def random_input(self, test_size=1):
if not self.one_hot:
X = np.zeros((test_size, len(self.input_features)))
else:
X = np.empty((test_size, len(self.input_features)), dtype=object)
for i, f in enumerate(self.input_features):
if self.feature_types[i] == "CATEGORICAL":
# choose random values for the categorical features
X[:, i] = np.random.choice(self.label_encoders[i].classes_,
size=test_size)
else:
if self.binning:
# choose uniformly random values for the continuous features
X[:, i] = [self.bins[f][j].unif_val()
for j in np.random.choice(len(self.bins[f]),
size=test_size)]
else:
X[:, i] = utils.gen_query_set(1, test_size)[:, 0]
return X
def extract(self,
budget,
steps=[],
adaptive_oracle=False,
baseline=False,
epsilon=1e-2,
reg_lambda=1e-16,
eps_factor=0.99,
epoch=100,
print_epoch=10,
batch_size=1,
num_passes=1000,
random_seed=0):
numpy.random.seed(random_seed)
assert not (adaptive_oracle and steps)
if steps:
step = steps[0]
else:
step = budget
if not adaptive_oracle:
X_ext = utils.gen_query_set(self.num_features(), step)
else:
X_ext = utils.line_search_oracle(self.num_features(), step,
self.query)
model = None
idx = 0
while budget > 0:
idx += 1
budget -= step
y_ext = self.query(X_ext)
if not baseline:
y_ext_p = self.query_probas(X_ext)
else:
num_classes = len(self.get_classes())
y_ext_p = numpy.zeros((len(y_ext), num_classes))
y_ext_p[numpy.arange(len(y_ext)), y_ext] = 1
print y_ext_p
print '{}'.format(-numpy.mean(y_ext_p * numpy.log(y_ext_p)))
model = build_model(X_ext,
y_ext_p,
epsilon=epsilon,
reg_lambda=reg_lambda,
num_passes=num_passes,
eps_factor=eps_factor,
epoch=epoch,
print_epoch=print_epoch,
batch_size=batch_size,
warm_start=model)
mtype = "base" if baseline else "extr"
mode = "adapt-local" if steps \
else "adapt-oracle" if adaptive_oracle \
else "passive"
save(
model, 'experiments/inversion/{}/models/{}_{}_{}.pkl'.format(
self.dataset, mode, mtype, len(X_ext)))
if budget > 0:
step = steps[idx] - steps[idx - 1]
X_local = utils.gen_query_set(n=X_repr.shape[1],
test_size=1000)
Y_local = predict(model, X_local)
assert len(pd.Series(Y_local).unique()) != 1
adaptive_budget = (min(step, budget) * 3) / 4
adaptive_budget += adaptive_budget % 2
random_budget = min(step, budget) - adaptive_budget
predict_func = lambda x: predict(model, x)
samples = utils.line_search(X_local, Y_local,
adaptive_budget / 2, predict_func)
X_random = utils.gen_query_set(X_ext.shape[1], random_budget)
X_ext = numpy.vstack((samples, X_random, X_ext))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('model_file', type=str, help='a pickled model file')
parser.add_argument('action', type=str, help='action to perform')
parser.add_argument('budget', type=str, help='query budget')
parser.add_argument('--num_passes',
type=int,
help='number of passes',
default=1000)
parser.add_argument('--epsilon',
type=float,
help='learning rate',
default=0.1)
parser.add_argument('--rounding', type=int, help='rounding digits')
parser.add_argument('--steps',
type=str,
nargs='+',
default=[],
help='adaptive active learning')
parser.add_argument('--adaptive_oracle',
dest='adaptive_oracle',
action='store_true',
help='adaptive active learning from oracle')
parser.add_argument('--force_reg',
dest='force_reg',
action='store_true',
help='train a regression layer only')
parser.add_argument('--batch_size', type=int, help='batch size', default=1)
parser.add_argument('--seed', type=int, default=0, help='random seed')
args = parser.parse_args()
model_file = args.model_file
action = args.action
budget = args.budget
num_passes = args.num_passes
rounding = args.rounding
steps = args.steps
adaptive_oracle = args.adaptive_oracle
epsilon = args.epsilon
batch_size = args.batch_size
force_reg = args.force_reg
seed = args.seed
np.random.seed(0)
X_train, y_train, X_test, y_test, _ = utils.prepare_data('att_faces')
ext = TheanoMLInversionExtractor(model_file)
num_unknowns = len(ext.get_classes()) * ext.num_features()
try:
budget = int(budget)
except ValueError:
