Skip to content

Latest commit

 

History

History
83 lines (69 loc) · 4.08 KB

README.md

File metadata and controls

83 lines (69 loc) · 4.08 KB
Raw

Neural Planets

Neural planets is a repository for rendering planet-like spheres with topography generated from random neural networks. The design-space is intentionally crafted to be large. With a mixture of 12 continuous and discrete design parameters (not including neural network weight parameters which can be arbitrarily large), the variety of planets is endless. This repository is dedicated to Manxin Chen.

animated animated

Installation

Usage of this software requires Python 3.6 or higher, pip, and access to a command line. To setup, run the following commands in a shell:

git clone https://github.com/jbial/neural-planets  # get the code
cd neural-planets  # move into the code directory 
pip install -r requirements.txt  # install required libraries
pip install -e .  # set up the package

Usage

0. Parameters

This software utilizes hydra-core for handling parameter settings. For a reference of every mutable parameter and it's description, refer to the YAML files in the config directory.

1. Generating planets

Use the scripts/generate_planet.py file to generate a rendering of a planet's 2D topographic projection and it's rendering onto a sphere. For example:

python scripts/generate_planet.py prefix=unique-identifying-prefix random_seed=30 noise_levels=8 model.radius=10 'model.layers=[64,64,64]' 'model.activations=[sigmoid,sin,tanh]' model.zdim=16 model.final_activation=tanh model.init_limit=2 model.min_delta=0.5 model.ffeats=-1 model.ffscale=1 model.noise_decay=0.35 model.noise_scale=2 colormap=gist_earth 'img_size=[500,500]'

2. Generating animations

Generating animations requires the same parameters as above (except for the planet interpolation animation) with the addition of the duration, view_distance, frames_per_second, and the animation_type (most important) parameters. Simply add these parameters to the command and change to the scripts/animate_planet.py file:

python scripts/animate_planet.py prefix=unique-identifying-prefix random_seed=30 noise_levels=8 duration=10 view_distance=100 frames_per_second=30 animation_type=rotation

For brevity, I left out the command line parameters for the model.

3. Planet interpolation

A special case of animations is the planet interpolation animation which corresponds to animation_type=planet_interpolation. This animation can interpolate between two different planet models, so it requires specification of parameters for a second neural network - this can be done manually by editing the parameters in the config/other.yaml file, or on the command line with additional planet parameters modified by the other.[PARAMETER] prefix.

animated

For this animation, you can run something like the above command but with animation_type=planet_interpolation, and again for brevity, you can specify the second planet's parameters in the config/other.yaml file:

other:

  noise_levels: 20

  model:
    in_dim: 3  # input dimension of model, this is fixed
    layers: [32, 16, 8, 4]
    radius: 20
    activations: [softplus, sigmoid, sin, tanh]
    final_activation: tanh
    zdim: 4
    ffeats: -1  # dimension of fourier basis space
    ffscale: 1  # scale of randomness: higher = very squiggly, lower = less squiggly
    init_limit: 5
    noise_decay: 0.25
    noise_scale: 1
    min_delta: 0.5

Note that the actual config/other.yaml file contains comments explaining each parameter shown above.

TODO

  • More _postprocess function varieties/heuristics in model.py
  • Finish Documentation
  • GPU support
  • Colormap interpolation between planets
  • More animation types (interpolation of any continuous parameter)
  • Flow maps on the surface