Dropping edges before message pass

[felipemello]

Hi, I wanted for each convolution to randomly ignore N edges. I guess that this would be equivalent to doing torch.nn.Dropout on the adjacency matrix. If i had to do it without DGL API, I would do something like below to drop half of the edges:

drop_edge_rate = 0.5
n_edges = len(h['edges'][0])
dropedge_filter = torch.randperm(n_edges, device = device)[:int(n_edges*drop_edge_rate)]

selected_eIDs_0 = h['edges'][0].index_select(dim = 0, index = dropedge_filter)
selected_eIDs_1 = h['edges'][1].index_select(dim = 0, index = dropedge_filter)

new_temporary_edges = torch.stack([selected_eIDs_0, selected_eIDs_1], dim = 1)

How can I drop the edges per convolution and still use DGL API’s like: g.update_all(fn.copy_u('x', 'm'), fn.sum('m', 'h')) or edge_softmax? Thanks!!

[mufeili]

Since 0.8, we now provide several transform modules that can be applied to built-in DGL datasets as described in the “Graph Dataset and Transforms” section here. The transform module DropEdge can meet your requirement.