def __sanity_checks(X, y=None):
if len(np.shape(X)) != 2:
raise ValueError(
f'Dataset X must be array of shape (n_datapoints, 2), was given {np.shape(X)}.'
)
n_datapoints, feature_dim = np.shape(X)
if feature_dim != 2:
raise ValueError(
f'Dataset X must have feature dimension of 2, was given {feature_dim}'
)
if y is not None:
if len(y) != n_datapoints:
raise ValueError(
f'Targets y must be of same length as X. Expected length {n_datapoints}, was given {len(y)}'
)
data_classes = np.unique(y)
if len(data_classes) != 2:
raise ValueError(
f'Currently only binary classification problems are supported!'
)
def __init__(self, embedding, X, y):
"""
Args:
embedding (BaseEmbedding): Instance of BaseEmbedding
X (np.array): Training dataset of shape (n_datapoints, feature_dim)
y (np.array): Training labels of shape (n_datapoints,)
"""
# check if the embedding object is of the correct type
if not isinstance(embedding, BaseEmbedding):
raise ValueError(
'Embedding must be an instance that inherits from BaseEmbedding class.'
)
self.embedding = embedding
# check if the dataset X is of the correct shape
if len(np.shape(X)) != 2:
raise ValueError(
f'Dataset X must be array of shape (n_datapoints, feature_dim), was given {np.shape(X)}.'
)
self.n_datapoints, self.feature_dim = np.shape(X)
# check if the dataset X and the training labels y have the same first dimension
if self.n_datapoints != np.shape(y)[0]:
raise ValueError(
f'Dataset X and training labels y must have the same first dimension. Got {self.n_datapoints} datapoints and {np.shape(y)[0]} labels.'
)
# check if the dataset feature dimension matches the embedding feature dimension
if self.feature_dim != embedding.feature_dim:
raise ValueError(
f'Dataset dimension does not match embedding feature dimension! Expected d={embedding.feature_dim}, was given d={self.feature_dim}.'
)
self.X = np.array(X, dtype=float, requires_grad=False)
self.y = np.array(y, requires_grad=False)
self.data_classes = np.unique(self.y)
self.n_data_classes = len(self.data_classes)
self.class_priors = np.array([
len(X[self.y == data_class]) / self.n_datapoints
for data_class in self.data_classes
])
if self.n_data_classes > 2:
raise NotImplementedError(
'EmbeddingTrainer currently only supports 2-class classification thesis_datasets!'
)
self.X_1 = self.X[self.y == self.data_classes[0]]
self.X_2 = self.X[self.y == self.data_classes[1]]
self.opt = None
def get_cell_centers(cells):
"""Get average coordinates per cell
Args:
cells (ndarray<int>): Cells as computed by `get_cells`
Returns:
centers (dict): Map from cell labels to mean coordinates of cells
"""
centers = {}
for _id in np.unique(cells):
wheres = np.where(cells == _id)
centers[_id] = np.array(
[np.mean(where.astype(float)) for where in wheres])
return centers
def get_cells(vert, horz, iterations=None):
"""Given boolean boundary matrices, obtain cells via spreading iterations
Args:
vert (ndarray<bool>, shape=(m,n-1)): Vertical boundaries
horz (ndarray<bool>, shape=(m-1,n)): Horizontal boundaries
iterations=None (int): Number of spreading iterations. If None, defaults to max(m, n)
Returns:
mat (ndarray<int>, shape=(m,n)): Cells, given as labeling of matrix elements. The labels are contiguous.
"""
num_rows = vert.shape[0] # This is m in the docstring.
num_cols = horz.shape[1] # This is n in the docstring.