budget = int(float(budget) * num_unknowns)
try:
steps = map(int, steps)
except ValueError:
steps = map(lambda x: int(float(x) * num_unknowns), steps)
print >> sys.stderr, 'Data: {}, Action: {}, Budget:{}, Seed: {}'.\
format(model_file, action, budget, seed)
print >> sys.stderr, 'Number of unknowns: {}'.format(num_unknowns)
if action == "extract":
ext.extract(budget,
steps=steps,
print_epoch=1,
adaptive_oracle=adaptive_oracle,
num_passes=num_passes,
epsilon=epsilon,
batch_size=batch_size,
random_seed=seed)
elif action == "compare":
X_test_u = utils.gen_query_set(X_test.shape[1], 1000)
ext.compare(X_test, X_test_u)
else:
raise ValueError('Unknown action')
def gen_query_set(self, n, test_size, force_input_space=True):
return utils.gen_query_set(n, test_size)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('model_file', type=str, help='a pickled model file')
parser.add_argument('action', type=str, help='action to perform')
parser.add_argument('budget', type=str, help='query budget')
parser.add_argument('--num_passes', type=int, help='number of passes',
default=1000)
parser.add_argument('--epsilon', type=float, help='learning rate',
default=0.1)
parser.add_argument('--rounding', type=int, help='rounding digits')
parser.add_argument('--steps', type=str, nargs='+', default=[],
help='adaptive active learning')
parser.add_argument('--adaptive_oracle', dest='adaptive_oracle',
action='store_true',
help='adaptive active learning from oracle')
parser.add_argument('--force_reg', dest='force_reg',
action='store_true',
help='train a regression layer only')
parser.add_argument('--batch_size', type=int, help='batch size', default=1)
parser.add_argument('--seed', type=int, default=0, help='random seed')
args = parser.parse_args()
model_file = args.model_file
action = args.action
budget = args.budget
num_passes = args.num_passes
rounding = args.rounding
steps = args.steps
adaptive_oracle = args.adaptive_oracle
epsilon = args.epsilon
batch_size = args.batch_size
force_reg = args.force_reg
seed = args.seed
np.random.seed(0)
X_train, y_train, X_test, y_test, _ = utils.prepare_data('att_faces')
ext = TheanoMLInversionExtractor(model_file)
num_unknowns = len(ext.get_classes()) * ext.num_features()
try:
budget = int(budget)
except ValueError:
budget = int(float(budget) * num_unknowns)
try:
steps = map(int, steps)
except ValueError:
steps = map(lambda x: int(float(x) * num_unknowns), steps)
print >> sys.stderr, 'Data: {}, Action: {}, Budget:{}, Seed: {}'.\
format(model_file, action, budget, seed)
print >> sys.stderr, 'Number of unknowns: {}'.format(num_unknowns)
if action == "extract":
ext.extract(budget, steps=steps, print_epoch=1,
adaptive_oracle=adaptive_oracle, num_passes=num_passes,
epsilon=epsilon, batch_size=batch_size, random_seed=seed)
elif action == "compare":
X_test_u = utils.gen_query_set(X_test.shape[1], 1000)
ext.compare(X_test, X_test_u)
else:
raise ValueError('Unknown action')
def extract(self,
X_train,
y_train,
num_repr,
budget,
steps=[],
adaptive_oracle=False,
use_labels_only=False,
gamma=1,
epsilon=1e-2,
reg_lambda=1e-16,
eps_factor=0.99,
epoch=100,
print_epoch=10,
batch_size=1,
num_passes=1000,
random_seed=0):
"""
Extracts a logistic regression multilabl classifier with RBF kernel
for self.query
:param X_train: For plotting
:param y_train: For plotting
:param num_repr: amount of vectors in the kernel base.
:param budget: total amount of queries to the oracle.
:param steps: monotoincally increasing list of queries amounts.
:param adaptive_oracle: whether to query the oracle using line search,
or use only random inputs.
:param use_labels_only: Whether to use only the labels in training.
:param gamma: coefficient for the RBF kernel
:param epsilon: initial learining rate.
:param reg_lambda: regularization coefficient.
:param eps_factor: update factor for the learning rate.
Once in every epoch:
learning_rate *= eps_factor
:param epoch: how many iterations over the whole training data count as a n epoch.
