Spaces:
Sleeping
Sleeping
import torch | |
import torch.nn as nn | |
import math | |
# PositionalEncoding | |
# Taken from https://pytorch.org/tutorials/beginner/transformer_tutorial.html | |
class PositionalEncoding(nn.Module): | |
def __init__(self, d_model, dropout=0.1, max_len=5000): | |
super(PositionalEncoding, 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) | |