if iterations is None:
iterations = max(num_rows, num_cols)
mat = np.arange(num_rows * num_cols, dtype=int).reshape(
(num_rows, num_cols))
for _ in range(iterations):
for i in range(num_rows):
for j in range(num_cols):
nghbhood = [(i, j)]
if j > 0 and not vert[i, j - 1]:
nghbhood.append((i, j - 1))
if j < num_cols - 1 and not vert[i, j]:
nghbhood.append((i, j + 1))
if i > 0 and not horz[i - 1, j]:
nghbhood.append((i - 1, j))
if i < num_rows - 1 and not horz[i, j]:
nghbhood.append((i + 1, j))
nghb_min = np.min([mat[_i, _j] for _i, _j in nghbhood])
for _i, _j in nghbhood:
mat[_i, _j] = nghb_min
_map = {val: count for count, val in enumerate(np.unique(mat))}
for i in range(num_rows):
for j in range(num_cols):
mat[i, j] = _map[mat[i, j]]
return mat
# visual check
plt.scatter(one_samples[0], one_samples[1])
plt.xlim(0, 28)
plt.ylim(0, 28)
plt.show()
# +
# split the two datasets, containing y=0 and y=1 respectively, into 5 datasets:
## distinct_zeros: samples unique to y=0
## distinct_ones: samples unique to y=1
## duplicates: samples that are both in y=0 and y=1
## not_ones: samples that are not in the "unique to y=0" set
## not_zeros: samples that are not in the "unique to y=1" set
zeros = np.unique(np.asarray(zero_samples)[:].T, axis=0)
ones = np.unique(np.asarray(one_samples)[:].T, axis=0)
distinct_zeros = []
distinct_ones = []
duplicates = []
# find unique zeros and duplicates
for sample in zeros:
first_index = np.where(ones[:, 0] == sample[0])
if (len(first_index) > 0):
second_index = np.where(ones[first_index][:, 1] == sample[1])
if (len(second_index[0]) > 0):
duplicates.append(sample)
else:
distinct_zeros.append(sample)
else:
distinct_zeros.append(sample)
def test_projector(self, device, tol, skip_if):
"""Test that a tensor product involving qml.Projector works correctly"""
n_wires = 3
dev = device(n_wires)
if dev.shots is None:
pytest.skip("Device is in analytic mode, cannot test sampling.")
if "Projector" not in dev.observables:
pytest.skip(
"Skipped because device does not support the Projector observable."
)
skip_if(dev, {"supports_tensor_observables": False})
theta = 1.432
phi = 1.123
varphi = -0.543
@qml.qnode(dev)
def circuit(basis_state):
qml.RX(theta, wires=[0])
qml.RX(phi, wires=[1])
qml.RX(varphi, wires=[2])
qml.CNOT(wires=[0, 1])
qml.CNOT(wires=[1, 2])
return qml.sample(
qml.PauliZ(wires=[0]) @ qml.Projector(basis_state,
wires=[1, 2]))
res = circuit([0, 0])
# res should only contain the eigenvalues of the projector matrix tensor product Z, i.e. {-1, 0, 1}
assert np.allclose(sorted(np.unique(res)), [-1, 0, 1], atol=tol(False))
mean = np.mean(res)
expected = (np.cos(varphi / 2) * np.cos(phi / 2) * np.cos(
theta / 2))**2 - (np.cos(varphi / 2) * np.sin(phi / 2) *
np.sin(theta / 2))**2
assert np.allclose(mean, expected, atol=tol(False))
var = np.var(res)
expected = (
(np.cos(varphi / 2) * np.cos(phi / 2) * np.cos(theta / 2))**2 +
(np.cos(varphi / 2) * np.sin(phi / 2) * np.sin(theta / 2))**2 -
((np.cos(varphi / 2) * np.cos(phi / 2) * np.cos(theta / 2))**2 -
(np.cos(varphi / 2) * np.sin(phi / 2) * np.sin(theta / 2))**2)**2)
assert np.allclose(var, expected, atol=tol(False))
res = circuit([0, 1])
assert np.allclose(sorted(np.unique(res)), [-1, 0, 1], atol=tol(False))
mean = np.mean(res)
expected = (np.sin(varphi / 2) * np.cos(phi / 2) * np.cos(
theta / 2))**2 - (np.sin(varphi / 2) * np.sin(phi / 2) *
np.sin(theta / 2))**2
assert np.allclose(mean, expected, atol=tol(False))
var = np.var(res)
expected = (
(np.sin(varphi / 2) * np.cos(phi / 2) * np.cos(theta / 2))**2 +
(np.sin(varphi / 2) * np.sin(phi / 2) * np.sin(theta / 2))**2 -
((np.sin(varphi / 2) * np.cos(phi / 2) * np.cos(theta / 2))**2 -
(np.sin(varphi / 2) * np.sin(phi / 2) * np.sin(theta / 2))**2)**2)
assert np.allclose(var, expected, atol=tol(False))
res = circuit([1, 0])
assert np.allclose(sorted(np.unique(res)), [-1, 0, 1], atol=tol(False))
