quantlab.model.predefined.master.PositionalEncoding

class quantlab.model.predefined.master.PositionalEncoding(d_model, max_len=100)

Bases: Module

Adds the fixed sinusoidal encoding of each step’s position, up to max_len steps.

Examples

>>> PositionalEncoding(d_model=4)(torch.zeros(2, 3, 4))[0, 1]
tensor([0.8415, 0.5403, 0.0100, 0.9999])

forward(x)

Return x ([S_t, N, D]) plus the encoding of its N steps.

Examples

>>> PositionalEncoding(d_model=4).forward(torch.zeros(1, 2, 4))[0, 0]
tensor([0., 1., 0., 1.])