def test_alce_lr_embed5(self):
cost_matrix = np.random.RandomState(1126).rand(3, 3)
np.fill_diagonal(cost_matrix, 0)
ds = Dataset(self.X + self.X_pool,
self.y[:3] + [None for _ in range(len(self.X_pool))])
qs = ALCE(ds,
cost_matrix,
LinearRegression(),
embed_dim=5,
random_state=1126)
qseq = run_qs(ds, qs, self.y_truth, self.quota)
assert_array_equal(qseq,
np.array([16, 63, 34, 122, 38, 35, 17, 24, 43, 18]))
def initialDataSetup(trainFeatures,
trainClasses,
testFeatures,
testClasses,
SNLabel='0'):
"""
Set up ideal labeler.
input: trainFeatures, array - train matrix
trainClasses, list - train labels
testFeatures, array - test (photometric) matrix
testClasses, list - test (photometric) labels
SNLabel, str - SN Ia flag
output: tuple, (train_dataset, fullLabels, labeler)
"""
# Concatenate features
fullFeatures = np.vstack([trainFeatures, testFeatures])
# Include None in place of labels from target sample
partialClasses = np.concatenate([(trainClasses[:,
2] == SNLabel).astype(int),
np.array([None] * testFeatures.shape[0])])
# Complete concatenated labels for train and target samples
fullClasses = np.concatenate([(trainClasses[:, 2] == SNLabel).astype(int),
(testClasses[:, 2] == SNLabel).astype(int)])
# Concatenate labels
fullLabels = np.concatenate([trainClasses, testClasses])
# Concatenated features and class labels with None on target data
train_dataset = Dataset(fullFeatures, partialClasses)
# Define ideal labeler
labeler = IdealLabeler(Dataset(fullFeatures, fullClasses))
return (train_dataset, fullLabels, labeler)
def test_multilabel_with_auxiliary_learner_shlr(self):
trn_ds = Dataset(self.X,
self.y[:5] + [None] * (len(self.y) - 5))
qs = MultilabelWithAuxiliaryLearner(trn_ds,
major_learner=BinaryRelevance(LogisticRegression(solver='liblinear',
multi_class="ovr")),
auxiliary_learner=BinaryRelevance(SVM(gamma="auto")),
criterion='shlr',
b=1.,
random_state=1126)
qseq = run_qs(trn_ds, qs, self.y, self.quota)
assert_array_equal(qseq,
np.array([1258, 805, 459, 550, 783, 964, 736, 1004, 38, 750]))
def split_train_test(n_classes):
from sklearn.datasets import load_digits
n_labeled = 5
digits = load_digits(n_class=n_classes) # consider binary case
X = digits.data
y = digits.target
print(np.shape(X))
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33)
while len(np.unique(y_train[:n_labeled])) < n_classes:
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.33)
trn_ds = Dataset(
X_train,
np.concatenate(
[y_train[:n_labeled], [None] * (len(y_train) - n_labeled)]))
tst_ds = Dataset(X_test, y_test)
return trn_ds, tst_ds, digits
def test_hs_subsampling(self):
ds = Dataset(self.X, self.y[:10] + [None] * (len(self.y) - 10))
sub_qs = UncertaintySampling(ds,
model=SVM(gamma='auto',
decision_function_shape='ovr'))
qs = HS(ds, self.classes, subsample_qs=sub_qs, random_state=1126)
qseq = run_qs(ds, qs, self.y, len(self.y) - 10)
assert_array_equal(
np.concatenate([qseq[:10], qseq[-10:]]),
np.array([
120, 50, 33, 28, 78, 133, 52, 124, 102, 109, 81, 108, 12, 10,
89, 114, 92, 126, 48, 25
]))
def initializeAWTLPool(X_source, y_source, X_target_train, y_target_train, n_labeled, sample_weights):
# create training set by appending the train sample from target to all instances from source
X_train = np.vstack([X_source,X_target_train])
y_train = np.append(y_source,y_target_train)
