Method __init__

OSUM/wenet/transformer/embedding.py:203–216

(self,
                 d_model: int,
                 head_dim: int,
                 dropout_rate: float,
                 max_len: int = 1500,
                 rope_theta=10000.0,
                 scale: bool = True)

class RopePositionalEncoding(PositionalEncoding):

    def __init__(self,
                 d_model: int,
                 head_dim: int,
                 dropout_rate: float,
                 max_len: int = 1500,
                 rope_theta=10000.0,
                 scale: bool = True):
        super().__init__(d_model, dropout_rate=dropout_rate, max_len=max_len)
        delattr(self, 'pe')
        self.max_len = max_len * 2
        pe = precompute_freqs_cis(head_dim, self.max_len, rope_theta)
        self.register_buffer("pe", torch.view_as_real(pe.unsqueeze(0)))
        self.dropout_rate = dropout_rate
        self.scale = scale

    def forward(
        self,