Where in the code are the natural gradients for the convolutional layers renormalized by _renorm_coeff (num_locations)?
Is this handled by the tensorflow optimizer class after the inverse covariances are applied to the gradients? Otherwise, I can't find this renormalization in the compute_cov / compute_inv / apply gradients functions.
It might be a bug, though this doesn't seem like the right conclusion given the effect this would have on performance.
Where in the code are the natural gradients for the convolutional layers renormalized by _renorm_coeff (num_locations)?
Is this handled by the tensorflow optimizer class after the inverse covariances are applied to the gradients? Otherwise, I can't find this renormalization in the compute_cov / compute_inv / apply gradients functions.
It might be a bug, though this doesn't seem like the right conclusion given the effect this would have on performance.
Thanks,
Max