hannesdm / the-bures-metric-for-generative-adversarial-networks Goto Github PK

K11 = fake_phi @ tf.transpose(fake_phi)
K11 = K11 + epsilon * tf.eye(batch_size, dtype=dtype)
K22 = real_phi @ tf.transpose(real_phi)
K22 = K22 + epsilon * tf.eye(batch_size, dtype=dtype)

Hello :), recently I am trying to use the bures kernel loss which mentioned in your paper in my research, but I am confused about the useage of "epsilon * tf.eye(batch_size, dtype=dtype)" in your code, I suppose this is used to prevent the matrix from being invertible, or it could be used for something else?