questions about input layer (model summary)

Hi, I’m new to graph cnn and doing some stellargraph tutorial. And I ran the GraphSAGE Cora Node Classification Example, graphsage-cora-example.py.
This is the part of the model summary.

__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_2 (InputLayer)            [(None, 20, 1433)]   0                                            
__________________________________________________________________________________________________
input_3 (InputLayer)            [(None, 200, 1433)]  0                                            
__________________________________________________________________________________________________
input_1 (InputLayer)            [(None, 1, 1433)]    0                                            
__________________________________________________________________________________________________

I have two questions.

1. Why are there multiple input layers?
2. What are these numbers of output shape indicates? I know the number 1433 is come from unique words of cora dataset.(right?)

I read the original paper of GraphSAGE, but still I don’t understand…

[youph]

@mimisen-boop yes, there can be an arbitrary number of layers in the model. However, adding more layers means increasing the “expressive power” of the resulting model (as each layer comes with its own learnable weights), and might lead to overfitting and worse generalisation power - e.g., see Figure 5 in the GCN paper. In our demos, we just chose 2-layer models as “good enough” models for demo purposes. But in general you are right - the number of layers, dimensionality of their outputs, etc. are model hyperparameters, and should be tuned to the dataset and the problem being solved, using proper hyperparameter tuning protocols.

[youph]

@mimisen-boop to change the number of layers in that example, you’d need to change the --neighbour_samples and --layer_size arguments. E.g., to have a 3-layer model, one can set them to something like --layer_size 20 20 20 and --neighbour_samples 20 10 5 , or whatever (these values are hyperparameters too). The notebook demos/node-classification/graphsage/graphsage-cora-node-classification-example.ipynb might be clearer on that.