def calculate_metrics(sample_data, probabilities, epoch):
"""Calculates the metrics."""
def convert_bitstring_array_to_probabilities(array):
return probabilities[
data_loader.convert_binary_digits_array_to_bitstrings(array)]
sampled_probabilities = convert_bitstring_array_to_probabilities(
sample_data)
mean_p = np.mean(sampled_probabilities)
logging.info('Min sampled probability %f', np.min(sampled_probabilities))
logging.info('Max sampled probability %f', np.max(sampled_probabilities))
logging.info('Mean sampled probability %f', mean_p)
logging.info('Space size %d', probabilities.size)
fidelity = probabilities.size * mean_p - 1
logistic_fidelity = np.log(probabilities.size) + np.euler_gamma + np.mean(
np.log(sampled_probabilities))
logging.info('Linear Fidelity: %f', fidelity)
logging.info('Logistic Fidelity: %f', logistic_fidelity)
int_array = data_loader.convert_binary_digits_array_to_bitstrings(
sample_data)
values_in_test, counts_in_test = np.unique(int_array, return_counts=True)
test_counts = np.zeros(2**FLAGS.num_qubits)
test_counts[values_in_test] = counts_in_test
chisquare_result = stats.chisquare(f_obs=test_counts,
f_exp=probabilities * len(int_array))
logging.info('chisquare p value: %f', chisquare_result.pvalue)
if FLAGS.save_test_counts:
if FLAGS.num_qubits > 25:
raise ValueError(
f'num_qubits ({FLAGS.num_qubits}) is too large to save test_counts'
)
with tf.io.gfile.GFile(
os.path.join(FLAGS.checkpoint_dir, f'f_obs_{epoch}.npy'),
'wb') as f:
np.save(f, test_counts)
with tf.io.gfile.GFile(
os.path.join(FLAGS.checkpoint_dir, f'f_exp_{epoch}.npy'),
'wb') as f:
np.save(f, probabilities * len(int_array))
train_array = np.zeros((2**FLAGS.num_qubits))
for i in data_loader.convert_binary_digits_array_to_bitstrings(
sample_data):
train_array[i] += 1
# Smooth and normalize
adj_train_array = (train_array + FLAGS.kl_smoothing) / np.sum(train_array)
adj_train_array = adj_train_array / np.sum(adj_train_array)
kl_div = np.sum(scipy.special.kl_div(probabilities, adj_train_array))
logging.info('KL Divergence: %f', kl_div)
return fidelity, logistic_fidelity, kl_div, chisquare_result.pvalue
def convert_bitstring_array_to_probabilities(array):
return probabilities[data_loader.convert_binary_digits_array_to_bitstrings(
array)]
def test_convert_binary_digits_array_to_bitstrings(self):
self.assertListEqual(
list(data_loader.convert_binary_digits_array_to_bitstrings(
np.array([[0, 0, 1, 1], [0, 1, 0, 0], [0, 0, 0, 1]]))),
[3, 4, 1])