# train_ds is the whole X_train but only the n_labeled (all source) instances are labeled
train_ds = le.AWTLDataset(X=X_train,y=np.concatenate([y_train[:n_labeled], [None] * (len(y_train) - n_labeled)]), da_weights=sample_weights)
# here we have the fully labeled training set
fully_labeled_trn_ds = Dataset(X=X_train,y=y_train)
return train_ds, fully_labeled_trn_ds
def test_multilabel_with_auxiliary_learner_mmr(self):
trn_ds = Dataset(self.X, self.y[:5] + [None] * (len(self.y) - 5))
qs = MultilabelWithAuxiliaryLearner(
trn_ds,
major_learner=BinaryRelevance(
LogisticRegression(solver='liblinear', multi_class="ovr")),
auxiliary_learner=BinaryRelevance(SVM(gamma="auto")),
criterion='mmr',
random_state=1126)
qseq = run_qs(trn_ds, qs, self.y, self.quota)
assert_array_equal(
qseq,
np.array([1258, 1461, 231, 1198, 1498, 1374, 955, 1367, 265, 144]))
def _E(args):
X, y, qx, clf, label_count, sigma, model = args
sigmoid = lambda x: 1 / (1 + np.exp(-x))
query_point = sigmoid(clf.predict_real([qx]))
feature_count = len(X[0])
ret = 0.0
for i in range(label_count):
clf_ = copy.copy(model)
clf_.train(Dataset(np.vstack((X, [qx])), np.append(y, i)))
PI = sigmoid(clf_.predict_real(np.vstack((X, [qx]))))
ret += query_point[-1][i] * _Phi(sigma, PI[:-1], X, PI[-1], qx,
label_count, feature_count)
return ret
def test_QueryByCommittee(self):
trn_ds = Dataset(
self.X, np.concatenate([self.y[:10], [None] * (len(self.y) - 10)]))
qs = QueryByCommittee(trn_ds,
models=[
LogisticRegression(C=1.0),
LogisticRegression(C=0.01),
LogisticRegression(C=100)
],
random_state=1126)
qseq = run_qs(trn_ds, qs, self.y, self.quota)
assert_array_equal(
qseq, np.array([267, 210, 229, 220, 134, 252, 222, 142, 245, 228]))
def main():
# n_labeled = 10 # number of samples that are initially labeled
# Load dataset
# val_data
# X_test = np.load('../ziqi/train_data_5000.npy')
# y_test_csv = pd.read_csv('../ziqi/train_data_label_5000.csv')
# y_test = np.array([i for i in y_test_csv['label']])
# tst_ds = Dataset(X_test, y_test)
#
# noisy_data
y_csv = pd.read_csv('data_label.csv')
y_test = y_csv['label']
y_test = [i for i in y_test]
y_test = np.array(y_test)
X_pool = np.load('data.npy')
y_pool = [None] * len(X_pool)
pool_ds = Dataset(X_pool, y_pool)
# Comparing UncertaintySampling strategy with entropy uncertainty sampling
model_folder = './model/vgg-x30-9982.pb'
# model predict
print('model predict....')
model = MyModelTF(model_folder, batch_size=256)
score = model.predict(X_pool)
score = pd.DataFrame(score)
score.to_csv('noisy_predict_all.csv', index=False)
# score = np.argmax(score, axis=1)
# all_csv = {'indexs':y_test_csv['filename'], 'pre_label': score, 'name_label': y_test}
# all_csv = pd.DataFrame(all_csv)
# all_csv.to_csv('../ziqi/result/train_predict_result.csv')
# acc_test = model.score(tst_ds)
# print('acc: ', acc_test)
print('make query')
qs = UncertaintySampling(pool_ds,
method='entropy',
model=MyModelTF(model_folder, batch_size=128))
ask_ids, labels = qs.make_query(return_label=True, n_instances=5000)
print('ask_ids : ', np.shape(ask_ids))
csv = {
'indexs': ask_ids,
'pre_label': labels,
'name_label': y_test[ask_ids]
}
csv = pd.DataFrame(csv)
csv.to_csv('./result/sample_result.csv', index=False)
def runrnn(trn_ds, tst_ds, val_ds, lbr, model, quota, best_val, batchsize):
E_in, E_out = [], []
intern = 0
finalnum = 0
print("[Important] Start the RNN Train:")
