The unignn from oneforward

UniGAT: inconsistency between paper formulations and code implementation

In your paper, the UniGAT computes the similarity alpha_ie between node i and the incident edge e. Then the attention is applied to passing messages from edges to nodes based on the alpha_ie.

However, the implementation seems to be different from the UniGAT in your paper. Only edge information is used for computing alpha_ie. Node information is not included. self.att_e is used instead of node information. Why?

UniGNN/model/UniGNN.py

Lines 246 to 282 in bea1fb5

 def forward(self, X, vertex, edges): 

 H, C, N = self.heads, self.out_channels, X.shape[0] 

 # X0 = X # NOTE: reserved for skip connection 

 X0 = self.W(X) 

 X = X0.view(N, H, C) 

 Xve = X[vertex] # [nnz, H, C] 

 Xe = scatter(Xve, edges, dim=0, reduce=self.args.first_aggregate) # [E, H, C] 

 alpha_e = (Xe * self.att_e).sum(-1) # [E, H, 1] 

 a_ev = alpha_e[edges] 

 alpha = a_ev # Recommed to use this 

 alpha = self.leaky_relu(alpha) 

 alpha = softmax(alpha, vertex, num_nodes=N) 

 alpha = self.attn_drop( alpha ) 

 alpha = alpha.unsqueeze(-1) 

 Xev = Xe[edges] # [nnz, H, C] 

 Xev = Xev * alpha 

 Xv = scatter(Xev, vertex, dim=0, reduce='sum', dim_size=N) # [N, H, C] 

 X = Xv 

 X = X.view(N, H * C) 

 if self.args.use_norm: 

 X = normalize_l2(X) 

 if self.skip_sum: 

 X = X + X0 

 # NOTE: concat heads or mean heads? 

 # NOTE: skip concat here? 

 return X

a problem in train.py "from data import data" in line 57, which package is used in the work?

I am very interested in this work and the method for hypergraph-based gnn is of great attraction to me.
However, there exists a minor issue, and I am confused about it.

After I read this paper, I clone the code and ran it in bash with 'python train.py --data=coauthorship --dataset=dblp --model-name=UniSAGE' , but the error ' ModuleNotFoundError: No module named 'data'' occurs.
I tried 'pip install data', but the package so-called 'data' is different from the one in this work, which has not 'data'/'data.load'......
So, I would like to know the source of the 'data' package, it seems like an individual-created .py file?
train.py "from data import data" in line 57, which package is used in work?

I'm looking forward to your answer, thank you！
Best wishes!

oneforward / unignn Goto Github PK

unignn's People

Contributors

Stargazers

Watchers

Forkers

unignn's Issues

UniGAT: inconsistency between paper formulations and code implementation

a problem in train.py "from data import data" in line 57, which package is used in the work?

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent

	def forward(self, X, vertex, edges):
	H, C, N = self.heads, self.out_channels, X.shape[0]

	# X0 = X # NOTE: reserved for skip connection

	X0 = self.W(X)
	X = X0.view(N, H, C)

	Xve = X[vertex] # [nnz, H, C]
	Xe = scatter(Xve, edges, dim=0, reduce=self.args.first_aggregate) # [E, H, C]


	alpha_e = (Xe * self.att_e).sum(-1) # [E, H, 1]
	a_ev = alpha_e[edges]
	alpha = a_ev # Recommed to use this
	alpha = self.leaky_relu(alpha)
	alpha = softmax(alpha, vertex, num_nodes=N)
	alpha = self.attn_drop( alpha )
	alpha = alpha.unsqueeze(-1)


	Xev = Xe[edges] # [nnz, H, C]
	Xev = Xev * alpha
	Xv = scatter(Xev, vertex, dim=0, reduce='sum', dim_size=N) # [N, H, C]
	X = Xv
	X = X.view(N, H * C)

	if self.args.use_norm:
	X = normalize_l2(X)

	if self.skip_sum:
	X = X + X0

	# NOTE: concat heads or mean heads?
	# NOTE: skip concat here?

	return X