underdoc-wang/mpgcn

[ICDE'20] Predicting Origin-Destination Flow via Multi-Perspective Graph Convolutional Network (Pytorch Replication)

mpgcn's Issues

Problems with data dimensions

Hello, the author. Excuse me. I want to ask about Data_ Container_ OD. Load in py_ What do the dimensions read from data represent? Why do I need to use reshape (- 1,47,47)?

About dataset

Hello, dear author, can you provide me with Dataset?

[about model arch] there's something different from the original paper?

Hi, author.

1. In the original paper, there's three grpahs: adj, poi, corre. while your replica has only two grpahs.

Your replica:

MPGCN(
  (branch_models): ModuleList(
    (0): ModuleDict(
      (temporal): LSTM(1, 32, batch_first=True)
      (spatial): ModuleList(
        (0): BDGCN(
          (activation): ReLU()
        )
        (1): BDGCN(
          (activation): ReLU()
        )
        (2): BDGCN(
          (activation): ReLU()
        )
      )
      (fc): Sequential(
        (0): Linear(in_features=32, out_features=1, bias=True)
        (1): ReLU()
      )
    )
    (1): ModuleDict(
      (temporal): LSTM(1, 32, batch_first=True)
      (spatial): ModuleList(
        (0): BDGCN(
          (activation): ReLU()
        )
        (1): BDGCN(
          (activation): ReLU()
        )
        (2): BDGCN(
          (activation): ReLU()
        )
      )
      (fc): Sequential(
        (0): Linear(in_features=32, out_features=1, bias=True)
        (1): ReLU()
      )
    )
  )
)

2. In the model input, what do the O_dyn_G and D_dyn_G means? it seems that you treat the origin and destination individually?

    def construct_dyn_G(self, OD_data: np.array,
                        perceived_period: int = 7):  # construct dynamic graphs based on OD history
        train_len = int(OD_data.shape[0] * self.params['split_ratio'][0] / sum(self.params['split_ratio']))
        num_periods_in_history = train_len // perceived_period  # dump the remainder
        OD_history = OD_data[:num_periods_in_history * perceived_period, :, :, :]

        O_dyn_G, D_dyn_G = [], []
        for t in range(perceived_period):
            OD_t_avg = np.mean(OD_history[t::perceived_period, :, :, :], axis=0).squeeze(axis=-1)
            O, D = OD_t_avg.shape

            O_G_t = np.zeros((O, O))  # initialize O graph at t
            for i in range(O):
                for j in range(O):
                    O_G_t[i, j] = distance.cosine(OD_t_avg[i, :], OD_t_avg[j, :])  # eq (6)
            D_G_t = np.zeros((D, D))  # initialize D graph at t
            for i in range(D):
                for j in range(D):
                    D_G_t[i, j] = distance.cosine(OD_t_avg[:, i], OD_t_avg[j, :])  # eq (7)
            O_dyn_G.append(O_G_t), D_dyn_G.append(D_G_t)

        return np.stack(O_dyn_G, axis=-1), np.stack(D_dyn_G, axis=-1)