:param print_epoch:
:param batch_size:
:param num_passes: how many iterations over the whole data should we do in each training.
:param random_seed:
:return:
"""
numpy.random.seed(random_seed)
assert not (adaptive_oracle and steps)
if steps:
step = steps[0]
else:
step = budget
if not adaptive_oracle:
X_ext = utils.gen_query_set(self.num_features(), step)
else:
X_ext = utils.line_search_oracle(self.num_features(), step,
self.query)
X_repr = utils.gen_query_set(self.num_features(), num_repr)
model = None
idx = 0
while budget > 0:
idx += 1
budget -= step
y_ext = self.query(X_ext)
if not use_labels_only:
y_ext_p = self.query_probas(X_ext)
else:
# When this a baseline, take the labels from the query result
# as one-hot vectors to be the probability vectors.
num_classes = len(self.get_classes())
y_ext_p = numpy.zeros((len(y_ext), num_classes))
y_ext_p[numpy.arange(len(y_ext)), y_ext] = 1
print y_ext_p
print '{} ({})'.format(
self.calculate_loss(X_ext, y_ext_p, reg_lambda),
self.calculate_loss(X_ext, y_ext_p, 0))
print >> sys.stderr, self.calculate_loss(X_ext, y_ext_p,
reg_lambda)
model = build_model(X_repr,
False,
X_ext,
y_ext_p,
gamma=gamma,
epsilon=epsilon,
reg_lambda=reg_lambda,
num_passes=num_passes,
eps_factor=eps_factor,
epoch=epoch,
print_epoch=print_epoch,
batch_size=batch_size,
warm_start=model)
mtype = "base" if use_labels_only else "extr"
mode = "adapt-local" if steps \
else "adapt-oracle" if adaptive_oracle \
else "passive"
save(
model, 'experiments/KLR/{}/models/{}_{}_{}.pkl'.format(
self.dataset, mode, mtype, len(X_ext)))
if X_train is not None and X_train.shape[1] == 2:
bounds = [-1.1, 1.1, -1.1, 1.1]
filename = 'experiments/KLR/{}/plots/{}_{}_{}_{}_boundary'.\
format(self.dataset, mode, mtype, len(X_ext), random_seed)
utils.plot_decision_boundary(lambda x: predict(model, x),
X_train.values, y_train, bounds,
filename)
filename = 'experiments/KLR/{}/plots/{}_{}_{}_{}_boundary_ext'.\
format(self.dataset, mode, mtype, len(X_ext), random_seed)
utils.plot_decision_boundary(lambda x: predict(model, x),
X_ext, y_ext, bounds, filename)
if budget > 0:
step = steps[idx] - steps[idx - 1]
X_local = utils.gen_query_set(n=X_repr.shape[1],
test_size=1000)
Y_local = predict(model, X_local)
assert len(pd.Series(Y_local).unique()) != 1
adaptive_budget = (min(step, budget) * 3) / 4
adaptive_budget += adaptive_budget % 2
random_budget = min(step, budget) - adaptive_budget
predict_func = lambda x: predict(model, x)
samples = utils.line_search(X_local, Y_local,
adaptive_budget / 2, predict_func)
X_random = utils.gen_query_set(X_ext.shape[1], random_budget)
X_ext = numpy.vstack((samples, X_random, X_ext))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('data', type=str, help='a dataset')
parser.add_argument('action', type=str, help='action to perform')
parser.add_argument('num_repr', type=int, help='number of representers')
parser.add_argument('budget', type=str, help='query budget')
parser.add_argument('--num_passes', type=int, help='number of passes',
default=1000)
parser.add_argument('--rounding', type=int, help='rounding digits')
parser.add_argument('--steps', type=str, nargs='+', default=[],
help='adaptive active learning')
parser.add_argument('--adaptive_oracle', dest='adaptive_oracle',
action='store_true',
help='adaptive active learning from oracle')
parser.add_argument('--gamma', type=float,
help='RBF kernel hyper-parameter')
parser.add_argument('--epsilon', type=float, help='learning rate',
default=0.1)
parser.add_argument('--seed', type=int, default=0, help='random seed')
