Image regression

A toy application that learns a mapping from (x, y) coordinates to color. Uses pytorch and pytorch-lightning.

Setup

Set up your python environment with conda like this:

conda env create

Usage

Train on a single greyscale image

python train.py -i keyboard.png --num-epochs 150

Train on a single color image (RGBA)

TODO

Train on multiple images

python train.py -i boxy_stripes2.png boxy_stripes2_30.png boxy_stripes2_60.png boxy_stripes2_90.png

Train on multiple color images (RGBA)

TODO

Interpolate between the images:

python interpolate_between_many.py --width 64 --height 64 --num-images 4 --model boxy_stripes2_boxy_stripes2_30_boxy_stripes2_60_bo_2f59b848.onnx --use-cuda 0

Examples

Name	Original	Learned image
Keyboard
8x8 Checkerboard

The following animation visualizes the output of a neural network that was trained on 12 different images (different rotations of boxy stripes). The input vectors are constructed to interpolate between the 12 images, so we get a kind of morphing effect.