remove Tensorflow dependency

[Edoardo-Pedicillo]

This PR closes #734

Checklist:

• Reviewers confirm new code works as expected.
• Tests are passing.
• Coverage does not decrease.
• Documentation is updated.
• Compatibility with Qibo modules (Please edit this section if the current pull request is not compatible with the following branches).
  • Qibo: master
  • Qibolab: main
  • Qibolab_platforms_qrc: main

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 96.73%. Comparing base (b811472) to head (00afd66).
Report is 7 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #750   +/-   ##
=======================================
  Coverage   96.73%   96.73%           
=======================================
  Files         108      108           
  Lines        7498     7498           
=======================================
  Hits         7253     7253           
  Misses        245      245           

Flag	Coverage Δ
unittests	96.73% <ø> (ø)

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

Files	Coverage Δ
src/qibocal/fitting/classifier/run.py	97.47% <ø> (ø)
...bocal/protocols/characterization/classification.py	100.00% <ø> (ø)

[alecandido]

Do we still need onnx? Is there a different way of serializing Scikit-Learn objects? (maybe even just dictionaries for the hyperparams, and dump in json)

[alecandido]

Thanks @Edoardo-Pedicillo, I only had that comment, and it could be split to a separate issue/PR as well. The PR itself seems alright.

[alecandido]

Consider that the dump() function, for which you need ONNX at the moment (even in SciKit you need the runtime), seems not to be used.

qibocal/src/qibocal/fitting/classifier/ada_boost.py

Line 41 in 29cc267

dump = scikit_utils.scikit_dump

At least not in run.py.

But please, also check yourself.

[Edoardo-Pedicillo]

The dump() function is used in the classification routine

qibocal/src/qibocal/protocols/characterization/classification.py

Line 122 in 1026f4f

classifier.dump()(self.models[qubit][i], dump_dir)

[alecandido]

The dump() function is used in the classification routine

Correct, that's why I was not finding it in classifiers. So, you need of course to dump the result of the fit. But is there no intrinsic way of Scikit-Learn to serialize this? (I mean, w/o ONNX)

[Edoardo-Pedicillo]

Pikle ;) btw onnx is suggested by scikit itself (https://scikit-learn.org/stable/model_persistence.html) and even the other options require extra dependencies

[alecandido]

I was wondering if there was something like a .params attribute to dump and load from JSON. But I would avoid going too much into the detail ourselves. Auditing dependencies is a good practice, but also software reuse (i.e. not reinventing the wheel).

There is also PMML (XML instead of binary) and skops, but they do not seem to be more maintained than scikit-onnx. And onnxruntime is incredibly more maintained and reliable than any of the previous ones.

Let's keep it as it is.