parser.add_argument('--batch_size', type=int, help='batch size')
args = parser.parse_args()
dataset = args.data
action = args.action
num_repr = args.num_repr
budget = args.budget
num_passes = args.num_passes
rounding = args.rounding
steps = args.steps
adaptive_oracle = args.adaptive_oracle
gamma = args.gamma
epsilon = args.epsilon
seed = args.seed
batch_size = args.batch_size
np.random.seed(0)
X_train, y_train, X_test, y_test, scaler = utils.prepare_data(dataset)
X_test_u = utils.gen_query_set(X_test.shape[1], 1000)
ext = LocalKernelExtractor(dataset, X_train, y_train, rounding=rounding)
num_unknowns = num_repr * X_train.shape[1] + \
len(ext.get_classes()) * (num_repr + 1)
try:
budget = int(budget)
except ValueError:
budget = int(float(budget) * num_unknowns)
try:
steps = map(int, steps)
except ValueError:
steps = map(lambda x: int(float(x) * num_unknowns), steps)
print >> sys.stderr, 'Data: {}, Action: {}, Budget:{}, Seed: {}'.\
format(dataset, action, budget, seed)
print >> sys.stderr, 'Number of unknowns: {}'.format(num_unknowns)
if action == "train":
ext.train(num_repr, X_test, y_test)
elif action == "extract":
if gamma is None:
gamma = ext.get_gamma()
print gamma
ext.extract(X_train, y_train, num_repr, budget, gamma=gamma, steps=steps,
adaptive_oracle=adaptive_oracle, num_passes=num_passes,
epsilon=epsilon, random_seed=seed, batch_size=batch_size)
elif action == "baseline":
if gamma is None:
gamma = ext.get_gamma()
ext.extract(X_train, y_train, num_repr, budget, gamma=gamma, steps=steps,
adaptive_oracle=adaptive_oracle, baseline=True,
num_passes=num_passes, epsilon=epsilon, random_seed=seed,
batch_size=batch_size, reg_lambda=1e-40)
elif action == "compare":
ext.compare(X_test, X_test_u, scaler=None)
def train(self, num_repr, X_test, y_test):
X_train = self.X_train
y_train = self.y_train
y_train_p = np.zeros((len(y_train), len(self.classes)))
y_train_p[np.arange(len(y_train)), y_train] = 1
"""
assert num_repr >= len(self.get_classes())
X_repr_bb = []
class_counter = Counter(y_train)
repr_per_class \
= (num_repr + len(self.get_classes()) - 1) / len(self.get_classes())
for (c, count) in sorted(class_counter.iteritems(), key=lambda _: _[1]):
print c, count, repr_per_class
reprs = X_train.values[np.where(y_train == c)[0][0:repr_per_class]]
X_repr_bb.append(reprs)
X_repr_bb = np.vstack(X_repr_bb)[0:num_repr]
print '{} representers'.format(len(X_repr_bb))
"""
X_repr_bb = X_train[0:num_repr].values
"""
import math
import matplotlib.pyplot as plt
side = math.sqrt(X_repr_bb.shape[1])
plt.figure()
for i in range(len(X_repr_bb)):
plt.subplot(2, len(X_repr_bb)/2, i + 1)
plt.axis('off')
image = X_repr_bb[i, :].reshape((side, side))
plt.imshow(image, cmap=plt.cm.gray_r, interpolation='nearest')
plt.show()
"""
"""
model_bb = krs.build_model(X_repr_bb, False, X_train, y_train_p,
epsilon=epsilon, reg_lambda=1e-4,
num_passes=num_passes, eps_factor=0.99,
epoch=100, print_epoch=10, batch_size=1,
gamma=gamma)
y_pred_bb = krs.predict(model_bb, X_test)
"""
best_model = None
best_acc = 0
best_gamma = None
for gamma in [10**x for x in range(-10, 10)]:
from sklearn.metrics.pairwise import rbf_kernel
from sklearn.linear_model import LogisticRegression
X_train_ker = rbf_kernel(X_train, X_repr_bb, gamma=gamma)
model = LogisticRegression(multi_class='multinomial',
solver='lbfgs').\
fit(X_train_ker, y_train)
y_pred = model.predict(rbf_kernel(X_test, X_repr_bb, gamma=gamma))
acc = accuracy_score(y_test, y_pred)
if acc > best_acc:
best_acc = acc
