The order of uppooling and skip-connection seems wrong

[LingxiaoShawn]

Hi Hongyang, Based on your paper, I think the following code in ops.py has some ordering problem

   def forward(self, g, h):
        adj_ms = []
        indices_list = []
        down_outs = []
        hs = []
        org_h = h
        for i in range(self.l_n):
            h = self.down_gcns[i](g, h)
            adj_ms.append(g)
            down_outs.append(h)
            g, h, idx = self.pools[i](g, h)
            indices_list.append(idx)

        h = self.bottom_gcn(g, h)

        for i in range(self.l_n):
            up_idx = self.l_n - i - 1
            g, idx = adj_ms[up_idx], indices_list[up_idx]
            g, h = self.unpools[i](g, h, down_outs[up_idx], idx)
            h = self.up_gcns[i](g, h)
            h = h.add(down_outs[up_idx])
            hs.append(h)
        h = h.add(org_h)
        hs.append(h)
        return hs

In my understanding the correct order should be

        for i in range(self.l_n):
            up_idx = self.l_n - i - 1
            g, idx = adj_ms[up_idx], indices_list[up_idx]

            g, h = self.unpools[up_idx](g, h, down_outs[up_idx], idx)

            h = h.add(down_outs[up_idx])
            hs.append(h)

            h = self.up_gcns[up_idx](g, h)

        h = h.add(org_h)
        hs.append(h)
        return hs`

please let me know whether I'm correct. Thank you again for sharing the structured code!

[HongyangGao]

Yes. Actually you can do either way. Due to the fact that graph data are very easy to overfit, using addition after gcn may help to avoid this problem.