Best Paper Award), 2019.

The code is tested with Python 3 and is meant to be run on Linux systems (such as a Google Cloud Deep Learning VM). It uses TensorFlow, Scipy, Numpy, Scikit-Learn, TFHub and Gin.

How does it work?

disentanglement_lib consists of several different steps:

All configuration details and experimental results of the different steps are saved and propagated along the steps (see below for a description). At the end, they can be aggregated in a single JSON file and analyzed with Pandas.


Installing disentanglement_lib

First, clone this repository with

git clone https://github.com/google-research/disentanglement_lib.git

Then, navigate to the repository (with cd disentanglement_lib) and run

pip install .[tf_gpu]

(or pip install .[tf] for TensorFlow without GPU support). This should install the package and all the required dependencies. To verify that everything works, simply run the test suite with


Downloading the data sets

To download the data required for training the models, navigate to any folder and run


which will install all the required data files (except for Shapes3D which is not publicly released) in the current working directory. For convenience, we recommend to set the environment variable DISENTANGLEMENT_LIB_DATA to this path, for example by adding

export DISENTANGLEMENT_LIB_DATA=<path to the data directory>

to your .bashrc file. If you choose not to set the environment variable DISENTANGLEMENT_LIB_DATA, disentanglement_lib will always look for the data in your current folder.

Reproducing prior experiments

To fully train and evaluate one of the 12'600 models in the paper Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations, simply run

dlib_reproduce --model_num=<?>

where <?> should be replaced with a model index between 0 and 12'599 which corresponds to the ID of which model to train. This will take a couple of hours and add a folder output/<?> which contains the trained model (including checkpoints and TFHub modules), the experimental results (in JSON format) and visualizations (including GIFs). To only print the configuration of that model instead of training, add the flag --only_print.

After having trained several of these models, you can aggregate the results by running the following command (in the same folder)


which creates a results.json file with all the aggregated results.

Running different configurations

Internally, disentanglement_lib uses gin to configure hyperparameters and other settings. To train one of the provided models but with different hyperparameters, you need to write a gin config such as examples/model.gin. Then, you may use the following command

dlib_train --gin_config=examples/model.gin --model_dir=<model_output_directory>

to train the model where --model_dir specifies where the results should be saved.

To evaluate the newly trained model consistent with the evaluation protocol in the paper Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations, simply run

dlib_reproduce --model_dir=<model_output_directory> --output_directory=<output>

Similarly, you might also want to look at dlib_postprocess and dlib_evaluate if you want to customize how representations are extracted and evaluated.

Starting your own research

disentanglement_lib is easily extendible and can be used to implement new models and metrics related to disentangled representations. To get started, simply go through examples/example.py which shows you how to create your own disentanglement model and metric and how to benchmark them against existing models and metrics.

Pretrained disentanglement_lib modules

Reproducing all the 12'600 models in the study Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations requires a substantial computational effort. To foster further research, disentanglement_lib includes 10'800 pretrained disentanglement_lib modules that correspond to the results of running dlib_reproduce with --model_num=<?> between 0 and 10'799 (the other models correspond to Shapes3D which is not publicly available). Each disentanglement_lib module contains the trained model (in the form of a TFHub module), the extracted representations (also as TFHub modules) and the recorded experimental results such as the different disentanglement scores (in JSON format). This makes it easy to compare new models to the pretrained ones and to compute new disentanglement metrics on the set of pretrained models.

To access the 10'800 pretrained disentanglement_lib modules, you may download individual ones using the following link:


where <?> corresponds to a model index between 0 and 10'799 (example).

Each ZIP file in the bucket corresponds to one run of dlib_reproduce with that model number. To learn more about the used configuration settings, look at the code in disentanglement_lib/config/unsupervised_study_v1/sweep.py or run:

dlib_reproduce --model_num=<?> --only_print

Frequently asked questions

How do I make pretty GIFs of my models?

If you run dlib_reproduce, they are automatically saved to the visualizations subfolder in your output directory. Otherwise, you can use the script dlib_visualize_dataset to generate them or call the function visualize(...) in disentanglement_lib/visualize/visualize_model.py.

How are results and models saved?

After each of the main steps (training/postprocessing/evaluation), an output directory is created. For all steps, there is a results folder which contains all the configuration settings and experimental results up to that step. The gin subfolder contains the operative gin config for each step in the gin format. The json subfolder contains files with the operative gin config and the experimental results of that step but in JSON format. Finally, the aggregate subfolder contains aggregated JSON files where each file contains both the configs and results from all preceding steps.

