flight-maneuvers 1.0.1

A library for flight maneuver classification.

Steerable Inductive Biases for Flight Maneuver Reconstruction

Clayton Curry

Todo List

Project

• implement utilities for loading flight scenarios, pre-processing and post-processing, and visualizing predicted maneuvers in a Flask application
• implement trainer(model, training_data) that relies on Adam to optimize model weights in a fixed number of gradient steps
• implement resnet.py (confirm convergence)
• implement se2_resnet.py on a 1D signal (confirm convergence)
• implement se3_resnet.py on a 1D signal (confirm convergence)
• implement a suitable strategy for neural_search
• for each flight maneuver, list the resulting precision and recall values for each maneuver class together with confidence intervals
• for each model, plot precision and recall values and show trends as training examples vary
• evolve candidate models for resnet.py, se2_resnet.py, and se3_resnet.py' by:
  • fitness function: - Binary Cross-Entropy Loss,
  • varying the number of convolution layers, kernel size, kernel width, fully connected layer width
  • constaining the a weight space to have fixed computer memory
  • the same training set, validation set, and test set
• with each candidate, obtain confidence intervals for precision and recall (per maneuver class) using bootstrapping
• plot a table that, for each flight maneuver, reports precision and recall for each maneuver class together with confidence intervals

Report

• describe related works and their relation to concepts in representation theory and harmonic function theory
• relate the physical symmetries of the label function to actions of $SE(3)$ and $SE(2)$
• show the derivation for one resnet
• show the derivation for one se2_resnet equivariant convolution layer
• show the derivation for one se3_resnet equivariant convolution layer
• describe how models are evolved
• explain trends, update claims in the abstract (if necessary), and describe how the findings support the hypothesis
• conclusions and future work

Description

TODO

Related Projects

• escnn
• e3nn
• geometric gnn dojo
• e3nn-jax
• equivariant-MLP
• SignNet-BasisNet
• mace-jax

Directory Structure and Usage

.
├── README.md
├── app.py                              # opens web app on localhost created using Dash and Flask 
├── main.py                             # runs experiments on local hardware after parsing CLI arguments
│
├── examples
│   ├── introduction.ipynb              # A gentle introduction to background and methodology used in project
│   └── TODO                            # TODO
│
│
└── src
    ├── model
    │   ├── resnet_1D.py                # ordinary residual blocks and network operating on a sampled timeseries
    │   ├── se2_resnet_1D.py            # TODO
    │   └── se3_resnet_1D.py            # TODO
    │
    ├── data_module.py                  # defines classes extending pytorch data utilities 
    ├── neural_search.py                # TODO
    └── utils.py                        # standard utilities for plotting results, obtaining statistics, and pre-/postprocessing data

How to run

First, create a new conda environment using Python 3.7 and activate it as follows

# create env
conda create -n flight-maneuvers python=3.7
# activate it
conda activate flight-maneuvers

Next, clone the repo, navigate to the project base, and install requirements,

# clone project   
git clone https://github.com/clay-curry/Flight-Maneuvers
# enter project
cd Flight-Maneuvers
# install dependencies
pip install -r requirements.txt

Finally, pass a supporting argument to the main module as follows

# run module (example: mnist as your main contribution)   
python main.py  COMMAND

where COMMAND is chosen from the following table:

COMMAND	Description
fetch	invokes 'curl' to download 1000 pre-simulated training examples
train	instantiates a model and invokes routines for training
evaluate	assesses model performance using a randomized ANOVA procedure
app	starts a web server on localhost and serves models using Flask