145170ca9bbd89aa01d8a40841e3c039d3683af8,stellargraph/layer/graph_attention.py,GraphAttention,call,#GraphAttention#Any#,211

Before Change

        // Convert A to dense tensor - needed for the mask to work
        // TODO: replace this dense implementation of GraphAttention layer with a sparse implementation
        if K.is_sparse(A):
            A = tf.sparse_tensor_to_dense(A, validate_indices=False)

        // For the GAT model to match that in the paper, we need to ensure that the graph has self-loops,
        // since the neighbourhood of node i in eq. (4) includes node i itself.
        // Adding self-loops to A via setting the diagonal elements of A to 1.0:
        if kwargs.get("add_self_loops", False):
            // get the number of nodes from inputs[1] directly
            N = K.int_shape(inputs[1])[-1]
            if N is not None:
                // create self-loops
                A = tf.linalg.set_diag(A, K.cast(np.ones((N,)), dtype="float"))
            else:
                raise ValueError(
                    "{}: need to know number of nodes to add self-loops; obtained None instead".format(
                        type(self).__name__
                    )
                )

        outputs = []
        for head in range(self.attn_heads):
            kernel = self.kernels[head]  // W in the paper (F x F")
            attention_kernel = self.attn_kernels[

After Change

            dense += mask

            // Apply softmax to get attention coefficients
            dense = K.softmax(dense, axis=1)  // (N x N), Eq. 3 of the paper

            // Apply dropout to features and attention coefficients
            dropout_feat = Dropout(self.in_dropout_rate)(features)  // (N x F")

Instances