# -*- coding: utf-8 -*-
# This code is part of Qiskit.
#
# (C) Copyright IBM 2019, 2020.
#
# This code is licensed under the Apache License, Version 2.0. You may
# obtain a copy of this license in the LICENSE.txt file in the root directory
# of this source tree or at http://www.apache.org/licenses/LICENSE-2.0.
#
# Any modifications or derivative works of this code must retain this
# copyright notice, and modified files need to carry a notice indicating
# that they have been altered from the originals.
"""Quantum Generative Adversarial Network."""
from typing import Optional, Union
import csv
import os
import logging
import numpy as np
from scipy.stats import entropy
from qiskit.circuit import QuantumCircuit
from qiskit.providers import BaseBackend
from qiskit.aqua import QuantumInstance, AquaError, aqua_globals
from qiskit.aqua.algorithms import QuantumAlgorithm
from qiskit.aqua.components.neural_networks.discriminative_network import DiscriminativeNetwork
from qiskit.aqua.components.neural_networks.generative_network import GenerativeNetwork
from qiskit.aqua.components.neural_networks.quantum_generator import QuantumGenerator
from qiskit.aqua.components.neural_networks.numpy_discriminator import NumPyDiscriminator
from qiskit.aqua.components.optimizers import Optimizer
from qiskit.aqua.components.uncertainty_models import UnivariateVariationalDistribution
from qiskit.aqua.components.uncertainty_models import MultivariateVariationalDistribution
from qiskit.aqua.utils.dataset_helper import discretize_and_truncate
from qiskit.aqua.utils.validation import validate_min
logger = logging.getLogger(__name__)
# pylint: disable=invalid-name
[docs]class QGAN(QuantumAlgorithm):
"""The Quantum Generative Adversarial Network algorithm.
The qGAN [1] is a hybrid quantum-classical algorithm used for generative modeling tasks.
This adaptive algorithm uses the interplay of a generative
:class:`~qiskit.aqua.components.neural_networks.GenerativeNetwork` and a
discriminative :class:`~qiskit.aqua.components.neural_networks.DiscriminativeNetwork`
network to learn the probability distribution underlying given training data.
These networks are trained in alternating optimization steps, where the discriminator tries to
differentiate between training data samples and data samples from the generator and the
generator aims at generating samples which the discriminator classifies as training data
samples. Eventually, the quantum generator learns the training data's underlying probability
distribution. The trained quantum generator loads a quantum state which is a model of the
target distribution.
**References:**
[1] Zoufal et al.,
`Quantum Generative Adversarial Networks for learning and loading random distributions
<https://www.nature.com/articles/s41534-019-0223-2>`_
"""
def __init__(self, data: np.ndarray, bounds: Optional[np.ndarray] = None,
num_qubits: Optional[np.ndarray] = None, batch_size: int = 500,
num_epochs: int = 3000, seed: int = 7,
discriminator: Optional[DiscriminativeNetwork] = None,
generator: Optional[GenerativeNetwork] = None,
tol_rel_ent: Optional[float] = None, snapshot_dir: Optional[str] = None,
quantum_instance: Optional[Union[QuantumInstance, BaseBackend]] = None) -> None:
"""
Args:
data: Training data of dimension k
bounds: k min/max data values [[min_0,max_0],...,[min_k-1,max_k-1]]
if univariate data: [min_0,max_0]
num_qubits: k numbers of qubits to determine representation resolution,
i.e. n qubits enable the representation of 2**n values
[num_qubits_0,..., num_qubits_k-1]
batch_size: Batch size, has a min. value of 1.
num_epochs: Number of training epochs
seed: Random number seed
discriminator: Discriminates between real and fake data samples
generator: Generates 'fake' data samples
tol_rel_ent: Set tolerance level for relative entropy.
