Conversation
|
|
There was a problem hiding this comment.
num_params = num_orbitals**2 # antisymmetric, specified by upper triangleThe comment should be deleted now.
There was a problem hiding this comment.
Just modified the docstrings to reflect the changes in the parameterization.
A couple of changes relative to #130 are present:
1.- Use ADAM as the optimizer instead of gradient descent with momentum
2.- Removal of some redundant permutations of indices in the eri and 2rdms
There was a problem hiding this comment.
The comment I mentioned above is actually in the notebook.
There was a problem hiding this comment.
Oh, I just saw it. I removed the comment :)
|
@kevinsung , I have removed the implementation of configuration recovery that uses broadcasting for performance to include it in a separate PR :) |
| # ``1.0`` as the occupation deviates further from the expected ratio | ||
| if ratio_exp == 1.0: | ||
| return eps | ||
| return 1.0 |
There was a problem hiding this comment.
This change is unrelated, right? Please revert it. Why did you make it though? Are you fixing something?
| k_flat: 1D array defining the orbital transform, ``K``. The array should specify the upper | ||
| triangle of the anti-symmetric transform operator in row-major order, excluding the diagonal. | ||
| k_flat: 1D array defining the orbital transform, ``K``. The anti-symmetric | ||
| matrix defining the transfomration is construccted by the anti-symmetrization |
There was a problem hiding this comment.
spelling mistakes
- "transformation"
- "constructed"
3 participants
The aim of this PR is to provide a stable implementation of the orbital optimization functions in the
fermionmodule. The optimizer uses gradient descent with the ADAM update rule.Additionally, I have included a beta implementation of configuration recovery using numpy broadcasting in the
configuration_recoverymodule:recover_configurations_bcast(). If we see that this later changes are not needed I can remove them.