Closed ghost closed 1 year ago
For example, in A3TGCN, should the forward function be written like this if H is None: H = torch.zeros(X.shape[0], X.shape[1]).to(X.device) H_accum = 0 probs = torch.nn.functional.softmax(self._attention, dim=0) for period in range(self.periods): H = self._base_tgcn(X[:, :, period], edge_index, edge_weight, H) H_accum = H_accum + probs[period] * H return H_accum
Dear author, your DCRNN model uses GRU in the time dimension, but it seems that during the training process of the model, the Hidden state matrix is 0 in each time dimension. The Hidden state matrix of a time dimension is then input into the forward function of the model to implement feature propagation in the time dimension. In other words, your current forward function does not pass in the Hidden state matrix, so is there no feature propagation in the time dimension?