This is a static, non-editable tutorial.
We recommend you install QuCumber if you want to run the examples locally. You can then get an archive file containing the examples from the relevant release here. Alternatively, you can launch an interactive online version, though it may be a bit slow:
Reconstruction of a complex wavefunction¶
In this tutorial, a walkthrough of how to reconstruct a complex wavefunction via training a Restricted Boltzmann Machine (RBM), the neural network behind QuCumber, will be presented.
The wavefunction to be reconstructed¶
The simple wavefunction below describing two qubits (coefficients stored in qubits_psi.txt
) will be reconstructed.
where the exact values of and used for this tutorial are
The example dataset, qubits_train.txt
, comprises of 500 measurements made in various bases (X, Y and Z). A corresponding file containing the bases for each data point in qubits_train.txt
, qubits_train_bases.txt
, is also required. As per convention, spins are represented in binary notation with zero and one denoting spin-down and spin-up, respectively.
Using qucumber to reconstruct the wavefunction¶
Imports¶
To begin the tutorial, first import the required Python packages.
[1]:
import numpy as np
import torch
import matplotlib.pyplot as plt
from qucumber.nn_states import ComplexWaveFunction
from qucumber.callbacks import MetricEvaluator
import qucumber.utils.unitaries as unitaries
import qucumber.utils.cplx as cplx
import qucumber.utils.training_statistics as ts
import qucumber.utils.data as data
import qucumber
# set random seed on cpu but not gpu, since we won't use gpu for this tutorial
qucumber.set_random_seed(1234, cpu=True, gpu=False)
The Python class ComplexWaveFunction
contains generic properties of a RBM meant to reconstruct a complex wavefunction, the most notable one being the gradient function required for stochastic gradient descent.
To instantiate a ComplexWaveFunction
object, one needs to specify the number of visible and hidden units in the RBM. The number of visible units, num_visible
, is given by the size of the physical system, i.e. the number of spins or qubits (2 in this case), while the number of hidden units, num_hidden
, can be varied to change the expressiveness of the neural network.
Note: The optimal num_hidden
: num_visible
ratio will depend on the system. For the two-qubit wavefunction described above, good results can be achieved when this ratio is 1.
On top of needing the number of visible and hidden units, a ComplexWaveFunction
object requires the user to input a dictionary containing the unitary operators (2x2) that will be used to rotate the qubits in and out of the computational basis, Z, during the training process. The unitaries
utility will take care of creating this dictionary.
The MetricEvaluator
class and training_statistics
utility are built-in amenities that will allow the user to evaluate the training in real time.
Lastly, the cplx
utility allows QuCumber to be able to handle complex numbers as they are not currently supported by PyTorch.
Training¶
To evaluate the training in real time, the fidelity between the true wavefunction of the system and the wavefunction that QuCumber reconstructs, , will be calculated along with the Kullback-Leibler (KL) divergence (the RBM’s cost function). First, the training data and the true wavefunction of this system need to be loaded using the data
utility.
[2]:
train_path = "qubits_train.txt"
train_bases_path = "qubits_train_bases.txt"
psi_path = "qubits_psi.txt"
bases_path = "qubits_bases.txt"
train_samples, true_psi, train_bases, bases = data.load_data(
train_path, psi_path, train_bases_path, bases_path
)
The file qubits_bases.txt
contains every unique basis in the qubits_train_bases.txt
file. Calculation of the full KL divergence in every basis requires the user to specify each unique basis.
As previously mentioned, a ComplexWaveFunction
object requires a dictionary that contains the unitary operators that will be used to rotate the qubits in and out of the computational basis, Z, during the training process. In the case of the provided dataset, the unitaries required are the well-known , and gates. The dictionary needed can be created with the following command.
[3]:
unitary_dict = unitaries.create_dict()
# unitary_dict = unitaries.create_dict(<unitary_name>=torch.tensor([[real part],
# [imaginary part]],
# dtype=torch.double)
If the user wishes to add their own unitary operators from their experiment to unitary_dict
, uncomment the block above. When unitaries.create_dict()
is called, it will contain the identity and the and gates by default under the keys “Z”, “X” and “Y”, respectively.
The number of visible units in the RBM is equal to the number of qubits. The number of hidden units will also be taken to be the number of visible units.
[4]:
nv = train_samples.shape[-1]
nh = nv
nn_state = ComplexWaveFunction(
num_visible=nv, num_hidden=nh, unitary_dict=unitary_dict, gpu=False
)
If gpu=True
(the default), QuCumber will attempt to run on a GPU if one is available (if one is not available, QuCumber will fall back to CPU). If one wishes to guarantee that QuCumber runs on the CPU, add the flag gpu=False
in the ComplexWaveFunction
object instantiation.
