Simulated calibration¶
Calibration of control pulses is the process of fine-tuning parameters in a feedback-loop with the experiment. We will simulate this process here by constructing a black-box simulation and interacting with it exactly like an experiment.
We have manange imports and creation of the black-box the same way as in
the previous example in a helper single_qubit_blackbox_exp.py.
from single_qubit_blackbox_exp import create_experiment
blackbox = create_experiment()
This blackbox is constructed the same way as in the OptimalControl example. The difference will be in how we interact with it. First, we decide on what experiment we want to perform and need to specify it as a python function. A general, minimal example would be
def exp_communication(params):
# Send parameters to experiment controller
# and receive a measurement result.
return measurement_result
Again, params is a linear vector of bare numbers. The measurement
result can be a single number or a set of results. It can also include
additional information about statistics, like averaging, standard
deviation, etc.
ORBIT - Single-length randomized benchmarking¶
The following defines an ORBIT
procedure. In short, we define sequences of gates that result in an
identity gate if our individual gates are perfect. Any deviation from
identity gives us a measure of the imperfections in our gates. Our
helper qt_utils provides these sequences.
from c3.utils import qt_utils
qt_utils.single_length_RB(
RB_number=1, RB_length=5, target=0
)
[['ry90m[0]',
'rx90p[0]',
'rx90m[0]',
'rx90p[0]',
'ry90p[0]',
'ry90p[0]',
'ry90p[0]',
'rx90p[0]',
'ry90m[0]',
'rx90p[0]']]
The desired number of 5 gates is selected from a specific set (the Clifford group) and has to be decomposed into the available gate-set. Here, this means 4 gates per Clifford, hence a sequence of 20 gates.
Communication with the experiment¶
Some of the following code is specific to the fact that this a
simulated calibration. The interface of \(C^2\) to the experiment
is simple: parameters in \(\rightarrow\) results out. Thus, we have
to wrap the blackbox by defining the target states and the opt_map.
import numpy as np
import tensorflow as tf
def ORBIT_wrapper(p):
def ORBIT(params, exp, opt_map, qubit_labels, logdir):
### ORBIT meta-parameters ###
RB_length = 60 # How long each sequence is
RB_number = 40 # How many sequences
shots = 1000 # How many averages per readout
################################
### Simulation specific part ###
################################
do_noise = False # Whether to add artificial noise to the results
qubit_label = list(qubit_labels.keys())[0]
state_labels = qubit_labels[qubit_label]
state_label = [tuple(l) for l in state_labels]
# Creating the RB sequences #
seqs = qt_utils.single_length_RB(
RB_number=RB_number, RB_length=RB_length, target=0
)
# Transmitting the parameters to the experiment #
exp.pmap.set_parameters(params, opt_map)
exp.set_opt_gates_seq(seqs)
# Simulating the gates #
U_dict = exp.compute_propagators()
# Running the RB sequences and read-out the results #
pops = exp.evaluate(seqs)
pop1s, _ = exp.process(pops, labels=state_label)
results = []
results_std = []
shots_nums = []
# Collecting results and statistics, add noise #
if do_noise:
for p1 in pop1s:
draws = tf.keras.backend.random_binomial(
[shots],
p=p1[0],
dtype=tf.float64,
)
results.append([np.mean(draws)])
results_std.append([np.std(draws)/np.sqrt(shots)])
shots_nums.append([shots])
else:
for p1 in pop1s:
results.append(p1.numpy())
results_std.append([0])
shots_nums.append([shots])
#######################################
### End of Simulation specific part ###
#######################################
goal = np.mean(results)
return goal, results, results_std, seqs, shots_nums
return ORBIT(
p, blackbox, gateset_opt_map, state_labels, "/tmp/c3logs/blackbox"
)
Optimization¶
We first import algorithms and the correct optimizer object.
