Image Compression with Flow Models
- Found new approaches of using flow models to do compression tasks for image‐like data. Wrote processing and feature extracting functions for 80K+ target data. Tested VQ‐VAE and basic Real NVP model to build unsupervising learning baselines;
- Based on open source codes, wrote hundreds lines of extra codes to restore Real NVP Compression model introduced by the SOTA papers. Did experiments with different feature extractors on the 80K+ 2D data; it turned out that the permutations of pixels will influence the results;
- Found Glow model’s learning potential; transplanted it from TensorFlow to PyTorch;
- Conclusion: flow models have great potential on distribution transformation, but better permutated data is also necessary.
Kelley Kan HUANG
Engineer (PI)
My research interests include distributed robotics, mobile computing and programmable matter.