Add parameter batch support to the AmplitudeEmbedding template

[dime10]

Context: In PR #1710, support for creating PennyLane QNodes with batches of input parameters via @qml.batch_params was added. However, not all templates have support for this feature yet.

Description of the Change:

• This modifies the AmplitudeEmbedding template to support a batch of parameters
  for the 'features' input. The pre-processing on the inputs is then applied for each
  feature_set in the batch.
• It adds the AmplitudeEmbedding template to the simple circuit test of the
  test_batch_params.py file. Also removes a test in test_amplitude.py that is no longer
  applicable.
• Fixes the wording in one the error messages of the template (unrelated to issue).

Benefits: It's possible to use batched parameters on the AmplitudeEmbedding template.

Possible Drawbacks: Less robust input checks.

Related GitHub Issues: Closes #1745

[dime10]

@josh146 An existing test, test_if_features_wrong_shape, in test_amplitude.py fails since it specifically checks that an error is thrown if features has more than one dimension. You mentioned adding tests to test_batch_params.py, but we'd also need to either remove this test, or have it check for >2 dimensions.

The other thing I was considering is that since the template is not aware of whether the @qml.batch_params decorator is actually present, we would not catch a situation where features is 2 dimensional without batching, which would previously throw an error. For example, this circuit should throw an error but now won't:

@qml.beta.qnode(dev)
def circuit(data, weights):
    qml.templates.AmplitudeEmbedding([[1,1,1,1], [2,2,2,2]], wires=[0, 1])

[josh146]

Thanks @dime10!

test_if_features_wrong_shape, in test_amplitude.py fails since it specifically checks that an error is thrown if features has more than one dimension. We'd also need to either remove this test, or have it check for >2 dimensions.

Yes, this is a very good catch. I ran into the same issue recently updating the StronglyEntanglingLayer to support batches of parameters, and I'm not sure I can think of a good solution. As you say, since the operation does not know ahead of time whether it is in a batched mode or not, it doesn't seem possible to validate this any more.

For now, I suggest removing this validation (and the corresponding test_if_features_wrong_shape).

For example, this circuit should throw an error but now won't:

Yes, this one deserves more thought. No need to worry about it now for this PR, but something we should consider and fix in a subsequent PR 🙂

We are currently working on rethinking the Operator class (tagging @mariaschuld for visibility), and potentially one solution would be to make the Operator class aware of potential batch dimensions ahead of time, e.g.,

@qml.beta.qnode(dev)
def circuit(data, weights):
    qml.templates.AmplitudeEmbedding([[1,1,1,1], [2,2,2,2]], wires=[0, 1], batch_axes=0)

In the meantime @dime10, let me know if you have any more questions, or otherwise, if the code is ready to be reviewed!

[codecov]

Codecov Report

Merging #1761 (7350486) into master (1bca151) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master    #1761   +/-   ##
=======================================
  Coverage   98.90%   98.90%           
=======================================
  Files         206      206           
  Lines       15385    15388    +3     
=======================================
+ Hits        15216    15219    +3     
  Misses        169      169           

Impacted Files	Coverage Δ
pennylane/templates/embeddings/amplitude.py	100.00% <100.00%> (ø)

---

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 1bca151...7350486. Read the comment docs.

[dime10]

@josh146 This is ready for review. Thank you for answering my questions!

[josh146]

Thanks @dime10! I will have a review out before the end of the day 🙂

[josh146]

Great work @dime10, very nice as well to see that you have kept in the length normalization! I liked your solution re: iterating over the batch dimension - I suppose this works since we are under the assumption that there is only one batch dimension present.

I think this is good to go on my end, just two final changes needed:

• From my reading, I think the test test_throws_exception_if_features_wrong_shape can be 'undeleted'.
• Don't forget to add your name, and a brief description of the change here, to the changelog! The changelog is in doc/changelog-dev.md.

[josh146]

@dime10 this looks great, happy to approve and merge it in!

Just before I do, don't forget to add a link to this PR in the changelog (it can be added to this existing entry

pennylane/doc/releases/changelog-dev.md

Line 19 in 1bca151

[(#1710)](https://github.com/PennyLaneAI/pennylane/pull/1710)

), and to add your name to the contributors section 🙂