best_gamma = gamma
best_model = model
W = best_model.coef_.T
b = best_model.intercept_
if len(self.classes) == 2:
W = np.hstack((np.zeros((len(X_repr_bb), 1)), W))
b = np.hstack((0, b))
W = theano.shared(
value=W,
name='W',
borrow=True
)
b = theano.shared(
value=b,
name='b',
borrow=True
)
model_bb = krs.KernelLog(T.matrix('x'), best_gamma, len(self.classes),
X_repr_bb, learn_Y=False, W=W, b=b)
y_pred_bb = krs.predict(model_bb, X_test)
print 'Best Gamma: {}'.format(best_gamma)
print 'Y_test: {}'.format(Counter(y_test))
print 'Y_pred: {}'.format(Counter(y_pred_bb))
acc = accuracy_score(y_test, y_pred_bb)
print 'Training accuracy: {}'.format(acc)
print >> sys.stderr, 'Training accuracy: {}'.format(acc)
X_test_u = utils.gen_query_set(X_test.shape[1], 10000)
y_pred_u = krs.predict(model_bb, X_test_u)
print 'Y_pred_u: {}'.format(Counter(y_pred_u))
if X_train.shape[1] == 2:
bounds = [-1.1, 1.1, -1.1, 1.1]
X_train = X_train.values
utils.plot_decision_boundary(
lambda x: krs.predict(model_bb, x), X_train, y_train, bounds)
krs.save(model_bb,
'experiments/KLR/{}/models/oracle.pkl'.format(self.dataset))
def train(self, num_repr, X_test, y_test):
X_train = self.X_train
y_train = self.y_train
y_train_p = np.zeros((len(y_train), len(self.classes)))
y_train_p[np.arange(len(y_train)), y_train] = 1
"""
assert num_repr >= len(self.get_classes())
X_repr_bb = []
class_counter = Counter(y_train)
repr_per_class \
= (num_repr + len(self.get_classes()) - 1) / len(self.get_classes())
for (c, count) in sorted(class_counter.iteritems(), key=lambda _: _[1]):
print c, count, repr_per_class
reprs = X_train.values[np.where(y_train == c)[0][0:repr_per_class]]
X_repr_bb.append(reprs)
X_repr_bb = np.vstack(X_repr_bb)[0:num_repr]
print '{} representers'.format(len(X_repr_bb))
"""
X_repr_bb = X_train[0:num_repr].values
"""
import math
import matplotlib.pyplot as plt
side = math.sqrt(X_repr_bb.shape[1])
plt.figure()
for i in range(len(X_repr_bb)):
plt.subplot(2, len(X_repr_bb)/2, i + 1)
plt.axis('off')
image = X_repr_bb[i, :].reshape((side, side))
plt.imshow(image, cmap=plt.cm.gray_r, interpolation='nearest')
plt.show()
"""
"""
model_bb = krs.build_model(X_repr_bb, False, X_train, y_train_p,
epsilon=epsilon, reg_lambda=1e-4,
num_passes=num_passes, eps_factor=0.99,
epoch=100, print_epoch=10, batch_size=1,
gamma=gamma)
y_pred_bb = krs.predict(model_bb, X_test)
"""
best_model = None
best_acc = 0
best_gamma = None
for gamma in [10**x for x in range(-10, 10)]:
from sklearn.metrics.pairwise import rbf_kernel
from sklearn.linear_model import LogisticRegression
X_train_ker = rbf_kernel(X_train, X_repr_bb, gamma=gamma)
model = LogisticRegression(multi_class='multinomial',
solver='lbfgs').\
fit(X_train_ker, y_train)
y_pred = model.predict(rbf_kernel(X_test, X_repr_bb, gamma=gamma))
acc = accuracy_score(y_test, y_pred)
if acc > best_acc:
best_acc = acc
best_gamma = gamma
best_model = model
W = best_model.coef_.T
b = best_model.intercept_
if len(self.classes) == 2:
W = np.hstack((np.zeros((len(X_repr_bb), 1)), W))
b = np.hstack((0, b))
W = theano.shared(
value=W,
name='W',
borrow=True
)
b = theano.shared(
value=b,
name='b',
borrow=True
)
model_bb = krs.KernelLog(T.matrix('x'), best_gamma, len(self.classes),
X_repr_bb, learn_kernel_base=False, W=W, b=b)
y_pred_bb = krs.predict(model_bb, X_test)
print 'Best Gamma: {}'.format(best_gamma)
print 'Y_test: {}'.format(Counter(y_test))
print 'Y_pred: {}'.format(Counter(y_pred_bb))
acc = accuracy_score(y_test, y_pred_bb)
print 'Training accuracy: {}'.format(acc)
