Tutorial 4. Adaptive Measurements

See also

The complete source code of this tutorial can be found in

Tutorial 4. Adaptive Measurements.ipynb

Tutorial 4. Adaptive Measurements.py

Following this Tutorial requires familiarity with the core concepts of Quantify, we highly recommended to consult the (short) usage. If you have some difficulties following the tutorial it might be worth reviewing the User guide!

We highly recommended to first follow Tutorial 1. Controlling a basic experiment using MeasurementControl and Tutorial 2. Advanced capabilities of the MeasurementControl.

In this tutorial, we explore the adaptive functionality of the MeasurementControl. With this mode, instead of predefining a grid of values to sweep through, we provide an optimization function and an initial state to the meas_ctrl. The meas_ctrl will then use this function to build the sweep. We import our usual modules and setup an meas_ctrl with visualization:

import time

import adaptive
import numpy as np
from qcodes import Instrument, ManualParameter
from scipy import optimize

import quantify_core.analysis.optimization_analysis as oa
import quantify_core.visualization.pyqt_plotmon as pqm
from quantify_core.analysis.interpolation_analysis import InterpolationAnalysis2D
from quantify_core.data.handling import set_datadir
from quantify_core.measurement.control import MeasurementControl
from quantify_core.utilities.examples_support import default_datadir
from quantify_core.visualization.instrument_monitor import InstrumentMonitor

Before instantiating any instruments or starting a measurement we change the directory in which the experiments are saved using the set_datadir() [get_datadir()] functions.

Data will be saved in:
meas_ctrl = MeasurementControl("meas_ctrl")
insmon = InstrumentMonitor("Instruments Monitor")
plotmon = pqm.PlotMonitor_pyqt("plotmon_meas_ctrl")

Finding a minimum

We will create a mock Instrument our meas_ctrl will interact with. In this case, it is a simple parabola centered at the origin.

para = Instrument("parabola")

para.add_parameter("x", unit="m", label="X", parameter_class=ManualParameter)
para.add_parameter("y", unit="m", label="Y", parameter_class=ManualParameter)

    "noise", unit="V", label="white noise amplitude", parameter_class=ManualParameter
    "acq_delay", initial_value=0.1, unit="s", parameter_class=ManualParameter

def _amp_model():
    )  # for display purposes, just so we can watch the live plot update
    return para.x() ** 2 + para.y() ** 2 + para.noise() * np.random.rand(1)

para.add_parameter("amp", unit="V", label="Amplitude", get_cmd=_amp_model)

Next, we will use the optimize package from scipy to provide our adaptive function. You can of course implement your own functions for this purpose, but for brevity we will use something standard and easily available.

Then, we set our Settables and Gettables as usual, and define a new dictionary af_pars. The only required key in this object is “adaptive_function”, the value of which being the adaptive function to use. The remaining fields in this dictionary are the arguments to the adaptive function itself. We also add some noise into the parabola to stress our adaptive function.

As such, it is highly recommended to thoroughly read the documentation around the adaptive function you are using.

We will use the optimize.minimize function (note this is passed by reference as opposed to calling the minimize function), which requires an initial state named "x0" and an algorithm to use named "method". In this case, we are starting at [-50, -50] and hope to minimize these values relative to our parabola function. Of course, this parabola has it’s global minimum at the origin, thus these values will tend towards 0 as our algorithm progresses.

meas_ctrl.settables([para.x, para.y])
af_pars = {
    "adaptive_function": optimize.minimize,  # used by meas_ctrl
    "x0": [-50, -50],  # used by `optimize.minimize` (in this case)
    "method": "Nelder-Mead",  # used by `optimize.minimize` (in this case)
    "options": {"maxfev": 100},  # limit the maximum evaluations of the gettable(s)
dset = meas_ctrl.run_adaptive("nelder_mead_optimization", af_pars)
Dimensions:  (dim_0: 100)
    x0       (dim_0) float64 -50.0 -52.5 -50.0 -47.5 ... -0.1848 -0.1848 -0.1847
    x1       (dim_0) float64 -50.0 -50.0 -52.5 -52.5 ... 0.2157 0.2156 0.2156
Dimensions without coordinates: dim_0
Data variables:
    y0       (dim_0) float64 5e+03 5.257e+03 5.256e+03 ... 0.5418 0.1151 0.3405
    tuid:                             20221013-182745-048-15ab14
    name:                             nelder_mead_optimization
    grid_2d:                          False
    grid_2d_uniformly_spaced:         False
    1d_2_settables_uniformly_spaced:  False
../_images/Tutorial 4. Adaptive Measurements_7_0.png
../_images/Tutorial 4. Adaptive Measurements_8_0.png

