| 49 | |
| 50 | |
| 51 | class Mlp(nn.Module): |
| 52 | def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.): |
| 53 | super().__init__() |
| 54 | out_features = out_features or in_features |
| 55 | hidden_features = hidden_features or in_features |
| 56 | self.fc1 = nn.Linear(in_features, hidden_features) |
| 57 | self.act = act_layer() |
| 58 | self.fc2 = nn.Linear(hidden_features, out_features) |
| 59 | self.drop = nn.Dropout(drop) |
| 60 | |
| 61 | def forward(self, x): |
| 62 | x = self.fc1(x) |
| 63 | x = self.act(x) |
| 64 | # x = self.drop(x) |
| 65 | # commit this for the orignal BERT implement |
| 66 | x = self.fc2(x) |
| 67 | x = self.drop(x) |
| 68 | return x |
| 69 | |
| 70 | |
| 71 | class Attention(nn.Module): |