djcrw / generalising-structural-knowledge

The Tolman-Eichenbaum Machine

· Report Bug · Request Feature ·

Table of Contents

• About the Project
• Getting Started
  • Installation
  • Running models
  • Pytorch version
• Contact
• Acknowledgements

About The Project

This project contains a tensorflow 1.9.0 and a tensorflow 2.3.0 implementation of the Tolman-Eichenbaum Machine (paper). The tensorflow 2 (tem_tf2) version is cleaner and easier to use.

In the tensorflow 2 (tem_tf2) version, it has all the relevant code for the simulations in the recent Nature Neuroscience review (paper).

Getting Started

You need to install python 3 and tensorflow 1.9.0 or tensorflow 2.3.0

Installation

Clone the repo

git clone https://github.com/djcrw/generalising-structural-knowledge.git

Running models

python3 run_tem.py

If using graph_mode parameter, there will be some time before training starts as graph optimisaiton is taking place.

Use notebook to load and visualise cell representations and to do behavioural analyses

Pytorch version

Jacob Bakermans has made a pytorch implementation of TEM found here https://github.com/jbakermans/torch_tem

Contact

James Whittington - @jcrwhittington - jcrwhittington at gmail.com

Project Link: https://github.com/djcrw/generalising-structural-knowledge

Acknowledgements

Thanks to Jacob Bakermans for the cover image!

generalising-structural-knowledge's Issues

Are there still any plans to upload the code?

Hi!

I really enjoyed reading this paper! I'm completing an honors thesis on the hippocampal-entorhinal circuit and ML, and the ideas you presented really stood out to me. I was hoping to play around with the code to better understand the model you created. Is releasing the code still a possibility?

Best,
Yoni

run_tem.py errors

Thanks for sharing your code! I'm trying to run it in a virtualenv (python3.6) and it looks like one of the files might be slightly out of date?

I'm running: python run_tem.py and when constructing default params there are a number of KeyError exceptions starting with:

Traceback (most recent call last):
  File "run_tem.py", line 21, in <module>
    pars = default_params()
  File "generalising-structural-knowledge/parameters.py", line 28, in default_params
    params['reward_value'] = params['n_states'][0]  # make same as predicting other sensory experiences
KeyError: 'n_states'

That error is fixed by shuffling line order around, but then there is another error:

Traceback (most recent call last):
  File "run_tem.py", line 21, in <module>
    pars = default_params()
  File "generalising-structural-knowledge/parameters.py", line 76, in default_params
    params['n_senses'] = [params['s_size']] * params['n_freq']
KeyError: 'n_freq'

I'm not familiar with the code, so it's not obvious how to fix these errors. Could you update the default_params function or point me in the right direction?

PyTorch code

Hi. I wonder, when do you think you'll be able to release the PyTorch implementation of your work?

Chen-Sun, Tim Behrens