dynamic.py 文件源码

def forward(self, x, lengths):
        batch_size = len(x)
        lengths = [len(s) for s in x]

        outputs = [Variable(torch.zeros(1, self.model_dim).float(), volatile=not self.training)
                   for _ in range(batch_size)]

        for t in range(max(lengths)):
            batch = []
            h = []
            idx = []
            for i, (s, l) in enumerate(zip(x, lengths)):
                if l >= max(lengths) - t:
                    batch.append(s.pop())
                    h.append(outputs[i])
                    idx.append(i)

            batch = np.concatenate(np.array(batch).reshape(-1, 1), 0)
            emb = Variable(torch.from_numpy(self.initial_embeddings.take(batch, 0)), volatile=not self.training)
            h = torch.cat(h, 0)
            h_next = self.rnn(emb, h)
            h_next = torch.chunk(h_next, len(idx))

            for i, o in zip(idx, h_next):
                outputs[i] = o

        outputs = torch.cat(outputs, 0)
        h = F.relu(self.l0(F.dropout(outputs, 0.5, self.training)))
        h = F.relu(self.l1(F.dropout(h, 0.5, self.training)))
        y = F.log_softmax(h)
        return y