Supporting counts from raw samples

[theodotk]

Context:

Fixing #2562

Description of the Change:

Adding the possibility to group samples into counts providing counts=True flag in qml.sample function. This required creation of a new measurement type Counts in measurements.py.

The implementation is based on editing the sample function in _qubit_device.py, though it was possible to base it on estimate_probability function (or to create something new that seems superfluous to me). Choice of editing the sample function seemed more natural for this task, but please tell me if that's not the best approach.

Examples

From the issue:

dev = qml.device('default.qubit', wires=3, shots=10)

@qml.qnode(dev)
def my_circ():
    qml.Hadamard(wires=0)
    qml.CNOT(wires=[0,1])
    qml.PauliX(wires=2) 

    return qml.sample(counts=True)```

```>>> my_counts = my_circ()
>>> print(my_counts)
{'001': 6, '111': 4}

It works also for observables:

@qml.qnode(dev)
def circuit():
    qml.Hadamard(wires=0)
    qml.CNOT(wires=[0, 1])
    return qml.sample(qml.PauliZ(0), counts=True), qml.sample(qml.PauliZ(1), counts=True)```

```>>>result = circuit()
>>>print(result)
[tensor({-1: 472, 1: 528}, dtype=object, requires_grad=True)
 tensor({-1: 472, 1: 528}, dtype=object, requires_grad=True)]

Benefits:

Better visibility of the sampling result.

Possible Drawbacks:

Shouldn't be, as the change scope was minimal.

Related GitHub Issues:
2562

[codecov]

Codecov Report

Merging #2686 (0fb203b) into master (38ca383) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master    #2686   +/-   ##
=======================================
  Coverage   99.61%   99.61%           
=======================================
  Files         255      255           
  Lines       20884    20913   +29     
=======================================
+ Hits        20804    20833   +29     
  Misses         80       80           

Impacted Files	Coverage Δ
pennylane/_qubit_device.py	99.18% <100.00%> (+0.04%)	⬆️
pennylane/interfaces/autograd.py	100.00% <100.00%> (ø)
pennylane/measurements.py	99.62% <100.00%> (+<0.01%)	⬆️
pennylane/qnode.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 38ca383...0fb203b. Read the comment docs.

[theodotk]

Hey @josh146, can you please re-launch the tests?
I'm also not sure how to interpret code coverage test results and if I can improve it.
Thanks

[theodotk]

Thanks @josh146 ! Sorry, I didn't think of checking the tests as well. Can you please relaunch?
Thanks again!

[theodotk]

Thanks, @josh146 !
Do you know whom should I ask for a review? The author of the original issue, or there is someone responsible for the measurement/qubit device part of the code?

[josh146]

Thanks @theodotk! Someone from the team will have a look shortly :)

[theodotk]

Hey @Jaybsoni @antalszava could you please review the PR?
It's my test assignment for the job interview hence the time limit, and I mention both of you because you are in the initial issue discussion.
Thanks!

[Jaybsoni]

Hey @theodotk,

Thanks for the contribution, I will take a look at it today! I think @antalszava will take a closer look at it at some point once he is back in the office as he is overlooking the interview process. Thanks for you patience.

Cheers,

[antalszava]

Hi @theodotk, it seems that changes made in the Autograd interface would be required for other interfaces too because they too cannot interpret dictionaries (e.g., trying to set interface="jax" fails on a jnp.array conversion).

If it doesn't seem to come around, no worries, I could help here tomorrow or latest next week. 👍

[theodotk]

Hi @antalszava ! I fixed the problem with jax, and both tensorflow and torch return expected result with the new version.

The problem arose because jax does not handle non-numeric types such as string. And in case of state vector sampling, keys of the counts dictionary are strings (because str is hashable and because converting binary to decimal instead is a less user-friendly option). But the dictionary itself seems to be handled okay.
So the code became slightly uglier 😢 but it works.

Just in case, I chose interface by setting qml.device("default.qubit.jax") (replacing "jax" by "tf" or "torch" for the others). Tell me if it isn't the right way to test.

[antalszava]

Hi @theodotk, that's awesome! 🙂

It would be nice then to have a simple example for each interface, e.g.,

import pennylane as qml

dev = qml.device('default.qubit', wires=3, shots=10)

@qml.qnode(dev, inteface="jax")
def circuit():
    return qml.sample(counts=True)

circuit()

and then this PR should be looking great! 🌟

[theodotk]

@antalszava you mean in the docs?

[antalszava]

@theodotk apologies! I meant such examples as test cases so that we can be sure things are and stay okay. 👍

[theodotk]

@antalszava done!

[antalszava]

Looks good @theodotk! 🥇 Great one with getting each detail right here, thanks for addressing the comments! 🙂

Made a couple of small changes to test docstrings and names, but will merge this soon.

[theodotk]

Thanks @antalszava , it is pleasant to see such a diligent review process!