TensorFlow-based rewrite of https://github.com/mzucker/imfit
You will need:
- TensorFlow (I used v1.3 and 1.7, should work with versions in between as well)
- PIL or Pillow (for Image loading/saving)
To do a full fit, just run fitme.sh. See that file for
example runtimes and objective function values collected with my
NVidia GTX 1080 GPU. Your runtimes and objective function values may
vary due to hardware differences.
If you just want to check the accuracy of an existing fit or visualize its output, you can run something like:
python ./tf-imfit.py -w images/zz_rect_weights.png images/zz_rect.png \
-s256 -i params/zz_rect_weighted_256px.txt -p512 -T 0
It wil create an out.png visualization that shows the approximated
image as well as the per-pixel error. Although you will want a fancy
GPU to do the error minimization, visualization/verification like the
command above should run in reasonable time on a CPU.