A trial at implementing the paper paper not an official implementation

Todo:

• Test grid_sample
• Increase Batch size -> increase training time (need recheck)
• Make tools/model.py more readable
• Train the entire Unsupervised Network together instead of train GazeRepresentationLearning separately
• Change Angle Encoder
• Add warping field regularization
• Recheck division by batch size in feature loss