This is the initial release of Fast Transformer and implements Fast Transformer as a subclassed TensorFlow model.
Classes
- `InvariantPointAttention`: Invariant Point Attention which was used in the structure module of Alphafold2 from the paper Highly accurate protein structure prediction with AlphaFold for coordinate refinement. Invariant Point Attention is a form of attention that acts on a set of frames and is invariant under global Euclidean transformations on said frames
- `IPABlock`: Invariant Point Attention Block which is an IPA followed by a feedforward and has normalization layers
- `IPATransformer`: IPA Based transformer which is a stack of `IPABlock` and feedforward layers:
Utility Methods
- `quaternion_raw_multiply`: multiply two quaternions
- `standardize_quaternion`: convert a unit quaternion to a standard form: one in which the real part is non-negative
- `quaternion_multiply`: multiply two quaternions representing rotations, returning the quaternion representing their composition, i.e. the versor with a nonnegative real part.
- `quaternion_to_matrix`: Convert rotations given as quaternions to rotation matrices.