mean = np.mean(res)
expected = (np.sin(varphi / 2) * np.sin(phi / 2) * np.cos(
theta / 2))**2 - (np.sin(varphi / 2) * np.cos(phi / 2) *
np.sin(theta / 2))**2
assert np.allclose(mean, expected, atol=tol(False))
var = np.var(res)
expected = (
(np.sin(varphi / 2) * np.sin(phi / 2) * np.cos(theta / 2))**2 +
(np.sin(varphi / 2) * np.cos(phi / 2) * np.sin(theta / 2))**2 -
((np.sin(varphi / 2) * np.sin(phi / 2) * np.cos(theta / 2))**2 -
(np.sin(varphi / 2) * np.cos(phi / 2) * np.sin(theta / 2))**2)**2)
assert np.allclose(var, expected, atol=tol(False))
res = circuit([1, 1])
assert np.allclose(sorted(np.unique(res)), [-1, 0, 1], atol=tol(False))
mean = np.mean(res)
expected = (np.cos(varphi / 2) * np.sin(phi / 2) * np.cos(
theta / 2))**2 - (np.cos(varphi / 2) * np.cos(phi / 2) *
np.sin(theta / 2))**2
assert np.allclose(mean, expected, atol=tol(False))
var = np.var(res)
expected = (
(np.cos(varphi / 2) * np.sin(phi / 2) * np.cos(theta / 2))**2 +
(np.cos(varphi / 2) * np.cos(phi / 2) * np.sin(theta / 2))**2 -
((np.cos(varphi / 2) * np.sin(phi / 2) * np.cos(theta / 2))**2 -
(np.cos(varphi / 2) * np.cos(phi / 2) * np.sin(theta / 2))**2)**2)
assert np.allclose(var, expected, atol=tol(False))
def test_hermitian(self, device, tol, skip_if):
"""Test that a tensor product involving qml.Hermitian works correctly"""
n_wires = 3
dev = device(n_wires)
if dev.shots is None:
pytest.skip("Device is in analytic mode, cannot test sampling.")
if "Hermitian" not in dev.observables:
pytest.skip(
"Skipped because device does not support the Hermitian observable."
)
skip_if(dev, {"supports_tensor_observables": False})
theta = 0.432
phi = 0.123
varphi = -0.543
A_ = 0.1 * np.array([
[-6, 2 + 1j, -3, -5 + 2j],
[2 - 1j, 0, 2 - 1j, -5 + 4j],
[-3, 2 + 1j, 0, -4 + 3j],
[-5 - 2j, -5 - 4j, -4 - 3j, -6],
])
@qml.qnode(dev)
def circuit():
qml.RX(theta, wires=[0])
qml.RX(phi, wires=[1])
qml.RX(varphi, wires=[2])
qml.CNOT(wires=[0, 1])
qml.CNOT(wires=[1, 2])
return qml.sample(
qml.PauliZ(wires=[0]) @ qml.Hermitian(A_, wires=[1, 2]))
res = circuit()
# res should only contain the eigenvalues of
# the hermitian matrix tensor product Z
Z = np.diag([1, -1])
eigvals = np.linalg.eigvalsh(np.kron(Z, A_))
assert np.allclose(sorted(np.unique(res)),
sorted(eigvals),
atol=tol(False))
mean = np.mean(res)
expected = (0.1 * 0.5 *
(-6 * np.cos(theta) *
(np.cos(varphi) + 1) - 2 * np.sin(varphi) *
(np.cos(theta) + np.sin(phi) - 2 * np.cos(phi)) +
3 * np.cos(varphi) * np.sin(phi) + np.sin(phi)))
assert np.allclose(mean, expected, atol=tol(False))
var = np.var(res)
expected = (
0.01 *
(1057 - np.cos(2 * phi) + 12 *
(27 + np.cos(2 * phi)) * np.cos(varphi) -
2 * np.cos(2 * varphi) * np.sin(phi) *
(16 * np.cos(phi) + 21 * np.sin(phi)) + 16 * np.sin(2 * phi) - 8 *
(-17 + np.cos(2 * phi) + 2 * np.sin(2 * phi)) * np.sin(varphi) -
8 * np.cos(2 * theta) *
(3 + 3 * np.cos(varphi) + np.sin(varphi))**2 - 24 * np.cos(phi) *
(np.cos(phi) + 2 * np.sin(phi)) * np.sin(2 * varphi) -
8 * np.cos(theta) *
(4 * np.cos(phi) *
(4 + 8 * np.cos(varphi) + np.cos(2 * varphi) -
(1 + 6 * np.cos(varphi)) * np.sin(varphi)) + np.sin(phi) *
(15 + 8 * np.cos(varphi) - 11 * np.cos(2 * varphi) +
42 * np.sin(varphi) + 3 * np.sin(2 * varphi)))) / 16)
assert np.allclose(var, expected, atol=tol(False))
def run_tree_architecture_search(config: dict, dev_type: str):
"""The main workhorse for running the algorithm
Args:
config: Dictionary with configuration parameters for the algorithm. Possible keys are:
- nqubits: Integer. The number of qubits in the circuit
- min_tree_depth: Integer. Minimum circuit depth before we start pruning
- max_tree_depth: Integer. Maximum circuit depth
- prune_rate: Integer. Percentage of nodes that we throw away when we prune
- prune_step: Integer. How often do we prune
- plot_trees: Boolean. Do we want to plot the tree at every depth?