print >> sys.stderr, 'Training accuracy: {}'.format(acc)
X_test_u = utils.gen_query_set(X_test.shape[1], 10000)
y_pred_u = krs.predict(model_bb, X_test_u)
print 'Y_pred_u: {}'.format(Counter(y_pred_u))
if X_train.shape[1] == 2:
bounds = [-1.1, 1.1, -1.1, 1.1]
X_train = X_train.values
utils.plot_decision_boundary(
lambda x: krs.predict(model_bb, x), X_train, y_train, bounds)
krs.save(model_bb,
'experiments/KLR/{}/models/oracle.pkl'.format(self.dataset))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('data', type=str, help='a dataset')
parser.add_argument('action', type=str, help='action to perform')
parser.add_argument('num_repr', type=int, help='number of representers')
parser.add_argument('budget', type=str, help='query budget')
parser.add_argument('--num_passes', type=int, help='number of passes',
default=1000)
parser.add_argument('--rounding', type=int, help='rounding digits')
parser.add_argument('--steps', type=str, nargs='+', default=[],
help='adaptive active learning')
parser.add_argument('--adaptive_oracle', dest='adaptive_oracle',
action='store_true',
help='adaptive active learning from oracle')
parser.add_argument('--gamma', type=float,
help='RBF kernel hyper-parameter')
parser.add_argument('--epsilon', type=float, help='learning rate',
default=0.1)
parser.add_argument('--seed', type=int, default=0, help='random seed')
parser.add_argument('--batch_size', type=int, help='batch size')
args = parser.parse_args()
dataset = args.data
action = args.action
num_repr = args.num_repr
budget = args.budget
num_passes = args.num_passes
rounding = args.rounding
steps = args.steps
adaptive_oracle = args.adaptive_oracle
gamma = args.gamma
epsilon = args.epsilon
seed = args.seed
batch_size = args.batch_size
np.random.seed(0)
X_train, y_train, X_test, y_test, scaler = utils.prepare_data(dataset)
X_test_u = utils.gen_query_set(X_test.shape[1], 1000)
ext = LocalKernelExtractor(dataset, X_train, y_train, rounding=rounding)
num_unknowns = num_repr * X_train.shape[1] + \
len(ext.get_classes()) * (num_repr + 1)
try:
budget = int(budget)
except ValueError:
budget = int(float(budget) * num_unknowns)
try:
steps = map(int, steps)
except ValueError:
steps = map(lambda x: int(float(x) * num_unknowns), steps)
print >> sys.stderr, 'Data: {}, Action: {}, Budget:{}, Seed: {}'.\
format(dataset, action, budget, seed)
print >> sys.stderr, 'Number of unknowns: {}'.format(num_unknowns)
if action == "train":
ext.train(num_repr, X_test, y_test)
elif action == "extract":
if gamma is None:
gamma = ext.get_gamma()
print gamma
ext.extract(X_train, y_train, num_repr, budget, gamma=gamma, steps=steps,
adaptive_oracle=adaptive_oracle, num_passes=num_passes,
epsilon=epsilon, random_seed=seed, batch_size=batch_size)
elif action == "baseline":
if gamma is None:
gamma = ext.get_gamma()
ext.extract(X_train, y_train, num_repr, budget, gamma=gamma, steps=steps,
adaptive_oracle=adaptive_oracle, use_labels_only=True,
num_passes=num_passes, epsilon=epsilon, random_seed=seed,
batch_size=batch_size, reg_lambda=1e-40)
elif action == "compare":
ext.compare(X_test, X_test_u, scaler=None)
def extract(self,
X_train,
y_train,
num_repr,
budget,
steps=[],
adaptive_oracle=False,
baseline=False,
gamma=1,
epsilon=1e-2,
reg_lambda=1e-16,
eps_factor=0.99,
epoch=100,
print_epoch=10,
batch_size=1,
num_passes=1000,
random_seed=0):
numpy.random.seed(random_seed)
assert not (adaptive_oracle and steps)
if steps:
step = steps[0]
else:
step = budget
if not adaptive_oracle:
X_ext = utils.gen_query_set(self.num_features(), step)
