cosmobot-deep-learning

Cosmobot deep learning models and helper code

Iterating on models

When to branch

This repo will contain one directory for each major type of model, with the best version of each checked into master.

If you are making tweaks to a model, do so in a branch. If the tweaks turn out to be an improvement on that model, land them. If you are creating an new type of model (rather than iterating on an existing model), create a new directory for it.

Documenting changes

Create a directory in the Experiments directory on Google Drive with a README detailing your changes and results.

Reference your branch name or changeset in the README.

Iterating with Jupyter notebooks

TODO: https://app.asana.com/0/819671808102776/1130875537031890/f

Datasets

The dataset csvs are generated independently and copied into this repo. For more context and a changelog, see: ML Dataset README

Terminology

Some standard terminology around our raw image data and how we process it

COPY-PASTA: These definitions have been copied from the cosmobot-process-experiment repo

• RAW image file - A JPEG+RAW image file as directly captured by a PiCam v2, saved as a .JPEG
• RGB image - A 3D numpy.ndarray: a 2D array of "pixels" (row-major), where each "pixel" is a 1D array of [red, green, blue] channels with a value between 0 and 1. This is our default format for interacting with images. An example 4-pixel (2x2) image would have this shape:

[
 [ [r1, g1, b1], [r2, g2, b2] ],
 [ [r3, g3, b3], [r4, g4, b4] ]
]

• ROI - An RGB image that has been cropped to a specific Region of Interest (ROI).
• ROI definition - A 4-tuple in the format provided by cv2.selectROI: (start_col, start_row, cols, rows), used to define a Region of Interest (ROI).