New metrics optimizations, refactoring and more
New features
- Visual Information Fidelity (VIF) and corresponding loss are implemented (#47)
- Inception Score (IS) is implemented (#42)
- TVLoss is changed. Now it is based on only predicted image, as it's used as a regularizer applied to the predicted images during NN training (#58)
Enchancements
- Purely PyTorch-based computation of KID score (#27)
- Assertion of minimal image resolution for MS-SSIM (#38)
- Scipy-free computation of FID score (#32)
- Fast InceptionNet initialization (#45)
Other
- More strict formatting rules via workflow with flake8 checks
- CI pipeline corrections: now workflows are triggered for PRs and merges into master
- PR templates
- API description and examples
- Small fixes and refactoring