Skip to main content

Implement GLOM, part-whole hierarchies in TensorFlow

Project description

GLOM TensorFlow Twitter

PyPI Flake8 Lint Upload Python Package Python Version


GitHub license PEP8 GitHub stars GitHub followers Twitter Follow

This Python package attempts to implement GLOM in TensorFlow, which allows advances made by several different groups transformers, neural fields, contrastive representation learning, distillation and capsules to be combined. This was suggested by Geoffrey Hinton in his paper "How to represent part-whole hierarchies in a neural network".

Further, Yannic Kilcher's video and Phil Wang's repo was very helpful for me to implement this project.


Run the following to install:

pip install glom-tf

Developing glom-tf

To install glom-tf, along with tools you need to develop and test, run the following in your virtualenv:

git clone
# or clone your own fork

cd GLOM-TensorFlow
pip install -e .[dev]

A bit about GLOM

The GLOM architecture is composed of a large number of columns which all use exactly the same weights. Each column is a stack of spatially local autoencoders that learn multiple levels of representation for what is happening in a small image patch. Each autoencoder transforms the embedding at one level into the embedding at an adjacent level using a multilayer bottom-up encoder and a multilayer top-down decoder. These levels correspond to the levels in a part-whole hierarchy.

<small>Interactions among the 3 levels in one column</small>

An example shared by the author was as an example when show a face image, a single column might converge on embedding vectors representing a nostril, a nose, a face, and a person.

At each discrete time and in each column separately, the embedding at a level is updated to be the weighted average of:

  • bottom-up neural net acting on the embedding at the level below at the previous time
  • top-down neural net acting on the embedding at the level above at the previous time
  • embedding vector at the previous time step
  • attention-weighted average of the embeddings at the same level in nearby columns at the previous time

For a static image, the embeddings at a level should settle down over time to produce similar vectors.

<small>A picture of the embeddings at a particular time</small>


from glomtf import Glom

model = Glom(dim = 512,
             levels = 5,
             image_size = 224,
             patch_size = 14)

img = tf.random.normal([1, 3, 224, 224])
levels = model(img, iters = 12) # (1, 256, 5, 12)
# 1 - batch
# 256 - patches
# 5 - levels
# 12 - dimensions

Use the return_all = True argument to get all the column and level states per iteration. This also gives you access to all the level data across iterations for clustering, from which you can inspect the islands too.

from glomtf import Glom

model = Glom(dim = 512,
             levels = 5,
             image_size = 224,
             patch_size = 14)

img = tf.random.normal([1, 3, 224, 224])
all_levels = model(img, iters = 12, return_all = True) # (13, 1, 256, 5, 12)
# 13 - time

# top level outputs after iteration 6
top_level_output = all_levels[7, :, :, -1] # (1, 256, 512)
# 1 - batch
# 256 - patches
# 512 - dimensions


    title   = {How to represent part-whole hierarchies in a neural network}, 
    author  = {Geoffrey Hinton},
    year    = {2021},
    eprint  = {2102.12627},
    archivePrefix = {arXiv},
    primaryClass = {cs.CV}

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

glom-tf-0.1.1.tar.gz (10.6 kB view hashes)

Uploaded source

Built Distribution

glom_tf-0.1.1-py3-none-any.whl (11.1 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page