Shots refactor I of III

[mariaschuld]

Context:

The shots refactor is a long-planned change that makes the execution of circuits with varying numbers of shots more flexible.

The first change adds an optional kwarg to the qnode's __call__ function which temporarily changes the number of shots that the device is executed with.

Description of the Change:

• Temporarily set the number of shots in qnode's __call__ function.

• Add a test.

• Change docs and changelog.

Benefits:

In every execution one can now overwrite the device's default shot argument.

dev = qml.device('default.qubit', wires=1, shots=10)

@qml.qnode(dev)
def circuit(a):
    qml.RX(a, wires=0)
    return qml.sample(qml.PauliZ(wires=0))

print(circuit(0.8))  # [ 1  1  1 -1 -1  1  1  1  1  1]
print(circuit(0.8, shots=3))  # [ 1  1  1]
print(circuit(0.8))  # [ 1  1  1 -1 -1  1  1  1  1  1]

Possible Drawbacks:

If a user creates a qfunc with a custom shots keyword argument used in the circuit, the kwarg will be popped from the argument list and is never fed into the circuit. This leads to unexpected behaviour, such as

dev = qml.device('default.qubit', wires=2, shots=10)

@qml.qnode(dev)
def circuit(a, shots=1):
    qml.RX(a, wires=shots)
    return qml.sample(qml.PauliZ(wires=0))

print(circuit(0.8, shots=0)) 
>>> "pennylane._device.DeviceError: The specified number of shots needs to be at least 1. Got 0."

In some sense, we make shots a reserved argument.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 97.72%. Comparing base (5dc52a5) to head (4f780f0).
Report is 2639 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #1075   +/-   ##
=======================================
  Coverage   97.72%   97.72%           
=======================================
  Files         154      154           
  Lines       11647    11658   +11     
=======================================
+ Hits        11382    11393   +11     
  Misses        265      265           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[josh146]

Looks good @mariaschuld! My only major comment I think was regarding warning vs. exception, the rest were all minor docstring suggestions :)

[josh146]

I would make this a warning rather than an exception. I can't quite explain, but it 'feels' more like a warning.

Like, we shouldn't stop the user from creating the qnode, but we should warn them that the shots argument won't act as they expect?

[mariaschuld]

The reason why I changed to the error is that I played around and all sorts of strange things happened until I realised that I used a shots argument in my code and I had just overlooked the warning...

I think a warning is more appropriate but an exception safer? Happy to change though!

[mariaschuld]

@josh146 is this the test you had in mind?

[josh146]

Yep looks good!

[josh146]

Yep looks good!

[mariaschuld]

[ch4089]

[mariaschuld]

@josh146 and @Thenerdstation, thanks for the comments!

I made a few more changes which ensure backwards compatibility. So we have:

• If users have a shots argument or keyword argument in the qfunc, the per-call-shot-mutation is switched off and their argument gets precedence. However, they will get a deprecation warning that discourages the use of shots args/kwargs.
• If they do not have a shots arg/kwarg we interpret (for example) shots=8 as a temporary overwrite to the device's shots argument

[chaserileyroberts]

I can't suggest it, but you need 3 backticks here.