Spaces:
Running
on
T4
Running
on
T4
File size: 1,367 Bytes
4275cae |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 |
"""
Various positional encodings for the transformer.
"""
import math
import torch
from torch import nn
def PE1d_sincos(seq_length, dim):
"""
:param d_model: dimension of the model
:param length: length of positions
:return: length*d_model position matrix
"""
if dim % 2 != 0:
raise ValueError("Cannot use sin/cos positional encoding with "
"odd dim (got dim={:d})".format(dim))
pe = torch.zeros(seq_length, dim)
position = torch.arange(0, seq_length).unsqueeze(1)
div_term = torch.exp((torch.arange(0, dim, 2, dtype=torch.float) *
-(math.log(10000.0) / dim)))
pe[:, 0::2] = torch.sin(position.float() * div_term)
pe[:, 1::2] = torch.cos(position.float() * div_term)
return pe.unsqueeze(1)
class PositionEmbedding(nn.Module):
"""
Absolute pos embedding (standard), learned.
"""
def __init__(self, seq_length, dim, dropout, grad=False):
super().__init__()
self.embed = nn.Parameter(data=PE1d_sincos(seq_length, dim), requires_grad=grad)
self.dropout = nn.Dropout(p=dropout)
def forward(self, x):
# x.shape: bs, seq_len, feat_dim
l = x.shape[1]
x = x.permute(1, 0, 2) + self.embed[:l].expand(x.permute(1, 0, 2).shape)
x = self.dropout(x.permute(1, 0, 2))
return x
|