Layer Adaptive Graph Neural Networks(LAGNN)

code for LAGNN

Environment Settings

• python == 3.8.10
• torch == 1.10.0

Parameter Settings

• middle_layer_num: number of hidden units
• value_layer_dropout: probility of layer_dropout (keep)
• temperature: temperature in gumbel softmax
• epoch: number of epochs to train the base model
• stage1_epoch: Number of epochs to pre-train
• seed: random seed
• lr: learning rate
• weight_decay: weight decay (L2 loss on parameters)
• hidden: embedding dimension
• dropout: dropout rate
• dataset: dataset cora, citeseer, pubmed, coauthor-cs, coauthor-phy, amazon-com
• epoch: number of epochs to train the base model
• stage1_epoch: Number of epochs to pre-train
• weight_choose_share: share choose weight between layer
• choose_weight_type: 0 for input, 1 for input and output, 2 for resnet, 3 for random dropout layer
• choose_weight_layernum: the number of weight layer
• linear_decay: whether probability decays linearly with the number of layers or not

Basic Usage

python ./LAGNN/train.py