torch-kaleidoscope 0.1

A tool kit for the use of non-RGB colour representations with torchvision and the PyTorch framework.

torch-kaleidoscope

A python library for providing PyTorch compatible colour representation utilities.

weights

Pretrained weights for PyTorch networks are available from my research group where they are hosted and downloadable from the following format of URL: https://cvl.cs.nott.ac.uk/resources/kaleidoscope/backbones/<network>/<colour_space>.pth

available weights

• ResNet
  • 50: RGB, HSV, H2SV, L*a*b*, L*u*v*, YUV
  • 101: RGB, H2SV
  • 152: RGB, H2SV
• ResNeXt
  • 50 32x4d: RGB, H2SV
  • 101 32x8d: RGB, H2SV
• EfficientNetV2
  • Small: RGB, H2SV
• ConvNext
  • Large: RGB, HSV, H2SV, H3SV, YUV
• Swin Transformer V2
  • Tiny: : RGB, H2SV

acknowledgements

The work that lead to needing this, as well as huge swathes of the conversions code and testing present in this repository owe their existence to the scikit-image library[1]. In particular Nicolas Pinto, Ralf Gommers, Travis Oliphant, Matt Terry, Alex Izvorski and everyone else who contributed to the color module.

references

[1] scikit-image, the scikit-image team (2014)