If the training achieves relative entropy equal or lower than tolerance it finishes.
snapshot_dir: Directory in to which to store cvs file with parameters,
if None (default) then no cvs file is created.
quantum_instance: Quantum Instance or Backend
Raises:
AquaError: invalid input
"""
validate_min('batch_size', batch_size, 1)
super().__init__(quantum_instance)
if data is None:
raise AquaError('Training data not given.')
self._data = np.array(data)
if bounds is None:
bounds_min = np.percentile(self._data, 5, axis=0)
bounds_max = np.percentile(self._data, 95, axis=0)
bounds = []
for i, _ in enumerate(bounds_min):
bounds.append([bounds_min[i], bounds_max[i]])
if np.ndim(data) > 1:
if len(bounds) != (len(num_qubits) or len(data[0])):
raise AquaError('Dimensions of the data, the length of the data bounds '
'and the numbers of qubits per '
'dimension are incompatible.')
else:
if (np.ndim(bounds) or len(num_qubits)) != 1:
raise AquaError('Dimensions of the data, the length of the data bounds '
'and the numbers of qubits per '
'dimension are incompatible.')
self._bounds = np.array(bounds)
self._num_qubits = num_qubits
# pylint: disable=unsubscriptable-object
if np.ndim(data) > 1:
if self._num_qubits is None:
self._num_qubits = np.ones[len(data[0])] * 3
else:
if self._num_qubits is None:
self._num_qubits = np.array([3])
self._data, self._data_grid, self._grid_elements, self._prob_data = \
discretize_and_truncate(self._data, self._bounds, self._num_qubits,
return_data_grid_elements=True,
return_prob=True, prob_non_zero=True)
self._batch_size = batch_size
self._num_epochs = num_epochs
self._snapshot_dir = snapshot_dir
self._g_loss = []
self._d_loss = []
self._rel_entr = []
self._tol_rel_ent = tol_rel_ent
self._random_seed = seed
if generator is None:
self.set_generator()
else:
self._generator = generator
if discriminator is None:
self.set_discriminator()
else:
self._discriminator = discriminator
self.seed = self._random_seed
self._ret = {}
@property
def seed(self):
""" Returns random seed """
return self._random_seed
@seed.setter
def seed(self, s):
"""
Sets the random seed for QGAN and updates the aqua_globals seed
at the same time
Args:
s (int): random seed
"""
self._random_seed = s
aqua_globals.random_seed = self._random_seed
self._discriminator.set_seed(self._random_seed)
@property
def tol_rel_ent(self):
""" Returns tolerance for relative entropy """
return self._tol_rel_ent
@tol_rel_ent.setter
def tol_rel_ent(self, t):
"""
Set tolerance for relative entropy
Args:
t (float): or None, Set tolerance level for relative entropy.
If the training achieves relative
entropy equal or lower than tolerance it finishes.
"""
self._tol_rel_ent = t
@property
def generator(self):
""" Returns generator """
return self._generator
# pylint: disable=unused-argument
[docs] def set_generator(self, generator_circuit: Optional[Union[QuantumCircuit,
UnivariateVariationalDistribution,
MultivariateVariationalDistribution]
] = None,
generator_init_params: Optional[np.ndarray] = None,
generator_optimizer: Optional[Optimizer] = None):
"""Initialize generator.
Args:
generator_circuit: parameterized quantum circuit which sets
the structure of the quantum generator
generator_init_params: initial parameters for the generator circuit
generator_optimizer: optimizer to be used for the training of the generator
"""
self._generator = QuantumGenerator(self._bounds, self._num_qubits,
generator_circuit, generator_init_params,
self._snapshot_dir)
@property
def discriminator(self):
""" Returns discriminator """
return self._discriminator
[docs] def set_discriminator(self, discriminator=None):
"""
Initialize discriminator.
