Neural Injective Geometry

Neural Injective Geometry networks (NIGnets) for non-self-intersecting geometry.

Installation

To pip install run:

$ pip install git+https://github.com/atharvaaalok/NIGnets.git

Basic Usage

To create a basic NIGnet and sample points on the shape represented by the network follow:

import torch
from torch import nn
from NIGnets import NIGnet

# Create NIGnet object
nig_net = NIGnet(layer_count = 4, act_fn = nn.Tanh)

# Generate sample points on the curve represented by NIGnet
num_pts = 1000
t = torch.linspace(0, 1, num_pts).reshape(-1, 1)
Xc = nig_net(t) # Shape: (num_pts, 2)

License

Distributed under the MIT License.

Project Plan

• Add .gitignore for the project.
• Create first cut documentation pages using Jupyterbooks and MyST markdown.
  • Motivation for non-self-intersecting geometry.
  • Add .gitignore for MyST markdown.
  • Launch web page for documentation using github pages.
• Create Injective Networks.
  • Basic architecture.
  • Generate proper documentation.
    • Proper docstrings. Follow Google python coding style guide and numpy style guide.
    • Use math equations.
    • Use type annotations.
  • Impossible intersection using matrix exponential.
• Use geosimilarity for loss functions.
• Add testing code.
  • Create automated training function.
  • Create plot function. Parameterized and target shape comparison.
  • Generate a bunch of target shapes. Use SVGs.
• Update documentation with Injective Networks and showcase.
  • Add documentation on Injective Networks.
  • Fit basic shapes using Injective Networks.
  • Create showcase for Injective Networks.
  • Create showcase for Injective Networks with impossible intersection.
• Add license.
• Create logos.
  • Create logo.
  • Create favicon.
  • Use on website.
• Add Monotonic networks.
  • Add Min-Max nets.
  • Add Smooth Min-Max nets.
  • Add documentation for Monotonic Nets.
  • Add showcase for Monotonic Nets.
• Create ResNet-like architecture using skip connections.
  • Add skip connections that preserve injectivity.
  • Add showcase for ResNet architecture.
• Add Auxilliary networks.
  • Add Pre-Aux nets.
  • Add Post-Aux nets.
  • Add documentation for Aux Nets.
  • Add showcase for Pre and Post nets separately and combined.
• Fit repeating and fractal shapes.
  • Use trignometric activations in Pre-Aux networks.
• Showcase for the full architecture.
• 3D NIGnets.
  • Create documentation for 3D NIGnets.
  • Create 3D surface point clouds to fit to.
  • Fit 3D geometric shapes for showcase.
• Experiment with different geometric loss functions.