🧶 Disent

A modular disentangled representation learning framework for pytorch

Visit the docs for more info, or browse the releases.

Contributions are welcome!

---

Overview

Disent is a modular disentangled representation learning framework for auto-encoders, built upon pytorch-lightning. This framework consists of various composable components that can be used to build and benchmark disentanglement pipelines.

The name of the framework is derived from both disentanglement and scientific dissent.

Goals

Disent aims to fill the following criteria:

1. Provide high quality, readable, consistent and easily comparable implementations of frameworks
2. Highlight difference between framework implementations by overriding hooks and minimising duplicate code
3. Use best practice eg. torch.distributions
4. Be extremely flexible & configurable

Citing Disent

Please use the following citation if you use Disent in your research:

@Misc{Michlo2021Disent,
  author =       {Nathan Juraj Michlo},
  title =        {Disent - A modular disentangled representation learning framework for pytorch},
  howpublished = {Github},
  year =         {2021},
  url =          {https://github.com/nmichlo/disent}
}

---

Warning ⚠️

Disent is still under active development. Features and APIs are not considered stable, and should be expected to change! A very limited set of tests currently exist which will be expanded upon in time.

---

Getting Started

The easiest way to use disent is by running experiement/run.py and changing the root config in experiements/config/config.yaml. Configurations are managed with Hydra Config. This mode is only available if you clone the repo directly.

Pypi:

1. Make sure pip3 is upgraded: pip3 install --upgrade pip

2. Install disent with: pip3 install disent (for up-to-date versions, rather clone the dev branch)

3. Visit the docs & examples!

Source:

1. Clone with: git clone --branch dev https://github.com/nmichlo/disent.git

2. Change your working directory to the root of the repo: cd disent

3. Install the requirements for python 3.8 with pip3 install -r requirements.txt

4. Run the default experiment after configuring experiment/config/config.yaml by running PYTHONPATH=. python3 experiment/run.py

---

Features

Disent includes implementations of modules, metrics and datasets from various papers. However modules marked with a "🧵" are introduced in disent for my MSc. research.

Frameworks

• Unsupervised:
  • VAE
  • Beta-VAE
  • DFC-VAE
  • DIP-VAE
  • InfoVAE
  • BetaTCVAE
• Weakly Supervised:
  • Ada-GVAE AdaVae(..., average_mode='gvae') Usually better than the Ada-ML-VAE
  • Ada-ML-VAE AdaVae(..., average_mode='ml-vae')
• Supervised:
  • TVAE
• Experimental:
  • 🧵 Ada-TVAE
    • Adaptive Triplet VAE
  • 🧵 DO-TVAE (DO-Ada-TVAE)
    • Data Overlap Adaptive Triplet VAE
  • various others not worth mentioning

Many popular disentanglement frameworks still need to be added, please submit an issue if you have a request for an additional framework.

todo

• FactorVAE
• GroupVAE
• MLVAE

Metrics

• Disentanglement:
  • FactorVAE Score
  • DCI
  • MIG
  • SAP
  • Unsupervised Scores
  • 🧵 Flatness Score
    • Measures max width (furthest two points) over path length (sum of distances between consecutive points) of factor traversal embeddings. A combined measure of linearity and ordering, (weighted towards axis alignment if l2 width over l1 path length is used).
  • 🧵 Flatness Components - Linearity & Axis Alignment
    • Measure linearity, how much the largest eigen vector explains a factor traversal, ie. the largest singular value of latent variables over the sum of singular values.
    • Measure axis-alignment, how much the largest standard basis vector explains a factor traversal, ie. the largest standard deviation of latent variables over the sum of standard deviations.

Some popular metrics still need to be added, please submit an issue if you wish to add your own, or you have a request.

todo

• DCIMIG
• Modularity and Explicitness

Datasets:

Various common datasets used in disentanglement research are implemented, as well as new sythetic datasets that are generated programatically on the fly. These are convenient and lightweight, not requiring storage space.

• Ground Truth:

  • Cars3D
  • dSprites
  • MPI3D
  • SmallNORB
  • Shapes3D

• Ground Truth Synthetic:

  • 🧵 XYSquares: (non-overlapping) 3 squares (R, G, B) that move across a non-overlapping grid. Obervations have no channel-wise loss overlap.
  • 🧵 XYObject: A simplistic version of dSprites with a single square.
  • 🧵 XYBlocks: 3 blocks of decreasing size that move across a grid. Blocks can be one of three colors R, G, B. if a smaller block overlaps a larger one and is the same color, the block is xor'd to black.

