DBF step optimization

[MatteoRobbiati]

A first possible optimization of the step is implemented.
We use hyperopt to optimize the step considering a single flow execution.

An example of usage:

import matplotlib.pyplot as plt
import numpy as np

from qibo.models.double_bracket import DoubleBracketFlow, FlowGeneratorType
from qibo import set_backend
from qibo.hamiltonians.hamiltonians import Hamiltonian
from qibo.quantum_info import random_hermitian

set_backend("numpy")

NSTEPS=20
nqubits = 4
h0 = random_hermitian(2**nqubits)
DELTA=0.005

np.random.seed(42)

norms = []
# No hyperopt vs hyperopt
methods = [None, "opt"]
init_step = 0.01

for i, meth in enumerate(methods):
    print(f"Testing opt: {meth}")
    norm_history = []

    dbf = DoubleBracketFlow(Hamiltonian(nqubits=nqubits, matrix=h0))
    norm_history.append(dbf.off_diagonal_norm)
    if meth is not None:
        print("Big optimization for setting the initial step:")
        step = dbf.hyperopt_step(step_min=0.005, step_max=0.02, max_evals=1000, verbose=True)

    for i in range(NSTEPS):
        # evolution under canonical commutator
        if meth is not None:
            # here a hyper-parameter space and an hyperopt algorithm can be passed
            # default are hp.uniform and tpe
            step = dbf.hyperopt_step(step_min=step-DELTA, step_max=step+DELTA, max_evals=100)
            print(f"New flow duration s={step}")
        dbf(step=init_step, mode=FlowGeneratorType.canonical)
        norm_history.append(dbf.off_diagonal_norm)
    norms.append(norm_history)


plt.figure(figsize=(5, 5*6/8))
plt.plot(norms[0], label="Fixed step", color="royalblue", alpha=0.7, lw=1.5)
plt.plot(norms[1], label="Optimized", color="red", alpha=0.7, lw=1.5)
plt.legend()
plt.xlabel("Steps")
plt.ylabel("Cost")
plt.savefig("test_fig.png")

And an example of benchmark between fixed step size (blue), and optimized step size:

Checklist:

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

[codecov]

Codecov Report

Attention: 1 lines in your changes are missing coverage. Please review.

Comparison is base (dbc8a12) 100.00% compared to head (f7d5bf8) 99.98%.

❗ Current head f7d5bf8 differs from pull request most recent head e7bdf50. Consider uploading reports for the commit e7bdf50 to get more accurate results

Files	Patch %	Lines
src/qibo/models/double_bracket.py	95.23%	1 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##               dbf    #1081      +/-   ##
===========================================
- Coverage   100.00%   99.98%   -0.02%     
===========================================
  Files           51       51              
  Lines         7583     7601      +18     
===========================================
+ Hits          7583     7600      +17     
- Misses           0        1       +1     

Flag	Coverage Δ
unittests	99.98% <95.23%> (-0.02%)	⬇️

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.

[andrea-pasquale]

Thanks @MatteoRobbiati for implementing this.
I have a few suggestions down below.

[andrea-pasquale]

Thanks for all the updates @MatteoRobbiati.
Just a small comment regarding how to deal with hyperopt again.

[andrea-pasquale]

Now that we removed the hyperopt dep from the main deps, when you try to import this module it will raise a ModuleNotFoundError.
I think that you should move this import directly inside the hyperopt_step possibly in a try-except block where you can tell the user to install hyperopt if they want to perform the optimiziation.
Perhaps at this point it is easier just to keep it as a dependency.
What do you recommend @scarrazza? Shall we add hyperopt to the main deps or we keep it as an optional dep?

[MatteoRobbiati]

If we like this setup, in which the user can set the space and the algorithm directly into the function, I think it is necessary to import hyperopt at the beginning of the double_bracket.py file.

Moreover, we never spent a lot of time in doing hyperoptimization of VQCs (number of layers, learning rate of the optimizers, etc).
I think adding an hyperoptimization tool as dependence of Qibo can be useful in general.
If you prefere, we can choose between hyperopt and Optuna, I am quite indifferent on this.

[Edoardo-Pedicillo]

I think that you should move this import directly inside the hyperopt_step possibly in a try-except block where you can tell the user to install hyperopt if they want to perform the optimiziation.

I agree with this option