import copy
from c3.experiment import Experiment as Exp
from c3.c3objs import Quantity as Qty
from c3.parametermap import ParameterMap as PMap
from c3.libraries import algorithms, envelopes
from c3.signal import gates, pulse
from c3.optimizers.calibration import Calibration
Representation of the experiment within \(C^3\)¶
At this point we have to make sure that the gates (“RX90p”, etc.) and drive line (“d1”) are compatible to the experiment controller operating the blackbox. We mirror the blackbox by creating an experiment in the \(C^3\) context:
t_final = 7e-9 # Time for single qubit gates
sideband = 50e6
lo_freq = 5e9 + sideband
# ### MAKE GATESET
gauss_params_single = {
'amp': Qty(
value=0.45,
min_val=0.4,
max_val=0.6,
unit="V"
),
't_final': Qty(
value=t_final,
min_val=0.5 * t_final,
max_val=1.5 * t_final,
unit="s"
),
'sigma': Qty(
value=t_final / 4,
min_val=t_final / 8,
max_val=t_final / 2,
unit="s"
),
'xy_angle': Qty(
value=0.0,
min_val=-0.5 * np.pi,
max_val=2.5 * np.pi,
unit='rad'
),
'freq_offset': Qty(
value=-sideband - 0.5e6,
min_val=-53 * 1e6,
max_val=-47 * 1e6,
unit='Hz 2pi'
),
'delta': Qty(
value=-1,
min_val=-5,
max_val=3,
unit=""
)
}
gauss_env_single = pulse.Envelope(
name="gauss",
desc="Gaussian comp for single-qubit gates",
params=gauss_params_single,
shape=envelopes.gaussian_nonorm
)
nodrive_env = pulse.Envelope(
name="no_drive",
params={
't_final': Qty(
value=t_final,
min_val=0.5 * t_final,
max_val=1.5 * t_final,
unit="s"
)
},
shape=envelopes.no_drive
)
carrier_parameters = {
'freq': Qty(
value=lo_freq,
min_val=4.5e9,
max_val=6e9,
unit='Hz 2pi'
),
'framechange': Qty(
value=0.0,
min_val= -np.pi,
max_val= 3 * np.pi,
unit='rad'
)
}
carr = pulse.Carrier(
name="carrier",
desc="Frequency of the local oscillator",
params=carrier_parameters
)
rx90p = gates.Instruction(
name="rx90p",
t_start=0.0,
t_end=t_final,
channels=["d1"]
)
QId = gates.Instruction(
name="id",
t_start=0.0,
t_end=t_final,
channels=["d1"]
)
rx90p.add_component(gauss_env_single, "d1")
rx90p.add_component(carr, "d1")
QId.add_component(nodrive_env, "d1")
QId.add_component(copy.deepcopy(carr), "d1")
QId.comps['d1']['carrier'].params['framechange'].set_value(
(-sideband * t_final * 2 * np.pi) % (2*np.pi)
)
ry90p = copy.deepcopy(rx90p)
ry90p.name = "ry90p"
rx90m = copy.deepcopy(rx90p)
rx90m.name = "rx90m"
ry90m = copy.deepcopy(rx90p)
ry90m.name = "ry90m"
ry90p.comps['d1']['gauss'].params['xy_angle'].set_value(0.5 * np.pi)
rx90m.comps['d1']['gauss'].params['xy_angle'].set_value(np.pi)
ry90m.comps['d1']['gauss'].params['xy_angle'].set_value(1.5 * np.pi)
parameter_map = PMap(instructions=[QId, rx90p, ry90p, rx90m, ry90m])
# ### MAKE EXPERIMENT
exp = Exp(pmap=parameter_map)
Next, we define the parameters we whish to calibrate. See how these gate
instructions are defined in the experiment setup example or in
single_qubit_blackbox_exp.py. Our gate-set is made up of 4 gates,
rotations of 90 degrees around the \(x\) and \(y\)-axis in
positive and negative direction. While it is possible to optimize each
parameters of each gate individually, in this example all four gates
share parameters. They only differ in the phase \(\phi_{xy}\) that
is set in the definitions.
gateset_opt_map = [
[
("rx90p[0]", "d1", "gauss", "amp"),
("ry90p[0]", "d1", "gauss", "amp"),
("rx90m[0]", "d1", "gauss", "amp"),
("ry90m[0]", "d1", "gauss", "amp")
],
[
("rx90p[0]", "d1", "gauss", "delta"),
("ry90p[0]", "d1", "gauss", "delta"),
("rx90m[0]", "d1", "gauss", "delta"),
("ry90m[0]", "d1", "gauss", "delta")
],
[
("rx90p[0]", "d1", "gauss", "freq_offset"),
("ry90p[0]", "d1", "gauss", "freq_offset"),
("rx90m[0]", "d1", "gauss", "freq_offset"),
("ry90m[0]", "d1", "gauss", "freq_offset")
],
[
("id[0]", "d1", "carrier", "framechange")
]
]
parameter_map.set_opt_map(gateset_opt_map)
As defined above, we have 16 parameters where 4 share their numerical value. This leaves 4 values to optimize.