Now the hyperparameters of the training process can be specified.
epochs
: the total number of training cycles that will be performed (default = 100)pos_batch_size
: the number of data points used in the positive phase of the gradient (default = 100)neg_batch_size
: the number of data points used in the negative phase of the gradient (default =pos_batch_size
)k
: the number of contrastive divergence steps (default = 1)lr
: the learning rate (default = 0.001)Note: For more information on the hyperparameters above, it is strongly encouraged that the user to read through the brief, but thorough theory document on RBMs. One does not have to specify these hyperparameters, as their default values will be used without the user overwriting them. It is recommended to keep with the default values until the user has a stronger grasp on what these hyperparameters mean. The quality and the computational efficiency of the training will highly depend on the choice of hyperparameters. As such, playing around with the hyperparameters is almost always necessary.
The two-qubit example in this tutorial should be extremely easy to train, regardless of the choice of hyperparameters. However, the hyperparameters below will be used.
[5]:
epochs = 500
pbs = 100 # pos_batch_size
nbs = pbs # neg_batch_size
lr = 0.1
k = 10
For evaluating the training in real time, the MetricEvaluator
will be called to calculate the training evaluators every 10 epochs. The MetricEvaluator
requires the following arguments.
period
: the frequency of the training evaluators being calculated (e.g.period=200
means that theMetricEvaluator
will compute the desired metrics every 200 epochs)A dictionary of functions you would like to reference to evaluate the training (arguments required for these functions are keyword arguments placed after the dictionary)
The following additional arguments are needed to calculate the fidelity and KL divergence in the training_statistics
utility.
target_psi
(the true wavefunction of the system)space
(the entire Hilbert space of the system)
The training evaluators can be printed out via the verbose=True
statement.
Although the fidelity and KL divergence are excellent training evaluators, they are not practical to calculate in most cases; the user may not have access to the target wavefunction of the system, nor may generating the Hilbert space of the system be computationally feasible. However, evaluating the training in real time is extremely convenient.
Any custom function that the user would like to use to evaluate the training can be given to the MetricEvaluator
, thus avoiding having to calculate fidelity and/or KL divergence. As an example, functions that calculate the the norm of each of the reconstructed wavefunction’s coefficients are presented. Any custom function given to MetricEvaluator
must take the neural-network state (in this case, the ComplexWaveFunction
object) and keyword arguments. Although the given example
requires the Hilbert space to be computed, the scope of the MetricEvaluator
’s ability to be able to handle any function should still be evident.
[6]:
def alpha(nn_state, space, **kwargs):
rbm_psi = nn_state.psi(space)
normalization = nn_state.normalization(space).sqrt_()
alpha_ = cplx.norm(
torch.tensor([rbm_psi[0][0], rbm_psi[1][0]], device=nn_state.device)
/ normalization
)
return alpha_
def beta(nn_state, space, **kwargs):
rbm_psi = nn_state.psi(space)
normalization = nn_state.normalization(space).sqrt_()
beta_ = cplx.norm(
torch.tensor([rbm_psi[0][1], rbm_psi[1][1]], device=nn_state.device)
/ normalization
)
return beta_
def gamma(nn_state, space, **kwargs):
rbm_psi = nn_state.psi(space)
normalization = nn_state.normalization(space).sqrt_()
gamma_ = cplx.norm(
torch.tensor([rbm_psi[0][2], rbm_psi[1][2]], device=nn_state.device)
/ normalization
)
return gamma_
def delta(nn_state, space, **kwargs):
rbm_psi = nn_state.psi(space)
normalization = nn_state.normalization(space).sqrt_()
delta_ = cplx.norm(
torch.tensor([rbm_psi[0][3], rbm_psi[1][3]], device=nn_state.device)
/ normalization
)
return delta_
Now the basis of the Hilbert space of the system must be generated in order to compute the fidelity, KL divergence, and the dictionary of functions the user would like to compute. These metrics will be evaluated every period
epochs, which is a parameter that must be given to the MetricEvaluator
.
Note that some of the coefficients are not being evaluated as they are commented out. This is simply to avoid cluttering the output, and may be uncommented by the user.
[7]:
period = 25
space = nn_state.generate_hilbert_space()
callbacks = [
MetricEvaluator(
period,
{
"Fidelity": ts.fidelity,
"KL": ts.KL,
"normα": alpha,
# "normβ": beta,
# "normγ": gamma,
# "normδ": delta,
},
target=true_psi,
bases=bases,
verbose=True,
space=space,
)
]
Now the training can begin. The ComplexWaveFunction
object has a function called fit
which takes care of this.
