| 191 | |
| 192 | |
| 193 | class PositionalEncoding(nn.Module): |
| 194 | |
| 195 | def __init__(self, d_model, dropout=0.1, max_len=5000): |
| 196 | super(PositionalEncoding, self).__init__() |
| 197 | self.dropout = nn.Dropout(p=dropout) |
| 198 | |
| 199 | pe = torch.zeros(max_len, d_model) |
| 200 | position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1) |
| 201 | div_term = torch.exp( |
| 202 | torch.arange(0, d_model, 2).float() * (-np.log(10000.0) / d_model)) |
| 203 | pe[:, 0::2] = torch.sin(position * div_term) |
| 204 | pe[:, 1::2] = torch.cos(position * div_term) |
| 205 | pe = pe.unsqueeze(0).transpose(0, 1) |
| 206 | |
| 207 | self.register_buffer('pe', pe) |
| 208 | |
| 209 | def forward(self, x): |
| 210 | # not used in the final model |
| 211 | x = x + self.pe[:x.shape[0], :] |
| 212 | return self.dropout(x) |
| 213 | |
| 214 | |
| 215 | class TimestepEmbedder(nn.Module): |