View Code? Open in Web Editor
NEW
The Tensorflow Implementation of paper Self-supervised graph convolutional clustering by preserving latent distribution
sgc-pld's Introduction
🌱 I'm currently pursuing the Ph.D degree at the Xidian University.
🔭 Multi-modal (view) representation learning, image processing, causal learning, self-supervised learning, unsupervised learning, etc.
🎈 From 2022-6 to 2022-12, I interned as an AI Engineer at Trustworthiness Theory Research Center, Huawei Technologies Co., Ltd, I was fortunately directed by Dr. Wei Zhou . During this period, I also work with Dr. Minglei Li in Huawei Cloud BU about 3D Hyper-realistic Digital Human Technology.
Email
sgc-pld's People
Contributors
Watchers
sgc-pld's Issues
在Graph_Attention_Encoder.py中,第34行:
self.z = self.__dense_layer(self.H)
self.XZ = self._dense_layer_Z(self.H)
输入怎么都是self.H, 自己和自己进行度量?
您好,我尝试按照论文中的描述,调整模型的参数值:
对应到代码中这一部分:
# Total lossA
self .loss = (features_loss + self .lambda_ * structure_loss ) + 0.1 * self .XZ_loss + 5 * self .C_loss + 1 * self .dense_loss
# self.loss = features_loss + self.lambda_ * structure_loss + 10 * self.C_loss + self.dense_loss # without pd 0.678
# self.loss = features_loss + self.lambda_ * structure_loss + 10 * self.C_loss + 100 * self.XZ_loss # without ss 0.64
# self.loss = features_loss + self.lambda_ * structure_loss + 100 * self.XZ_loss + self.dense_loss # without clustering loss
return self .loss , self .H , self .C , self .pred , self .dense_loss , self .z , features_loss , structure_loss #, self.C_loss, self.XZ_loss
最终在Cora数据集上得到如下结果:
Epoch--995: time: 501.838:
acc: 0.67872969 nmi: 0.51988349 f_score: 0.60493104 ari: 0.45633046
Epoch--996: time: 502.479:
acc: 0.67909897 nmi: 0.51989684 f_score: 0.60486178 ari: 0.45696338
Epoch--997: time: 503.102:
acc: 0.67836041 nmi: 0.51759465 f_score: 0.60393833 ari: 0.45571426
Epoch--998: time: 503.696:
acc: 0.67651403 nmi: 0.51489665 f_score: 0.60190142 ari: 0.45244131
Epoch--999: time: 504.392:
acc: 0.67762186 nmi: 0.51728817 f_score: 0.60408646 ari: 0.45316671
而预期的准确率能达到 0.76。请问我是否还应该修改代码的其他部分?另外如果可以的话,请您提供一下Citesser 和 Pubmed 的参数配置!非常感谢!