| 38 | |
| 39 | |
| 40 | class VideoEncoder(nn.Module): |
| 41 | def __init__(self,emb_dim,fea_dim): |
| 42 | super(VideoEncoder, self).__init__() |
| 43 | |
| 44 | self.emb_dim = emb_dim |
| 45 | self.linear1 = torch.nn.Linear(self.emb_dim, self.emb_dim, bias=False) |
| 46 | self.linear2 = nn.Sequential(torch.nn.Linear(self.emb_dim, fea_dim),torch.nn.ReLU()) |
| 47 | |
| 48 | def forward(self, input_thumb, input_L): |
| 49 | input_ALL = torch.cat((input_L, input_thumb),1) #（bs,len+1,4096） |
| 50 | fea_A = torch.bmm(input_thumb,self.linear1(input_ALL).permute(0,2,1)) # (bs, 1, len+1) |
| 51 | fea_alpha = F.softmax(fea_A) # (bs, 1, len+1) |
| 52 | fea_V = torch.matmul(fea_alpha,input_ALL).squeeze() # (bs, 4096) |
| 53 | fea = self.linear2(fea_V) |
| 54 | return fea |
| 55 | |
| 56 | class ReverseLayerF(Function): |
| 57 | #@staticmethod |