Autodifferentiation support for PytorchBackend

[Simone-Bordoni]

Complete pytorch backend, close issue #1266

Checklist:

• Reviewers confirm new code works as expected.
• Tests are passing.
• Coverage does not decrease.
• Documentation is updated.

[Simone-Bordoni]

The only differentiation method that was not working for the pytorch backend was the parameter shift rule.
However now i have unmuted the test and removed the check on the backend and everything is working correctly for the pytorch backend. Probably the problem was linked to some error in the initial implementation of the pytorch backend and was fixed by correcting all the other tests.

[codecov]

Codecov Report

Attention: Patch coverage is 98.24561% with 5 lines in your changes are missing coverage. Please review.

Project coverage is 99.75%. Comparing base (363a6e5) to head (c95a9ef).

❗ Current head c95a9ef differs from pull request most recent head abd9382

Please upload reports for the commit abd9382 to get more accurate results.

Files	Patch %	Lines
src/qibo/backends/pytorch.py	88.23%	2 Missing ⚠️
src/qibo/gates/gates.py	85.71%	2 Missing ⚠️
...qibo/quantum_info/superoperator_transformations.py	96.87%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1276      +/-   ##
==========================================
- Coverage   99.82%   99.75%   -0.07%     
==========================================
  Files          72       72              
  Lines       10560    10645      +85     
==========================================
+ Hits        10541    10619      +78     
- Misses         19       26       +7     

Flag	Coverage Δ
unittests	99.75% <98.24%> (-0.07%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

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

[renatomello]

The only differentiation method that was not working for the pytorch backend was the parameter shift rule. However now i have unmuted the test and removed the check on the backend and everything is working correctly for the pytorch backend. Probably the problem was linked to some error in the initial implementation of the pytorch backend and was fixed by correcting all the other tests.

There's also all the classes inside qibo.models.variational. They need to work properly with the optmizers. RIght now the tests are not passing, they are being skipped.

[Simone-Bordoni]

The only differentiation method that was not working for the pytorch backend was the parameter shift rule. However now i have unmuted the test and removed the check on the backend and everything is working correctly for the pytorch backend. Probably the problem was linked to some error in the initial implementation of the pytorch backend and was fixed by correcting all the other tests.

There's also all the classes inside qibo.models.variational. They need to work properly with the optmizers. RIght now the tests are not passing, they are being skipped.

ok, I forgot these tests, I will try to fix them tomorrow

[Simone-Bordoni]

I have encountered a problem while coding the SGD training procedure of the VQE for the pytorch backend.
The problem is that when executing a circuit with parametrized gates that require gradients the parameters of the gates are initially casted as numpy and this is not compatible with the gradients (see failed test).
Fixing the issue requires modifying numpy_matrices.py, as this module is of fundamental importance @renatomello @BrunoLiegiBastonLiegi @MatteoRobbiati maybe we can discuss how to modify it in order to make it work correctly with torch.

[BrunoLiegiBastonLiegi]

Do you mean this cast?
https://github.com/qiboteam/qibo/blob/7b4489e6c95e5844796e8c08b6dbaacc9e28984e/src/qibo/backends/npmatrices.py#L76C1-L76C72

Because this is overloaded by the TorchMatrices object I believe.

[Simone-Bordoni]

Do you mean this cast? https://github.com/qiboteam/qibo/blob/7b4489e6c95e5844796e8c08b6dbaacc9e28984e/src/qibo/backends/npmatrices.py#L76C1-L76C72

Because this is overloaded by the TorchMatrices object I believe.

That cast is fine. The problem is in the two lines before:

cos = self.np.cos(theta / 2.0) + 0j
isin = -1j * self.np.sin(theta / 2.0)

when theta is a torch tensor with requires_grad = True you can't compute cos/sin using numpy.

[BrunoLiegiBastonLiegi]

Do you mean this cast? https://github.com/qiboteam/qibo/blob/7b4489e6c95e5844796e8c08b6dbaacc9e28984e/src/qibo/backends/npmatrices.py#L76C1-L76C72
Because this is overloaded by the TorchMatrices object I believe.

That cast is fine. The problem is in the two lines before: cos = self.np.cos(theta / 2.0) + 0j isin = -1j * self.np.sin(theta / 2.0) when theta is a torch tensor with requires_grad = True you can't compute cos/sin using numpy.

but then why isn't self.np = torch for TorchMatrices as it is for the TorchBackend?

[Simone-Bordoni]

Do you mean this cast? https://github.com/qiboteam/qibo/blob/7b4489e6c95e5844796e8c08b6dbaacc9e28984e/src/qibo/backends/npmatrices.py#L76C1-L76C72
Because this is overloaded by the TorchMatrices object I believe.

That cast is fine. The problem is in the two lines before: cos = self.np.cos(theta / 2.0) + 0j isin = -1j * self.np.sin(theta / 2.0) when theta is a torch tensor with requires_grad = True you can't compute cos/sin using numpy.

but then why isn't self.np = torch for TorchMatrices as it is for the TorchBackend?

I can try to fix things this way. With Renato we just wanted to have the matrices defined with numpy and then casted to the correct backend.

[renatomello]

Do you mean this cast? https://github.com/qiboteam/qibo/blob/7b4489e6c95e5844796e8c08b6dbaacc9e28984e/src/qibo/backends/npmatrices.py#L76C1-L76C72
Because this is overloaded by the TorchMatrices object I believe.

That cast is fine. The problem is in the two lines before: cos = self.np.cos(theta / 2.0) + 0j isin = -1j * self.np.sin(theta / 2.0) when theta is a torch tensor with requires_grad = True you can't compute cos/sin using numpy.

but then why isn't self.np = torch for TorchMatrices as it is for the TorchBackend?

This solution should be fine