[8]:
nn_state.fit(
train_samples,
epochs=epochs,
pos_batch_size=pbs,
neg_batch_size=nbs,
lr=lr,
k=k,
input_bases=train_bases,
callbacks=callbacks,
time=True,
)
Epoch: 25 Fidelity = 0.940240 KL = 0.032256 normα = 0.258429
Epoch: 50 Fidelity = 0.974944 KL = 0.017143 normα = 0.260490
Epoch: 75 Fidelity = 0.984727 KL = 0.012232 normα = 0.270684
Epoch: 100 Fidelity = 0.987769 KL = 0.010389 normα = 0.269163
Epoch: 125 Fidelity = 0.988929 KL = 0.009581 normα = 0.261813
Epoch: 150 Fidelity = 0.989075 KL = 0.009273 normα = 0.271764
Epoch: 175 Fidelity = 0.989197 KL = 0.008928 normα = 0.267943
Epoch: 200 Fidelity = 0.989451 KL = 0.008817 normα = 0.259327
Epoch: 225 Fidelity = 0.990894 KL = 0.007215 normα = 0.269941
Epoch: 250 Fidelity = 0.991517 KL = 0.006804 normα = 0.261673
Epoch: 275 Fidelity = 0.991808 KL = 0.006408 normα = 0.261002
Epoch: 300 Fidelity = 0.992318 KL = 0.005788 normα = 0.274654
Epoch: 325 Fidelity = 0.992078 KL = 0.005881 normα = 0.266831
Epoch: 350 Fidelity = 0.991938 KL = 0.006020 normα = 0.262980
Epoch: 375 Fidelity = 0.991670 KL = 0.006181 normα = 0.270877
Epoch: 400 Fidelity = 0.992082 KL = 0.005945 normα = 0.255576
Epoch: 425 Fidelity = 0.992678 KL = 0.005130 normα = 0.259746
Epoch: 450 Fidelity = 0.993102 KL = 0.004702 normα = 0.259373
Epoch: 475 Fidelity = 0.993109 KL = 0.004765 normα = 0.255803
Epoch: 500 Fidelity = 0.992805 KL = 0.004785 normα = 0.261486
Total time elapsed during training: 49.059 s
All of these training evaluators can be accessed after the training has completed, as well. The code below shows this, along with plots of each training evaluator versus the training cycle number (epoch).
[9]:
# Note that the key given to the *MetricEvaluator* must be
# what comes after callbacks[0].
fidelities = callbacks[0].Fidelity
# Alternatively, we may use the usual dictionary/list subscripting
# syntax. This is useful in cases where the name of the metric
# may contain special characters or spaces.
KLs = callbacks[0]["KL"]
coeffs = callbacks[0]["normα"]
epoch = np.arange(period, epochs + 1, period)
[10]:
# Some parameters to make the plots look nice
params = {
"text.usetex": True,
"font.family": "serif",
"legend.fontsize": 14,
"figure.figsize": (10, 3),
"axes.labelsize": 16,
"xtick.labelsize": 14,
"ytick.labelsize": 14,
"lines.linewidth": 2,
"lines.markeredgewidth": 0.8,
"lines.markersize": 5,
"lines.marker": "o",
"patch.edgecolor": "black",
}
plt.rcParams.update(params)
plt.style.use("seaborn-deep")
[11]:
fig, axs = plt.subplots(nrows=1, ncols=3, figsize=(14, 3))
ax = axs[0]
ax.plot(epoch, fidelities, "o", color="C0", markeredgecolor="black")
ax.set_ylabel(r"Fidelity")
ax.set_xlabel(r"Epoch")
ax = axs[1]
ax.plot(epoch, KLs, "o", color="C1", markeredgecolor="black")
ax.set_ylabel(r"KL Divergence")
ax.set_xlabel(r"Epoch")
ax = axs[2]
ax.plot(epoch, coeffs, "o", color="C2", markeredgecolor="black")
ax.set_ylabel(r"$\vert\alpha\vert$")
ax.set_xlabel(r"Epoch")
plt.tight_layout()
plt.show()
It should be noted that one could have just run nn_state.fit(train_samples)
using the default hyperparameters and no training evaluators, which would induce different convergence behavior.
At the end of the training process, the network parameters (the weights, visible biases, and hidden biases) are stored in the ComplexWaveFunction
object. One can save them to a pickle file, which will be called saved_params.pt
, with the following command.
[12]:
nn_state.save("saved_params.pt")
This saves the weights, visible biases and hidden biases as torch tensors under the following keys: weights
, visible_bias
, hidden_bias
.