benchmarks/experimental/experimental_async_approaches.py [63:78]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
class PositionalEncodingLayer(nn.Module):
    def __init__(self, d_model, dropout=0.1, max_len=5000):
        super(PositionalEncodingLayer, self).__init__()
        self.dropout = nn.Dropout(p=dropout)

        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = pe.unsqueeze(0).transpose(0, 1)
        self.register_buffer("pe", pe)

    def forward(self, x):
        x = x + self.pe[: x.size(0), :]
        return self.dropout(x)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



benchmarks/models/transformer_lm.py [28:45]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
class PositionalEncodingLayer(nn.Module):
    """PositionalEncoding layer for a given Transformer model."""

    def __init__(self, d_model, dropout=0.1, max_len=5000):
        super(PositionalEncodingLayer, self).__init__()
        self.dropout = nn.Dropout(p=dropout)

        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = pe.unsqueeze(0).transpose(0, 1)
        self.register_buffer("pe", pe)

    def forward(self, x):
        x = x + self.pe[: x.size(0), :]
        return self.dropout(x)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -