Skip to content

Latest commit

 

History

History
98 lines (75 loc) · 4.69 KB

README.md

File metadata and controls

98 lines (75 loc) · 4.69 KB
Raw

About

This repository contains research code for exploring the capabilities of simple neural networks to learn structured information about visual objects.

Usage

Getting Started

The code is designed to run on multiple machines, at the UIUC Learning & Language Lab using a custom job submission system called Ludwig. If you have access to the lab's file server, you can submit jobs with Ludwig.

To install Ludwig and all required dependencies, create a new virtual Python environment, then

 pip install -r requirements.txt

Note: The machines that will execute jobs may not have the same version of torch or numpy. Check that you are using the same versions used by Ludwig workers here.

Running jobs

To define which jobs will be executed, edit polyomino_world/params.py. To run each job 10 times,

ludwig -r10

Alternatively, jobs can be run locally, and results will be saved on the shared drive:

ludwig --local

To develop locally, without moving results to the shared drive:

ludwig --isolated

Visualizing results

Clone LudwigViz, navigate to the root directory, then

python -m flask run

Alternatively, use the provided script, plot_results.py in scripts. Before results can be plotted, create an environment variable, LUDWIG_MNT that points to the path where the shared drive is mounted on your machine.

Changelog

February 21, 2021

  • integrate with Ludwig==3.0.0 job submission system.
  • consolidated training and data creation scripts into a single function, job.main.
  • simplify evaluation of models by saving performance data with Ludwig, and making names of performance curves more accessible.
  • pass world.World data directly to dataset.DataSet instead of writing to and reading from disk.
  • isolate instances of shapes.Shape from world.World.
  • simplify generation of world data by explicitly generating event sequences with a single shape. Generation of events involving multiple shapes were never implemented directly.
  • move logic for computing wold_cells in shapes.Shape to world.World.
  • add dataclasses in helpers.py to enforce consistent attribute access and naming, and simplify attribute access.
  • manipulate paths with pathlib instead of os.
  • use raise instead of sys.exit() to exit program.
  • use Python3.7 string formatting, e.g. learning rate={}.format(params.lr) -> flearning rate={params.lr}
  • import module-level objects instead of a module, e.g. from polyominoworld import herlpers -> from polyominoworld.helpers import Action.
  • shorten variable names : e.g. feature_list -> features.
  • add all instance initialization logic to __init__ to make it easier to find the logic responsible for populating previously empty/placeholder attributes.
  • remove duplicated code in shapes.py.
  • remove any custom logic for writing, and reading data in world.py, network.py, and evaluate.py.
  • use saving and loading functions provided by torch to save a model's data.
  • remove option to generate shapes in random world locations to temporarily consolidate logic for generating event sequences.
  • add README.md, and requirements.txt.
  • remove outdated notes.txt.
  • add Python3 type hints to function arguments.
  • add brief summaries to the start of most Python files.
  • clearly demarcate library code from scripting code by moving scripts into scripts.
  • remove subdirectories from source code folder to simplify navigation.
  • remove unused __init__.py files.
  • remove many redundant or unnecessary dict and list objects to reduce chances of bug that would result by a mismatch in data that available in multiple objects.
  • implemented option to train on gpu.
  • remove unnecessary .float() calls on torch objects.
  • rename 'HiddenState' to 'hidden', 'WorldState' to 'world', and 'FeatureVector' to 'features' to match lower-cased strings used everywhere else.
  • use binary cross-entropy instead of MSE loss when y_type='features
  • prepend _ to method names to indicate they are used to perform "private" logic, and to distinguish them from methods which may be called externally.
  • add doc strings to functions
  • remove shape.id_number and populate active world cells with shape.color instead which can be directly used when creating WorldVector.
  • consolidate single and multi-layer networks into a single network capable of both.

Compatibility

Developed on Ubuntu 18.04 with Python 3.7.9 Versions of Pytorch above 1.6 are not available for versions of Python above Python3.7.9.