- data_set: String. Which dataset are we learning? Can be 'moons' or 'circles'
- nsteps: Integer. The number of steps for training.
- opt: qml.Optimizer. Pennylane optimizer
- batch_size: Integer. Batch size for training.
- n_samples: Integer. Number of samples that we want to take from the data set.
- learning_rate: Float. Optimizer learning rate.
- save_frequency: Integer. How often do we want to save the tree? Set to 0 for no saving.
- save_path: String. Location to store the data.
Returns:
"""
# build in: circuit type
# if circuit_type=='schuld' use controlled rotation gates and cycle layout for entangling layers
# if circuit_type=='hardware' use minimal gate set and path layout for entangling layers
# Parse configuration parameters.
NQUBITS = config['nqubits']
NSAMPLES = config['n_samples']
PATH = config['save_path']
if dev_type == "local":
dev = qml.device("default.qubit.autograd", wires=NQUBITS)
elif dev_type == "remote":
my_bucket = "amazon-braket-0fc49b964f85" # the name of the bucket
my_prefix = PATH.split('/')[1] # name of the folder in the bucket is the same as experiment name
s3_folder = (my_bucket, my_prefix)
device_arn = "arn:aws:braket:::device/quantum-simulator/amazon/sv1"
dev = qml.device("braket.aws.qubit", device_arn=device_arn, wires=NQUBITS, s3_destination_folder=s3_folder,
parallel=True, max_parallel=10, poll_timeout_seconds=30)
MIN_TREE_DEPTH = config['min_tree_depth']
MAX_TREE_DEPTH = config['max_tree_depth']
SAVE_FREQUENCY = config['save_frequency']
PRUNE_DEPTH_STEP = config['prune_step'] # EVERY ith step is a prune step
PRUNE_RATE = config['prune_rate'] # Percentage of nodes to throw away at each layer
PLOT_INTERMEDIATE_TREES = config['plot_trees']
assert MIN_TREE_DEPTH < MAX_TREE_DEPTH, 'MIN_TREE_DEPTH must be smaller than MAX_TREE_DEPTH'
assert 0.0 < PRUNE_RATE < 1.0, f'The PRUNE_RATE must be between 0 and 1, found {PRUNE_RATE}'
if config['data_set'] == 'circles':
X_train, y_train = datasets.make_circles(n_samples=NSAMPLES, factor=.5, noise=.05)
elif config['data_set'] == 'moons':
X_train, y_train = datasets.make_moons(n_samples=NSAMPLES, noise=.05)
# rescale data to -1 1
X_train = np.multiply(1.0, np.subtract(np.multiply(np.divide(np.subtract(X_train, X_train.min()),
(X_train.max() - X_train.min())), 2.0), 1.0))
if config['readout_layer'] == 'one_hot':
# one hot encode labels
y_train_ohe = np.zeros((y_train.size, y_train.max() + 1))
y_train_ohe[np.arange(y_train.size), y_train] = 1
elif config['readout_layer'] == 'weighted_neuron':
y_train_ohe = y_train
# automatically determine the number of classes
NCLASSES = len(np.unique(y_train))
assert NQUBITS >= NCLASSES, 'The number of qubits must be equal or larger than the number of classes'
save_timing = config.get('save_timing', False)
if save_timing:
print('saving timing info')
