Not sure how to use Non Object Optimizers

[mxv001]

Hello experts,
Can you help with the guidelines on how to use non-object based minimization?

Thank you very much!

Steps to reproduce

Run the code below and observe sensible outputs. Then uncomment the COBYLA method and the code breaks.

Observed Results

I get an error:

TypeError: NonObjectOptimizer.__init__() got an unexpected keyword argument 'parametrization'

Expected Results

I expect this to break, because the minimizer is "non object" and cannot be instantiated as an object. But I do not understand how to get a for loop and ask-tell architecture from nevergrad. How can I adapt my code to accomplish this?

Note that I have read and implemented the functional scipy version, but my question is how to use the nevergrad wrapping so that I can have ask-tell format in a for loop.

Also is it possible to pass in a Nevergrad parameterization to the wrapped method?

Relevant Code

import nevergrad as ng

def square(x, y=12):
  return sum((x - 0.5) ** 2) + abs(y)


# IT WOULD BE GREAT TO USE NEVERGRAD NATIVE PARAMETERIZATION
parametrization = ng.p.Instrumentation(
  ng.p.Array(shape=(2,)).set_bounds(lower=-12, upper=12),
  y=ng.p.Scalar()
)

def print_candidate_and_value(optimizer, candidate, value):
  print(candidate,value)

# THIS WORKS JUST FINE
optimizer  = ng.optimizers.NGOpt(parametrization=parametrization, budget=50, num_workers=1)

# UNCOMMENT THIS AND IT BREAKS
#optimizer = ng.families.NonObjectOptimizer(method= 'COBYLA',parametrization=parametrization, budget=50, num_workers=1)

optimizer.register_callback("tell", print_candidate_and_value)

# GETTING THE LOOP BELOW TO WORK IS THE GOAL 
for i in range(optimizer.budget):
  x = optimizer.ask()
  loss = square(*x.args, **x.kwargs)
  optimizer.tell(x, loss)

[mxv001]

Well, I was able to figure it out! Here is the working code if anyone else has the same goal:

def square(x, y=12):
    return sum((x - 0.5) ** 2) + abs(y)


# IT WOULD BE GREAT TO USE NEVERGRAD NATIVE PARAMETERIZATION
parametrization = ng.p.Instrumentation(
  x=ng.p.Array(shape=(2,)).set_bounds(lower=-12, upper=12),
  y=ng.p.Scalar()
)

def print_candidate_and_value(optimizer, candidate, value):
    print(candidate,value)

optimizer=nevergrad.families.NonObjectOptimizer(method = 'Nelder-Mead')(parametrization=parametrization,budget=100)
optimizer.register_callback("tell", print_candidate_and_value)

for i in range(optimizer.budget):
    x = optimizer.ask()
    loss = square(*x.args, **x.kwargs)
    optimizer.tell(x, loss)

[mxv001]

Closed