Improved functionalities
- Accuracies comparable with those claimed in the original paper (see below)
- Released pretrained models for both state-description and from-pixels
- Added features extraction routines
Accuracies
- state-description: 97.9%
- from-pixels: 93.6%