import time
# Create a directed graph.
G = nx.DiGraph()
# Add the root
G.add_node("ROOT")
G.nodes['ROOT']["W"] = 0.0
# nx.set_node_attributes(G, {'ROOT': 0.0}, 'W')
# Define allowed layers
ct_ = config.get('circuit_type', None)
if ct_ == 'schuld':
possible_layers = ['ZZ', 'X', 'Y', 'Z']
config['parameterized_gates'] = ['ZZ', 'X', 'Y', 'Z']
if ct_ == 'hardware':
possible_layers = ['hw_CNOT', 'X', 'Y', 'Z']
config['parameterized_gates'] = ['X', 'Y', 'Z']
possible_embeddings = [config['embedding'], ]
assert all([l in string_to_layer_mapping.keys() for l in
possible_layers]), 'No valid mapping from string to function found'
assert all([l in string_to_embedding_mapping.keys() for l in
possible_embeddings]), 'No valid mapping from string to function found'
leaves_at_depth_d = dict(zip(range(MAX_TREE_DEPTH), [[] for _ in range(MAX_TREE_DEPTH)]))
leaves_at_depth_d[0].append('ROOT')
# Iteratively construct tree, pruning at set rate
### PICKLE ALL STUFF FIRST
pickled_data_for_MPI = [NQUBITS, NCLASSES, dev, config, X_train, y_train_ohe]
with open(config['save_path'] + '/MPI_data.pickle', 'wb') as pdata:
pickle.dump(pickled_data_for_MPI, pdata)
for d in range(1, MAX_TREE_DEPTH):
print(f"Depth = {d}")
# Save trees
if (SAVE_FREQUENCY > 0) & ~(d % SAVE_FREQUENCY):
nx.write_gpickle(G, config['save_path'] + f'/tree_depth_{d}.pickle')
# Plot trees
if PLOT_INTERMEDIATE_TREES:
plot_tree(G)
# If we are not passed MIN_TREE_DEPTH, don't prune
if d < MIN_TREE_DEPTH:
# First depth connects to root
if d == 1:
tree_grow_root(G, leaves_at_depth_d, possible_embeddings)
# At the embedding level we don't need to train because there are no params.
for v in leaves_at_depth_d[d]:
G.nodes[v]['W'] = 1.0
print('current graph: ', list(G.nodes(data=True)))
# nx.set_node_attributes(G, {v: 1.0}, 'W')
else:
tree_grow(G, leaves_at_depth_d, d, possible_layers)
best_arch = max(nx.get_node_attributes(G, 'W').items(), key=operator.itemgetter(1))[0]
print('Current best architecture: ', best_arch)
print('max W:', G.nodes[best_arch]['W'])
# For every leaf, create a circuit and run the optimization.
train_all_leaves_parallel(G, leaves_at_depth_d, d, config)
else:
# Check that we are at the correct prune depth step.
if not (d - MIN_TREE_DEPTH) % PRUNE_DEPTH_STEP:
print('Prune Tree')
best_arch = max(nx.get_node_attributes(G, 'W').items(), key=operator.itemgetter(1))[0]
print('Current best architecture: ', best_arch)
print('max W:', G.nodes[best_arch]['W'])
# print(nx.get_node_attributes(G,'W'))
tree_prune(G, leaves_at_depth_d, d, PRUNE_RATE)
print('Grow Pruned Tree')
tree_grow(G, leaves_at_depth_d, d, possible_layers)
# For every leaf, create a circuit and run the optimization.
train_all_leaves_parallel(G, leaves_at_depth_d, d, config)
else:
print('Grow Tree')
best_arch = max(nx.get_node_attributes(G, 'W').items(), key=operator.itemgetter(1))[0]
print('Current best architecture: ', best_arch)
print('max W:', G.nodes[best_arch]['W'])
tree_grow(G, leaves_at_depth_d, d, possible_layers)
train_all_leaves_parallel(G, leaves_at_depth_d, d, config)
best_arch = max(nx.get_node_attributes(G, 'W').items(), key=operator.itemgetter(1))[0]
print('architecture with max W: ', best_arch)
print('max W:', G.nodes[best_arch]['W'])
print('weights: ', G.nodes[best_arch]['weights'])
import pandas as pd
pd.DataFrame.from_dict(nx.get_node_attributes(G, 'W'), orient='index').to_csv('tree_weights.csv')
for i in range(len(a)):
if a[i] * b[i] >= 0:
total_correct += 1
#else:
# print("incorrect")
return (len(a) - total_correct) / len(a)
# ## Init and visualize
# take the dataset, and parse it do that all data is in R
X, Y = create_data()
# PCA
X, Y = prep_data(X, Y, feature_dim)
# check if we still have unique data after dimensionality reduction
if not (len(X) == len(np.unique(X, axis=0))):
print(len(X), len(np.unique(X, axis=0)))
assert False, "DATA NOT UNIQUE, DUPLICATES DETECTED!"
# make sure we have balanced data
X_neg = []
X_pos = []
Y_neg = []
Y_pos = []
for i in range(len(Y)):
if Y[i] < 0:
X_neg.append(X[i])
Y_neg.append(Y[i])
else:
X_pos.append(X[i])
Y_pos.append(Y[i])