  Input Transforms + Input/Target Augmentations

  • Input based transforms are supported.
  • Input and Target CPU and GPU based augmentations are supported.

Schedules/Annealing:

Hyper-parameter annealing is supported through the use of schedules. The currently implemented schedules include:

• Linear Schedule
• Cyclic Schedule
• Cosine Wave Schedule
• Various other wrapper schedules

---

Why?

• Created as part of my Computer Science MSc scheduled for completion in 2021.

• I needed custom high quality implementations of various VAE's.

• A pytorch version of disentanglement_lib.

• I didn't have time to wait for Weakly-Supervised Disentanglement Without Compromises to release their code as part of disentanglement_lib. (As of September 2020 it has been released, but has unresolved discrepencies).

• disentanglement_lib still uses outdated Tensorflow 1.0, and the flow of data is unintuitive because of its use of Gin Config.

---

Architecture

disent

• disent/data: raw groundtruth datasets
• disent/dataset: dataset wrappers & sampling strategies
• disent/framework: frameworks, including Auto-Encoders and VAEs
• disent/metrics: metrics for evaluating disentanglement using ground truth datasets
• disent/model: common encoder and decoder models used for VAE research
• disent/schedule: annealing schedules that can be registered to a framework
• disent/transform: transform operations for processing & augmenting input and target data from datasets

experiment

• experiment/run.py: entrypoint for running basic experiments with hydra config
• experiment/config: root folder for hydra config files
• experiment/util: various helper code, pytorch lightning callbacks & visualisation tools for experiments

---

Example Code

The following is a basic working example of disent that trains a BetaVAE with a cyclic beta schedule and evaluates the trained model with various metrics.

Basic Example

import pytorch_lightning as pl
from torch.optim import Adam
from torch.utils.data import DataLoader
from disent.data.groundtruth import XYObjectData
from disent.dataset.groundtruth import GroundTruthDataset
from disent.frameworks.vae import BetaVae
from disent.metrics import metric_dci, metric_mig
from disent.model.ae import EncoderConv64, DecoderConv64, AutoEncoder
from disent.schedule import CyclicSchedule
from disent.transform import ToStandardisedTensor

# We use this internally to test this script.
# You can remove all references to this in your own code.
from disent.util import is_test_run

# create the dataset & dataloaders
# - ToStandardisedTensor transforms images from numpy arrays to tensors and performs checks
data = XYObjectData()
dataset = GroundTruthDataset(data, transform=ToStandardisedTensor())
dataloader = DataLoader(dataset=dataset, batch_size=4, shuffle=True)

# create the BetaVAE model
# - adjusting the beta, learning rate, and representation size.
module = BetaVae(
    make_optimizer_fn=lambda params: Adam(params, lr=5e-4),
    make_model_fn=lambda: AutoEncoder(
        # z_multiplier is needed to output mu & logvar when parameterising normal distribution
        encoder=EncoderConv64(x_shape=dataset.x_shape, z_size=6, z_multiplier=2),
        decoder=DecoderConv64(x_shape=dataset.x_shape, z_size=6),
    ),
    cfg=BetaVae.cfg(beta=0.004)
)

# cyclic schedule for target 'beta' in the config/cfg. The initial value from the
# config is saved and multiplied by the ratio from the schedule on each step.
# - based on: https://arxiv.org/abs/1903.10145
module.register_schedule('beta', CyclicSchedule(
    period=1024,  # repeat every: trainer.global_step % period
))

# train model
# - for 65536 batches/steps
trainer = pl.Trainer(logger=False, checkpoint_callback=False, max_steps=65536, fast_dev_run=is_test_run())
trainer.fit(module, dataloader)

# compute disentanglement metrics
# - we cannot guarantee which device the representation is on
# - this will take a while to run
get_repr = lambda x: module.encode(x.to(module.device))

metrics = {
    **metric_dci(dataset, get_repr, num_train=10 if is_test_run() else 1000, num_test=5 if is_test_run() else 500, show_progress=True),
    **metric_mig(dataset, get_repr, num_train=20 if is_test_run() else 2000),
}

# evaluate
print('metrics:', metrics)

Visit the docs for more examples!

---