else:
X_ext = utils.line_search_oracle(self.num_features(), step,
self.query)
X_repr = utils.gen_query_set(self.num_features(), num_repr)
model = None
idx = 0
while budget > 0:
idx += 1
budget -= step
y_ext = self.query(X_ext)
if not baseline:
y_ext_p = self.query_probas(X_ext)
else:
num_classes = len(self.get_classes())
y_ext_p = numpy.zeros((len(y_ext), num_classes))
y_ext_p[numpy.arange(len(y_ext)), y_ext] = 1
print y_ext_p
print '{} ({})'.format(
self.calculate_loss(X_ext, y_ext_p, reg_lambda),
self.calculate_loss(X_ext, y_ext_p, 0))
print >> sys.stderr, self.calculate_loss(X_ext, y_ext_p,
reg_lambda)
model = build_model(X_repr,
False,
X_ext,
y_ext_p,
gamma=gamma,
epsilon=epsilon,
reg_lambda=reg_lambda,
num_passes=num_passes,
eps_factor=eps_factor,
epoch=epoch,
print_epoch=print_epoch,
batch_size=batch_size,
warm_start=model)
mtype = "base" if baseline else "extr"
mode = "adapt-local" if steps \
else "adapt-oracle" if adaptive_oracle \
else "passive"
save(
model, 'experiments/KLR/{}/models/{}_{}_{}.pkl'.format(
self.dataset, mode, mtype, len(X_ext)))
if X_train is not None and X_train.shape[1] == 2:
bounds = [-1.1, 1.1, -1.1, 1.1]
filename = 'experiments/KLR/{}/plots/{}_{}_{}_{}_boundary'.\
format(self.dataset, mode, mtype, len(X_ext), random_seed)
utils.plot_decision_boundary(lambda x: predict(model, x),
X_train.values, y_train, bounds,
filename)
filename = 'experiments/KLR/{}/plots/{}_{}_{}_{}_boundary_ext'.\
format(self.dataset, mode, mtype, len(X_ext), random_seed)
utils.plot_decision_boundary(lambda x: predict(model, x),
X_ext, y_ext, bounds, filename)
if budget > 0:
step = steps[idx] - steps[idx - 1]
X_local = utils.gen_query_set(n=X_repr.shape[1],
test_size=1000)
Y_local = predict(model, X_local)
assert len(pd.Series(Y_local).unique()) != 1
adaptive_budget = (min(step, budget) * 3) / 4
adaptive_budget += adaptive_budget % 2
random_budget = min(step, budget) - adaptive_budget
predict_func = lambda x: predict(model, x)
samples = utils.line_search(X_local, Y_local,
adaptive_budget / 2, predict_func)
X_random = utils.gen_query_set(X_ext.shape[1], random_budget)
X_ext = numpy.vstack((samples, X_random, X_ext))
def extract(self, X_train, y_train, budget, steps=[],
adaptive_oracle=False, baseline=False, epsilon=1e-1,
num_passes=1000, reg_lambda=1e-8, eps_factor=0.99, epoch=100,
print_loss=True, print_epoch=10, batch_size=20, random_seed=0):
numpy.random.seed(random_seed)
assert not (adaptive_oracle and steps)
if steps:
step = steps[0]
else:
step = budget
if not adaptive_oracle:
X_ext = utils.gen_query_set(X_train.shape[1], step)
else:
X_ext = utils.line_search_oracle(X_train.shape[1], step, self.query)
idx = 0
while budget > 0:
idx += 1
budget -= step
y_ext = self.query(X_ext)
if not baseline:
y_ext_p = self.query_probas(X_ext)
else:
num_classes = len(self.get_classes())
y_ext_p = numpy.zeros((len(y_ext), num_classes))
y_ext_p[numpy.arange(len(y_ext)), y_ext] = 1
# Loss with correct parameters:
print self.calculate_loss(X_ext, y_ext_p, reg_lambda)
print >> sys.stderr, self.calculate_loss(X_ext, y_ext_p, reg_lambda)
model = build_model(self.hidden_nodes, X_ext, y_ext_p,
epsilon=epsilon, num_passes=num_passes,
reg_lambda=reg_lambda, epoch=epoch,
eps_factor=eps_factor, print_loss=print_loss,
print_epoch=print_epoch, batch_size=batch_size)
m_type = "base" if baseline else "extr"
mode = "adapt-local" if steps \
else "adapt-oracle" if adaptive_oracle \
else "passive"
save(model, 'experiments/{}/models/{}_{}_{}_{}_{}.pkl'.