Args:
discriminator (Discriminator): discriminator
"""
if discriminator is None:
self._discriminator = NumPyDiscriminator(len(self._num_qubits))
else:
self._discriminator = discriminator
self._discriminator.set_seed(self._random_seed)
@property
def g_loss(self):
""" Returns generator loss """
return self._g_loss
@property
def d_loss(self):
""" Returns discriminator loss """
return self._d_loss
@property
def rel_entr(self):
""" Returns relative entropy between target and trained distribution """
return self._rel_entr
[docs] def get_rel_entr(self):
""" Get relative entropy between target and trained distribution """
samples_gen, prob_gen = self._generator.get_output(self._quantum_instance)
temp = np.zeros(len(self._grid_elements))
for j, sample in enumerate(samples_gen):
for i, element in enumerate(self._grid_elements):
if sample == element:
temp[i] += prob_gen[j]
prob_gen = temp
prob_gen = [1e-8 if x == 0 else x for x in prob_gen]
rel_entr = entropy(prob_gen, self._prob_data)
return rel_entr
def _store_params(self, e, d_loss, g_loss, rel_entr):
with open(os.path.join(self._snapshot_dir, 'output.csv'), mode='a') as csv_file:
fieldnames = ['epoch', 'loss_discriminator',
'loss_generator', 'params_generator', 'rel_entropy']
writer = csv.DictWriter(csv_file, fieldnames=fieldnames)
writer.writerow({'epoch': e, 'loss_discriminator': np.average(d_loss),
'loss_generator': np.average(g_loss), 'params_generator':
self._generator.generator_circuit.params, 'rel_entropy': rel_entr})
self._discriminator.save_model(self._snapshot_dir) # Store discriminator model
[docs] def train(self):
"""
Train the qGAN
"""
if self._snapshot_dir is not None:
with open(os.path.join(self._snapshot_dir, 'output.csv'), mode='w') as csv_file:
fieldnames = ['epoch', 'loss_discriminator', 'loss_generator', 'params_generator',
'rel_entropy']
writer = csv.DictWriter(csv_file, fieldnames=fieldnames)
writer.writeheader()
for e in range(self._num_epochs):
aqua_globals.random.shuffle(self._data)
index = 0
while (index + self._batch_size) <= len(self._data):
real_batch = self._data[index: index + self._batch_size]
index += self._batch_size
generated_batch, generated_prob = self._generator.get_output(self._quantum_instance,
shots=self._batch_size)
# 1. Train Discriminator
ret_d = self._discriminator.train([real_batch, generated_batch],
[np.ones(len(real_batch)) / len(real_batch),
generated_prob])
d_loss_min = ret_d['loss']
# 2. Train Generator
self._generator.set_discriminator(self._discriminator)
ret_g = self._generator.train(self._quantum_instance, shots=self._batch_size)
g_loss_min = ret_g['loss']
self._d_loss.append(np.around(float(d_loss_min), 4))
self._g_loss.append(np.around(g_loss_min, 4))
rel_entr = self.get_rel_entr()
self._rel_entr.append(np.around(rel_entr, 4))
self._ret['params_d'] = ret_d['params']
self._ret['params_g'] = ret_g['params']
self._ret['loss_d'] = np.around(float(d_loss_min), 4)
self._ret['loss_g'] = np.around(g_loss_min, 4)
self._ret['rel_entr'] = np.around(rel_entr, 4)
if self._snapshot_dir is not None:
self._store_params(e, np.around(d_loss_min, 4),
np.around(g_loss_min, 4), np.around(rel_entr, 4))
logger.debug('Epoch %s/%s...', e + 1, self._num_epochs)
logger.debug('Loss Discriminator: %s', np.around(float(d_loss_min), 4))
logger.debug('Loss Generator: %s', np.around(g_loss_min, 4))
logger.debug('Relative Entropy: %s', np.around(rel_entr, 4))
if self._tol_rel_ent is not None:
if rel_entr <= self._tol_rel_ent:
break
def _run(self):
"""
Run qGAN training
Returns:
dict: with generator(discriminator) parameters & loss, relative entropy
Raises:
AquaError: invalid backend
"""
if self._quantum_instance.backend_name == ('unitary_simulator' or 'clifford_simulator'):
raise AquaError(
'Chosen backend not supported - '
'Set backend either to statevector_simulator, qasm_simulator'
' or actual quantum hardware')
self.train()
return self._ret