Sorted tape.get_parameters

[dwierichs]

Context:
The tape method get_parameters relies on tape.trainable_params (if trainable_only=True), which is a set.
This means that if for some reason the internal ordering of trainable_params is not preserved (which is not guaranteed in python, in contrast to Python 3.7+ dicts), get_parameters will return the parameters in an incorrect order. In particular, the classical_jacobian outputs a permutation if this happens.

Description of the Change:
Adding sorted to the indices retrieved from tape.trainable_params within tape.get_parameters

Benefits:
The order of the parameters is stable.

Possible Drawbacks:
For large tapes this incurs an overhead, but as trainable_params typically is "sorted" already, this overhead is very small:

a, b, c = set(range(int(1e3))), set(range(int(1e4))), set(range(int(1e5)))
sorted(a) # Takes ~0.3 millisecs
sorted(b) # Takes ~1 millisecs
sorted(c) # Takes ~8 millisecs

[josh146]

Thanks @dwierichs! Although, this seems to have broken three tests:

tests/transforms/test_metric_tensor.py::TestMetricTensor::test_evaluate_block_diag_metric_tensor[parameter-shift]
tests/transforms/test_metric_tensor.py::TestMetricTensor::test_evaluate_block_diag_metric_tensor[backprop]
qchem/tests/test_excitations_to_wires.py::test_integration_with_uccsd[weights0-singles0-doubles0-expected0]

Super strange 🤔

[codecov]

Codecov Report

❗ No coverage uploaded for pull request base (v0.19.0-rc0@f674439). Click here to learn what that means.
The diff coverage is n/a.

@@              Coverage Diff               @@
##             v0.19.0-rc0    #1836   +/-   ##
==============================================
  Coverage               ?   98.96%           
==============================================
  Files                  ?      218           
  Lines                  ?    16364           
  Branches               ?        0           
==============================================
  Hits                   ?    16195           
  Misses                 ?      169           
  Partials               ?        0           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update f674439...2fb3ce4. Read the comment docs.

[josh146]

Thanks @dwierichs! I suppose a more long term solution is to switch to returning a list for trainable_params.

This way we ideally won't need to sort (assuming trainable_params is always constructed by iterating over operations in order), but:

• it is not a guarantee if a developer constructs trainable_params without sorting
• we lose the ability to assume no repeated elements in the list

So I guess it's a tradeoff either way 🤔