Closed abieler closed 2 years ago
IDK if its worth introducing a new abstact type for the two GAT layers as some of the functions are identical.
Could replicate the results from Figure 1 in the paper regarding the attention scores:
GATConv Layer
GATv2 Layer
note the difference is that here we have self-loops required, which results in the 1st alpha score not being 0 for all queries.
Would you mind adding some tests?
@abieler Thank you for contribution. I will merge this first then fix the tests.
IDK if its worth introducing a new abstact type for the two GAT layers as some of the functions are identical.
Could replicate the results from Figure 1 in the paper regarding the attention scores:
GATConv Layer![gat-attention-scores](https://user-images.githubusercontent.com/7083141/147884965-407cfc08-d8d2-4f63-8675-7b7a701aab2c.png)
GATv2 Layer![gatv2-attention-scores](https://user-images.githubusercontent.com/7083141/147884984-3efe928e-91aa-494a-8462-a3dbbdcd787e.png)
note the difference is that here we have self-loops required, which results in the 1st alpha score not being 0 for all queries.