We can see from the graphs that the values of the settables in the dataset snake towards 0 as expected. Success!


There are several analysis classes available in quantify which can be used to visualize and extract relevant information from the results of these adaptive measurements.

The OptimizationAnalysis class searches the dataset for the optimal datapoint and provides a number of useful plots to visualize the convergence of the measurement result around the minimum.

a_obj = oa.OptimizationAnalysis(dset)
../_images/Tutorial 4. Adaptive Measurements_9_0.png ../_images/Tutorial 4. Adaptive Measurements_9_1.png ../_images/Tutorial 4. Adaptive Measurements_9_2.png

The analysis generates plots of each of the variables versus the number of iteration steps completed. The figures show the data converging on the optimal value.

The InterpolationAnalysis2D class can be used to generate a 2-dimensional heatmap which interpolates between a set of irregularly spaced datapoints.

a_obj = InterpolationAnalysis2D(dset)
../_images/Tutorial 4. Adaptive Measurements_10_0.png

Adaptive Sampling

Quantify is designed to be modular and the adaptive functions support is no different. To this end, the meas_ctrl has first class support for the adaptive package. Let’s see what the same experiment looks like with this module. Note the fields of the af_pars dictionary have changed to be compatible with the different adaptive function we are using.

As a practical example, let’s revisit a Resonator Spectroscopy experiment. This time we only know our device has a resonance in 6-7 GHz range. We really don’t want to sweep through a million points, so instead let’s use an adaptive sampler to quickly locate our peak.

res = Instrument("Resonator")

res.add_parameter("freq", unit="Hz", label="Frequency", parameter_class=ManualParameter)
res.add_parameter("amp", unit="V", label="Amplitude", parameter_class=ManualParameter)
_fwhm = 15e6  # pretend you don't know what this value is
_res_freq = 6.78e9  # pretend you don't know what this value is
_noise_level = 0.1

def lorenz():
    """A Lorenz model function."""
    time.sleep(0.02)  # for display purposes, just so we can watch the graph update
    return (
        - (
            * ((_fwhm / 2.0) ** 2)
            / ((res.freq() - _res_freq) ** 2 + (_fwhm / 2.0) ** 2)
        + _noise_level * np.random.rand(1)

res.add_parameter("S21", unit="V", label="Transmission amp. S21", get_cmd=lorenz)
_noise_level = 0.0
af_pars = {
    "adaptive_function": adaptive.learner.Learner1D,
    "goal": lambda l: l.npoints > 99,
    "bounds": (6.0e9, 7.0e9),
dset = meas_ctrl.run_adaptive("adaptive sample", af_pars)
Dimensions:  (dim_0: 100)
    x0       (dim_0) float64 6e+09 7e+09 6.5e+09 ... 6.79e+09 6.77e+09 6.754e+09
Dimensions without coordinates: dim_0
Data variables:
    y0       (dim_0) float64 0.9999 0.9988 0.9993 ... 0.6186 0.6391 0.9237
    tuid:                             20221013-182801-129-845547
    name:                             adaptive sample
    grid_2d:                          False
    grid_2d_uniformly_spaced:         False
    1d_2_settables_uniformly_spaced:  False
../_images/Tutorial 4. Adaptive Measurements_14_0.png


Can I return multi-dimensional data from a Gettable in Adaptive Mode?

Yes, but only first dimension (y0) will be considered by the adaptive function; the remaining dimensions will merely be saved to the dataset.