Visual Mesh 2.1.1

This upgrade is a minor upgrade fixing a few bugs and adding the ability for the OpenCL engine to cache its built code

What's Changed

• Fix data shape for classification images by @Bidski in #9
• Fix interpolation pixel selection by @Bidski in #10
• Add ability to cache built binaries in OpenCL engine by @ysims in #11

Full Changelog: v2.1.0...v2.1.1

Visual Mesh 2.1.0

This release upgrades and improves several parts of the Visual Mesh

Upgrades:

• Upgraded the target TensorFlow version to 2.5
• Improved the quality and reliability of seeker networks by changing the loss function and allowing blank seekers
• Removed the Radial networks as they could not pass clang-tidy and are similar in performance to ring
• Add the ability for the CPU engine to process Bayer pattern images
• Make image generation a part of the flavours system so new progress images can be added easily

Bug Fixes:

• Allowed the library to build/install without examples or the plugin
• Add k to the make_dataset script
• Output images progress individually so it can work when the dataset has images of different sizes
• Ran clang-tidy over the codebase and made many fixes

Visual Mesh 2.0.1

This is a minor bugfix release

• Ran Include What You Use over the codebase and fixed a number of include related issues
• Properly disables OpenCL when it is disabled manually (rather than just not found)
• Manually find and set the preferred workgroup size in OpenCL
• Cached more of the memory creation operations in OpenCL to improve the performance further
• Fixed the Rectilinear projection model so that it does not project rays that are coming from behind the camera
• Dumped training history to YAML rather than a python pickle
• Changed the mesh lookup signature to match the VisualMesh class

Visual Mesh 2.0

The second major version of the visual mesh, based on Tensorflow 2.3

Visual Mesh 1.0

The original TensorFlow 1.14 compatible release