parameter_map.print_parameters()
rx90p[0]-d1-gauss-amp : 450.000 mV
ry90p[0]-d1-gauss-amp
rx90m[0]-d1-gauss-amp
ry90m[0]-d1-gauss-amp
rx90p[0]-d1-gauss-delta : -1.000
ry90p[0]-d1-gauss-delta
rx90m[0]-d1-gauss-delta
ry90m[0]-d1-gauss-delta
rx90p[0]-d1-gauss-freq_offset : -50.500 MHz 2pi
ry90p[0]-d1-gauss-freq_offset
rx90m[0]-d1-gauss-freq_offset
ry90m[0]-d1-gauss-freq_offset
id[0]-d1-carrier-framechange : 4.084 rad
It is important to note that in this example, we are transmitting only these four parameters to the experiment. We don’t know how the blackbox will implement the pulse shapes and care has to be taken that the parameters are understood on the other end. Optionally, we could specifiy a virtual AWG within \(C^3\) and transmit pixilated pulse shapes directly to the physiscal AWG.
Algorithms¶
As an optimization algoritm, we choose CMA-Es and set up some options specific to this algorithm.
alg_options = {
"popsize" : 10,
"maxfevals" : 300,
"init_point" : "True",
"tolfun" : 0.01,
"spread" : 0.25
}
We define the subspace as both excited states \(\{|1>,|2>\}\), assuming read-out can distinguish between 0, 1 and 2.
state_labels = {
"excited" : [(1,), (2,)]
}
In the real world, this setup needs to be handled in the experiment controller side. We construct the optimizer object with the options we setup:
import os
import tempfile
# Create a temporary directory to store logfiles, modify as needed
log_dir = os.path.join(tempfile.TemporaryDirectory().name, "c3logs")
opt = Calibration(
dir_path=log_dir,
run_name="ORBIT_cal",
eval_func=ORBIT_wrapper,
pmap=parameter_map,
exp_right=exp,
algorithm=algorithms.cmaes,
options=alg_options
)
opt.set_exp(exp)
And run the calibration:
x = parameter_map.get_parameters_scaled()
opt.optimize_controls()
C3:STATUS:Saving as: /tmp/tmpicnnbliz/c3logs/ORBIT_cal/2021_01_28_T_15_17_30/calibration.log
(5_w,10)-aCMA-ES (mu_w=3.2,w_1=45%) in dimension 4 (seed=912463, Thu Jan 28 15:17:30 2021)
C3:STATUS:Adding initial point to CMA sample.
Iterat #Fevals function value axis ratio sigma min&max std t[m:s]
1 10 1.446744168975211e-01 1.0e+00 2.11e-01 2e-01 2e-01 1:18.9
2 20 2.074359374665050e-01 1.4e+00 1.96e-01 1e-01 2e-01 2:28.5
3 30 1.042216610303495e-01 1.5e+00 1.76e-01 1e-01 2e-01 3:36.4
4 40 1.720244494886762e-01 1.9e+00 1.88e-01 1e-01 2e-01 4:46.5
5 50 9.761264536669531e-02 2.2e+00 2.05e-01 1e-01 2e-01 6:15.4
6 60 1.956493007802809e-01 2.8e+00 1.75e-01 8e-02 2e-01 7:17.9
7 70 6.625917264980545e-02 3.0e+00 2.20e-01 9e-02 3e-01 8:22.8
8 80 7.697621753428294e-02 4.1e+00 2.19e-01 8e-02 3e-01 9:25.8
9 90 8.826758030850271e-02 4.7e+00 1.85e-01 6e-02 3e-01 10:28.7
10 100 9.099567192014653e-02 5.3e+00 1.59e-01 4e-02 2e-01 11:32.7
11 110 6.673347151005890e-02 6.9e+00 1.49e-01 3e-02 2e-01 12:27.9
12 120 6.822093884865452e-02 7.6e+00 1.68e-01 4e-02 2e-01 13:26.6
13 130 6.307315835232992e-02 8.1e+00 1.42e-01 3e-02 2e-01 14:22.8
14 140 6.301017013241370e-02 7.8e+00 1.42e-01 2e-02 2e-01 15:18.7
