Method forward

models.py:119–158

(self, occ, prc)

        self.adj2 = adj2

    def forward(self, occ, prc):  # occ.shape = [batch,node, seq]
        b, n, s = occ.shape
        data = torch.stack([occ, prc], dim=3).reshape(b*n, s, -1).unsqueeze(1)
        data = self.conv2d(data)
        data = data.squeeze().reshape(b, n, -1)

        # first layer
        atts_mat = self.gat_lyr(data)  # attention matrix, dense(nodes, nodes)
        occ_conv1 = torch.matmul(atts_mat, data)  # (b, n, s)
        occ_conv1 = self.dropout(self.LeakyReLU(self.gcn(occ_conv1)))

        # second layer
        atts_mat2 = self.gat_lyr(occ_conv1)  # attention matrix, dense(nodes, nodes)
        occ_conv2 = torch.matmul(atts_mat2, occ_conv1)  # (b, n, s)
        occ_conv2 = self.dropout(self.LeakyReLU(self.gcn(occ_conv2)))

        occ_conv1 = (1 - self.alpha) * occ_conv1 + self.alpha * data
        occ_conv2 = (1 - self.alpha) * occ_conv2 + self.alpha * occ_conv1
        occ_conv1 = occ_conv1.view(b * n, self.seq)
        occ_conv2 = occ_conv2.view(b * n, self.seq)

        x = torch.stack([occ_conv1, occ_conv2], dim=2)  # best
        lstm_out, (_, _) = self.lstm(x)  # b*n, s, 2

        # TPA
        ht = lstm_out[:, -1, :]  # ht
        hw = lstm_out[:, :-1, :]  # from h(t-1) to h1
        hw = torch.transpose(hw, 1, 2)
        Hc = self.fc1(hw)
        Hn = self.fc2(Hc)
        ht = torch.unsqueeze(ht, dim=2)
        a = torch.bmm(Hn, ht)
        a = torch.sigmoid(a)
        a = torch.transpose(a, 1, 2)
        vt = torch.matmul(a, Hc)
        ht = torch.transpose(ht, 1, 2)
        hx = torch.cat((vt, ht), dim=2)
        y = self.fc3(hx)
        y = y.view(b, n)
        return y