def fit_metric(model_name, subject_data_file, threshold, C, nfolds, normalize,
pca, method, rank):
subject_data = pd.read_table(subject_data_file, sep=" ")
net, layer = model_name
row_labels = eval(open('outputs/{0:s}_row_labels.txt'.format(net)).read())
if not pca:
x = np.load('outputs/{0:s}_{1:s}.npy'.format(net, layer))
else:
x = np.load('outputs/{0:s}_{1:s}_pca.npy'.format(net, layer))
scaler = pp.StandardScaler()
accuracies = []
for n in range(nfolds):
train_x, train_relations, test_x, test_relations = split_data(
subject_data,
x,
row_labels,
objects_in_train=7,
threshold=threshold)
if normalize:
train_x = scaler.fit_transform(train_x)
test_x = scaler.transform(test_x)
if method == 'low_rank':
A, objective, accuracy, converged = metric.learn_low_rank_metric(
train_x,
train_relations,
cost=C,
rank=rank,
method='SLSQP',
tol=1e-6,
verbose=False)
elif method == 'diag':
A, objective, accuracy, converged = metric.learn_diagonal_metric(
train_x,
train_relations,
cost=C,
method='L-BFGS-B',
tol=1e-6,
verbose=False)
test_accuracy = metric.calculate_accuracy(test_x, A, test_relations)
accuracies.append(test_accuracy)
print("{0:35s} ({1:4.2f}): {2:.4f}+-{3:.4f}".format(
net + '_' + layer, threshold, np.mean(accuracies),
np.std(accuracies) * 2))
results = {
'Accuracy': np.mean(accuracies),
'AccuracySD': np.std(accuracies)
}
return results
def test():
subject_data_file = 'HumanPredictions_20151118.txt'
outputs = np.load('outputs/alexnet_prob_pca.npy')
subject_data = pd.read_table(subject_data_file, sep=" ")
row_labels = eval(open('outputs/alexnet_row_labels.txt').read())
C = 1.0
normalize = False
scaler = pp.StandardScaler()
train_accuracies = []
test_accuracies = []
for n in range(20):
print('.'),
train_x, train_rels, test_x, test_rels = split_data(subject_data,
outputs,
row_labels,
objects_in_train=7,
threshold=0.8)
if normalize:
train_x = scaler.fit_transform(train_x)
test_x = scaler.transform(test_x)
"""
A, objective, accuracy, converged = metric.learn_diagonal_metric(train_x, train_rels, cost=C, method='L-BFGS-B',
tol=1e-6, verbose=False)
"""
A, objective, accuracy, converged = metric.learn_low_rank_metric(
train_x,
train_rels,
cost=C,
method='SLSQP',
rank=5,
tol=1e-6,
verbose=False)
test_accuracy = metric.calculate_accuracy(test_x, A, test_rels)
train_accuracies.append(accuracy)
test_accuracies.append(test_accuracy)
print("\nTrain: {0:f}+-{1:f}".format(np.mean(train_accuracies),
np.std(train_accuracies) * 2))
print("Test: {0:f}+-{1:f}".format(np.mean(test_accuracies),
np.std(test_accuracies) * 2))
def fit_metric(model_name, subject_data_file, threshold, C, nfolds, normalize, pca, method, rank):
subject_data = pd.read_table(subject_data_file, sep=" ")
net, layer = model_name
row_labels = eval(open("outputs/{0:s}_row_labels.txt".format(net)).read())
if not pca:
x = np.load("outputs/{0:s}_{1:s}.npy".format(net, layer))
else:
x = np.load("outputs/{0:s}_{1:s}_pca.npy".format(net, layer))
scaler = pp.StandardScaler()
accuracies = []
for n in range(nfolds):
train_x, train_relations, test_x, test_relations = split_data(
subject_data, x, row_labels, objects_in_train=7, threshold=threshold
)
if normalize:
train_x = scaler.fit_transform(train_x)
test_x = scaler.transform(test_x)
if method == "low_rank":
A, objective, accuracy, converged = metric.learn_low_rank_metric(
train_x, train_relations, cost=C, rank=rank, method="SLSQP", tol=1e-6, verbose=False
)
elif method == "diag":
A, objective, accuracy, converged = metric.learn_diagonal_metric(
train_x, train_relations, cost=C, method="L-BFGS-B", tol=1e-6, verbose=False
)
test_accuracy = metric.calculate_accuracy(test_x, A, test_relations)
accuracies.append(test_accuracy)
print(
"{0:35s} ({1:4.2f}): {2:.4f}+-{3:.4f}".format(
net + "_" + layer, threshold, np.mean(accuracies), np.std(accuracies) * 2
)
)
results = {"Accuracy": np.mean(accuracies), "AccuracySD": np.std(accuracies)}
return results
def test():
subject_data_file = "HumanPredictions_20151118.txt"
outputs = np.load("outputs/alexnet_prob_pca.npy")
subject_data = pd.read_table(subject_data_file, sep=" ")
row_labels = eval(open("outputs/alexnet_row_labels.txt").read())
C = 1.0
normalize = False
scaler = pp.StandardScaler()
train_accuracies = []
test_accuracies = []
for n in range(20):
print("."),
train_x, train_rels, test_x, test_rels = split_data(
subject_data, outputs, row_labels, objects_in_train=7, threshold=0.8
)
if normalize:
train_x = scaler.fit_transform(train_x)
test_x = scaler.transform(test_x)
"""
A, objective, accuracy, converged = metric.learn_diagonal_metric(train_x, train_rels, cost=C, method='L-BFGS-B',
tol=1e-6, verbose=False)
"""
A, objective, accuracy, converged = metric.learn_low_rank_metric(
train_x, train_rels, cost=C, method="SLSQP", rank=5, tol=1e-6, verbose=False
)
test_accuracy = metric.calculate_accuracy(test_x, A, test_rels)
train_accuracies.append(accuracy)
test_accuracies.append(test_accuracy)
print("\nTrain: {0:f}+-{1:f}".format(np.mean(train_accuracies), np.std(train_accuracies) * 2))
print("Test: {0:f}+-{1:f}".format(np.mean(test_accuracies), np.std(test_accuracies) * 2))