15 150 6.795728963072037e-02 9.3e+00 1.32e-01 2e-02 2e-01 16:15.8
16 160 7.675314380135559e-02 9.2e+00 1.03e-01 2e-02 1e-01 17:12.9
17 170 6.806172046778505e-02 9.1e+00 8.05e-02 1e-02 1e-01 18:11.5
18 180 5.698438523961635e-02 1.0e+01 7.42e-02 9e-03 9e-02 19:06.1
19 190 5.536707419037251e-02 1.1e+01 6.89e-02 8e-03 9e-02 20:00.6
20 200 4.924177790655197e-02 1.2e+01 7.31e-02 8e-03 9e-02 20:58.2
21 210 5.836136870997249e-02 1.2e+01 8.20e-02 8e-03 1e-01 21:55.1
22 220 5.463139088536284e-02 1.3e+01 8.29e-02 9e-03 1e-01 22:51.0
23 230 4.562693294212217e-02 1.4e+01 8.66e-02 9e-03 1e-01 23:48.3
24 240 5.188441161313757e-02 1.6e+01 7.74e-02 7e-03 1e-01 24:46.1
25 250 5.199237655967553e-02 1.7e+01 7.41e-02 6e-03 9e-02 25:47.1
26 260 5.684400595430246e-02 1.6e+01 6.41e-02 5e-03 9e-02 26:43.7
27 270 4.441763519087279e-02 1.8e+01 5.12e-02 4e-03 7e-02 27:36.2
28 280 4.994977609185950e-02 1.8e+01 5.51e-02 5e-03 8e-02 28:33.9
29 290 6.108777009078262e-02 1.8e+01 5.14e-02 4e-03 7e-02 29:30.4
30 300 5.658962789881571e-02 1.8e+01 4.65e-02 4e-03 6e-02 30:28.0
31 310 5.765354335022381e-02 1.8e+01 4.77e-02 4e-03 6e-02 31:26.9
termination on maxfevals=300
final/bestever f-value = 5.765354e-02 4.441764e-02
incumbent solution: [-0.4739081748676816, -0.09828275146514219, -1.0504851431889897, 0.9108808620989909]
std deviation: [0.013780217516583012, 0.0038070906112681576, 0.02460767003734409, 0.05816700836608336]
Analysis¶
The following code uses matplotlib to create an ORBIT plot from the logfile.
import json
from matplotlib.ticker import MaxNLocator
from matplotlib import rcParams
from matplotlib import cycler
import matplotlib as mpl
import matplotlib.pyplot as plt
rcParams['xtick.direction'] = 'in'
rcParams['axes.grid'] = True
rcParams['grid.linestyle'] = '--'
rcParams['markers.fillstyle'] = 'none'
rcParams['axes.prop_cycle'] = cycler(
'linestyle', ["-", "--"]
)
rcParams['text.usetex'] = True
rcParams['font.size'] = 16
rcParams['font.family'] = 'serif'
logfilename = opt.logdir + "calibration.log"
with open(logfilename, "r") as filename:
log = filename.readlines()
options = json.loads(log[7])
goal_function = []
batch = 0
batch_size = options["popsize"]
eval = 0
for line in log[9:]:
if line[0] == "{":
if not eval % batch_size:
batch = eval // batch_size
goal_function.append([])
eval += 1
point = json.loads(line)
if 'goal' in point.keys():
goal_function[batch].append(point['goal'])
# Clean unfinished batch
if len(goal_function[-1])<batch_size:
goal_function.pop(-1)
fig, ax = plt.subplots(1)
means = []
bests = []
for ii in range(len(goal_function)):
means.append(np.mean(np.array(goal_function[ii])))
bests.append(np.min(np.array(goal_function[ii])))
for pt in goal_function[ii]:
ax.plot(ii+1, pt, color='tab:blue', marker="D", markersize=2.5, linewidth=0)
ax.xaxis.set_major_locator(MaxNLocator(integer=True))
ax.set_ylabel('ORBIT')
ax.set_xlabel('Iterations')
ax.plot(
range(1, len(goal_function)+1), bests, color="tab:red", marker="D",
markersize=5.5, linewidth=0, fillstyle='full'
)