The training step further saves the TensorFlow checkpoint (in a tf_checkpoint subfolder) and the trained model as a TFHub module (in a tfhub subfolder). Similarly, the postprocessing step saves the representation function as a TFHub module (in a tfhub subfolder). If you run dlib_reproduce, it will create subfolders for all the different substeps that you ran. In particular, it will create an output directory for each metric that you computed.

How do I access the results?

To access the results, first aggregate all the results using dlib_aggregate_results by specifying a glob pattern that captures all the results files. For example, after training a couple of different models with dlib_reproduce, you would specify

dlib_aggregate --output_path=<...>.json \

The first in the glob pattern would capture the different models, the second different representations and the last * the different metrics. Finally, you may access the aggregated results with:

from disentanglement_lib.utils import aggregate_results
df = aggregate_results.load_aggregated_json_results(output_path)

Where to look in the code?

The following provides a guide to the overall code structure:

(1) Training step:

(2) Postprocessing step:

(3) Evaluation step:

Hyperparameters and configuration files:

Shared functionality:

NeurIPS 2019 Disentanglement Challenge

The library is also used for the NeurIPS 2019 Disentanglement challenge. The challenge consists of three different datasets.

  1. Simplistic rendered images (mpi3d_toy)
  2. Realistic rendered images (mpi3d_realistic): not yet published
  3. Real world images (mpi3d_real): not yet published

    Currently, only the simplistic rendered dataset is publicly available and will be automatically downloaded by running the following command.


    Other datasets will be made available at the later stages of the competition. For more information on the competition kindly visit the competition website. More information about the dataset can be found here or in the arXiv preprint On the Transfer of Inductive Bias from Simulation to the Real World: a New Disentanglement Dataset.

Abstract reasoning experiments

The library also includes the code used for the experiments of the following paper in the disentanglement_lib/evaluation/abstract_reasoning subdirectory:

Are Disentangled Representations Helpful for Abstract Visual Reasoning? Sjoerd van Steenkiste, Francesco Locatello, Jürgen Schmidhuber, Olivier Bachem. NeurIPS, 2019.

The experimental protocol consists of two parts: First, to train the disentanglement models, one may use the the standard replication pipeline (dlib_reproduce), for example via the following command:

dlib_reproduce --model_num=<?> --study=abstract_reasoning_study_v1

where <?> should be replaced with a model index between 0 and 359 which corresponds to the ID of which model to train.

Second, to train the abstract reasoning models, one can use the automatically installed pipeline dlib_reason. To configure the model, copy and modify disentanglement_lib/config/abstract_reasoning_study_v1/stage2/example.gin as needed. Then, use the following command to train and evaluate an abstract reasoning model:

dlib_reason --gin_config=<?> --input_dir=<?> --output_dir=<?>

The results can then be found in the results subdirectory of the output directory.

Fairness experiments

The library also includes the code used for the experiments of the following paper in disentanglement_lib/evaluation/metrics/fairness.py:

On the Fairness of Disentangled Representations Francesco Locatello, Gabriele Abbati, Tom Rainforth, Stefan Bauer, Bernhard Schoelkopf, Olivier Bachem. NeurIPS, 2019.

To train and evaluate all the models, simply use the following command:

dlib_reproduce --model_num=<?> --study=fairness_study_v1

where <?> should be replaced with a model index between 0 and 12'599 which corresponds to the ID of which model to train.

If you only want to reevaluate an already trained model using the evaluation protocol of the paper, you may use the following command:

dlib_reproduce --model_dir=<model_output_directory> --output_directory=<output> --study=fairness_study_v1

UDR experiments

The library also includes the code for the Unsupervised Disentanglement Ranking (UDR) method proposed in the following paper in disentanglement_lib/bin/dlib_udr:

Unsupervised Model Selection for Variational Disentangled Representation Learning Sunny Duan, Loic Matthey, Andre Saraiva, Nicholas Watters, Christopher P. Burgess, Alexander Lerchner, Irina Higgins.

UDR can be applied to newly trained models (e.g. obtained by running dlib_reproduce) or to the existing pretrained models. After the models have been trained, their UDR scores can be computed by running:

dlib_udr --model_dirs=<model_output_directory1>,<model_output_directory2> \

The scores will be exported to <output>/results/aggregate/evaluation.json under the model_scores attribute. The scores will be presented in the order of the input model directories.


Please send any feedback to bachem@google.com and francesco.locatello@tuebingen.mpg.de.


If you use disentanglement_lib, please consider citing:

  title={Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations},
  author={Locatello, Francesco and Bauer, Stefan and Lucic, Mario and Raetsch, Gunnar and Gelly, Sylvain and Sch{\"o}lkopf, Bernhard and Bachem, Olivier},
  booktitle={International Conference on Machine Learning},

This is not an officially supported Google product.