| 191 | |
| 192 | |
| 193 | class Mlp(nn.Module): |
| 194 | def __init__(self, in_features, hidden_features=None, |
| 195 | out_features=None, act_layer=nn.GELU, drop=0.): |
| 196 | super().__init__() |
| 197 | out_features = out_features or in_features |
| 198 | hidden_features = hidden_features or in_features |
| 199 | self.norm = nn.LayerNorm(in_features) |
| 200 | self.fc1 = nn.Linear(in_features, hidden_features) |
| 201 | self.fc2 = nn.Linear(hidden_features, out_features) |
| 202 | self.act = act_layer() |
| 203 | self.drop = nn.Dropout(drop) |
| 204 | |
| 205 | def forward(self, x): |
| 206 | x = self.norm(x) |
| 207 | |
| 208 | x = self.fc1(x) |
| 209 | x = self.act(x) |
| 210 | x = self.drop(x) |
| 211 | x = self.fc2(x) |
| 212 | x = self.drop(x) |
| 213 | return x |
| 214 | |
| 215 | |
| 216 | class Attention(torch.nn.Module): |