format(self.dataset, mode, m_type,
self.hidden_nodes, len(X_ext), random_seed))
if X_train is not None and X_train.shape[1] == 2:
bounds = [-1.1, 1.1, -1.1, 1.1]
filename = 'experiments/{}/plots/{}_{}_{}_{}_{}_boundary'.\
format(self.dataset, mode, m_type,
self.hidden_nodes, len(X_ext), random_seed)
utils.plot_decision_boundary(lambda x: predict(model, x),
X_train.values, y_train,
bounds, filename)
filename = 'experiments/{}/plots/{}_{}_{}_{}_{}_boundary_ext'.\
format(self.dataset, mode, m_type,
self.hidden_nodes, len(X_ext), random_seed)
utils.plot_decision_boundary(lambda x: predict(model, x),
X_ext, y_ext, bounds, filename)
if budget > 0:
step = steps[idx] - steps[idx-1]
X_local = utils.gen_query_set(n=X_ext.shape[1], test_size=1000)
Y_local = predict(model, X_local)
assert len(pd.Series(Y_local).unique()) != 1
adaptive_budget = (min(step, budget)*3)/4
adaptive_budget += adaptive_budget % 2
random_budget = min(step, budget) - adaptive_budget
predict_func = lambda x: predict(model, x)
samples = utils.line_search(X_local, Y_local, adaptive_budget/2,
predict_func)
X_random = utils.gen_query_set(X_ext.shape[1], random_budget)
X_ext = numpy.vstack((samples, X_random, X_ext))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('data', type=str, help='a dataset')
parser.add_argument('hidden_nodes', type=int, help='number of hidden nodes')
parser.add_argument('action', type=str, help='action to perform')
parser.add_argument('budget', type=str, help='query budget')
parser.add_argument('--num_passes', type=int, help='number of passes',
default=1000)
parser.add_argument('--epsilon', type=float, help='learning rate',
default=0.1)
parser.add_argument('--rounding', type=int, help='rounding digits')
parser.add_argument('--steps', type=str, nargs='+', default=[],
help='adaptive active learning')
parser.add_argument('--adaptive_oracle', dest='adaptive_oracle',
action='store_true',
help='adaptive active learning from oracle')
parser.add_argument('--force_reg', dest='force_reg',
action='store_true',
help='train a regression layer only')
parser.add_argument('--batch_size', type=int, help='batch size', default=1)
parser.add_argument('--seed', type=int, default=0, help='random seed')
args = parser.parse_args()
dataset = args.data
action = args.action
hidden_nodes = args.hidden_nodes
budget = args.budget
num_passes = args.num_passes
rounding = args.rounding
steps = args.steps
adaptive_oracle = args.adaptive_oracle
epsilon = args.epsilon
batch_size = args.batch_size
force_reg = args.force_reg
seed = args.seed
np.random.seed(0)
X_train, y_train, X_test, y_test, scaler = utils.prepare_data(dataset)
if force_reg:
dataset += "_reg"
ext = LocalPerceptronExtractor(dataset, hidden_nodes, X_train, y_train,
rounding=rounding, force_reg=force_reg)
num_unknowns = hidden_nodes * (X_train.shape[1] + 1) + \
len(ext.get_classes()) * (hidden_nodes + 1)
try:
budget = int(budget)
except ValueError:
budget = int(float(budget) * num_unknowns)
try:
steps = map(int, steps)
except ValueError:
steps = map(lambda x: int(float(x) * num_unknowns), steps)
print >> sys.stderr, 'Data: {}, Action: {}, Budget:{}, Seed: {}'.\
format(dataset, action, budget, seed)
print >> sys.stderr, 'Number of unknowns: {}'.format(num_unknowns)
if action == "train":
ext.train(X_test, y_test, num_passes=num_passes)
elif action == "extract":
ext.extract(X_train, y_train, budget, steps=steps,
adaptive_oracle=adaptive_oracle, num_passes=num_passes,
epsilon=epsilon, batch_size=batch_size, random_seed=seed)
elif action == "baseline":
ext.extract(X_train, y_train, budget, steps=steps,
adaptive_oracle=adaptive_oracle, baseline=True,
num_passes=num_passes, epsilon=epsilon,
batch_size=batch_size, random_seed=seed,
reg_lambda=1e-40)
elif action == "compare":
X_test_u = utils.gen_query_set(X_test.shape[1], 10000)
ext.compare(X_test, X_test_u, force_reg=force_reg)
else:
raise ValueError('Unknown action')