Question about DGNN

[inst15]

Hi RManLuo, thanks for your nice work! I have a question about DGNN. How can I change DGNN model of decoder? I would greatly appreciate your response!

[RManLuo]

Hi, you can change the decoder at here:

GSNOP/code/layers.py

Lines 45 to 81 in aaf6059

	class EdgePredictor(torch.nn.Module):
	def __init__(self, dim_in, dim_out=1):
	super(EdgePredictor, self).__init__()
	self.dim_in = dim_in
	self.src_fc = torch.nn.Linear(dim_in, dim_in)
	self.dst_fc = torch.nn.Linear(dim_in, dim_in)
	self.out_fc = torch.nn.Linear(dim_in, 1)
	self.droup = torch.nn.Dropout(0.2)
	self.out = torch.nn.Sequential(
	torch.nn.Linear(dim_in * 2, dim_in),
	torch.nn.ReLU(),
	torch.nn.Dropout(0.2),
	torch.nn.Linear(dim_in, dim_in),
	torch.nn.ReLU(),
	torch.nn.Dropout(0.2),
	torch.nn.Linear(dim_in, dim_out),
	)
	def forward(self, h, neg_samples=1):
	num_edge = h.shape[0] // (neg_samples + 2)
	h_src = h[:num_edge]
	h_pos_dst = h[num_edge:2 * num_edge]
	h_neg_dst = h[2 * num_edge:]
	h_pos_edge = torch.cat([h_src, h_pos_dst], dim=-1)
	if len(h.shape) > 2:
	h_neg_edge = torch.cat([h_src.tile(neg_samples, 1, 1), h_neg_dst], dim=-1)
	else:
	h_neg_edge = torch.cat([h_src.tile(neg_samples, 1), h_neg_dst], dim=-1)
	return self.out(h_pos_edge), self.out(h_neg_edge)
	def predict(self, src_h, dst_h):
	h_src = self.src_fc(src_h).unsqueeze(1)
	h_dst = self.dst_fc(dst_h)
	h_edge = torch.nn.functional.relu(h_src + h_dst)
	return self.out_fc(h_edge)

[inst15]

Thank you for your answer! But sorry, I made a mistake. My actual question is how can I change DGNN model of encoder.