start_time = time.time()
if quota % batchsize == 0:
intern = int(quota / batchsize)
else:
intern = int(quota / batchsize) + 1
finalnum = int(quota % batchsize)
for t in range(intern):
print("[RNN]this is the " + str(t) + " times to ask")
x_first_train = []
y_first_train = []
scores = model.predict_pro(trn_ds)
unlabeled_entry_ids, X_pool = zip(*trn_ds.get_unlabeled_entries())
if t == intern - 1 and finalnum != 0:
max_n = heapq.nsmallest(finalnum, range(len(scores)), scores.take)
else:
max_n = heapq.nsmallest(batchsize, range(len(scores)), scores.take)
X, _ = zip(*trn_ds.data)
print(max_n)
for ask_id in max_n:
real_id = unlabeled_entry_ids[ask_id]
lb = lbr.label(X[real_id])
trn_ds.update(real_id, lb)
x_first_train.append(X[real_id])
y_first_train.append(lb)
x_first_train = np.array(x_first_train)
y_first_train = np.array(y_first_train)
first_train = Dataset(x_first_train, y_first_train)
best_val = model.retrain(trn_ds, val_ds, best_val, first_train)
# E_in = np.append(E_in, 1 - model.score(trn_ds))
E_out = np.append(E_out, model.score(tst_ds))
# print (E_in)
print(E_out)
E_time = get_time_dif(start_time)
return E_out, E_time
def split_train_test(train_dir, vocab_dir, test_size, n_labeled, wordslength):
#train_dir = './data/labeled1.txt'
#vocab_dir = './data/vocab_yinan_test_rnn4.txt'
if not os.path.exists(vocab_dir):
build_vocab(train_dir,vocab_dir,1000)
categories, cat_to_id = read_category()
words, word_to_id = read_vocab(vocab_dir)
x,y = process_file(train_dir, word_to_id, cat_to_id, wordslength)
# x_rnn, y_rnn = process_file_rnn(train_dir, word_to_id, cat_to_id, 600)
listy = []
for i in range(np.shape(y)[0]):
for j in range(np.shape(y)[1]):
if y[i][j] == 1:
listy.append(j)
listy = np.array(listy)
X_train, X_test, y_train, y_test = \
train_test_split(x, listy, test_size=test_size)
# X_train = X_train[:(n_labeled + 24)]
trn_ds = Dataset(X_train, np.concatenate(
[y_train[:n_labeled], [None] * (len(y_train) - n_labeled)]))
# trn_ds = Dataset(X_train, np.concatenate(
# [y_train[:n_labeled], [None] * (len(y_train) - n_labeled)]))
fully_tst_ds = Dataset(X_test, y_test)
X_val, X_real_test, y_val, y_real_test = \
train_test_split(X_test, y_test, test_size=0.5)
tst_ds = Dataset(X_real_test, y_real_test)
val_ds = Dataset(X_val, y_val)
fully_labeled_trn_ds = Dataset(X_train, y_train)
# print (fully_labeled_trn_ds.get_entries()[0])
return trn_ds, tst_ds, y_train, fully_labeled_trn_ds, fully_tst_ds, val_ds
def test_multilabel_with_auxiliary_learner_hlr(self):
trn_ds = Dataset(self.X,
self.y[:5] + [None] * (len(self.y) - 5))
qs = MultilabelWithAuxiliaryLearner(trn_ds,
major_learner=BinaryRelevance(
LogisticRegression(solver='liblinear',
multi_class="ovr",
random_state=1126)),
auxiliary_learner=BinaryRelevance(SVM(gamma="auto")),
criterion='hlr',
random_state=1126)
qseq = run_qs(trn_ds, qs, self.y, self.quota)
assert_array_equal(qseq,
np.array([701, 1403, 147, 897, 974, 1266, 870, 703, 292, 1146]))
def test_QueryByCommittee(self):
#import ipdb; ipdb.set_trace()
random.seed(1126)
trn_ds = Dataset(
self.X, np.concatenate([self.y[:10], [None] * (len(self.y) - 10)]))
qs = QueryByCommittee(trn_ds,
models=[
LogisticRegression(C=1.0),
LogisticRegression(C=0.01),
LogisticRegression(C=100)
])
qseq = run_qs(trn_ds, qs, self.y, self.quota)
assert_array_equal(
qseq, np.array([11, 207, 101, 30, 116, 108, 83, 172, 211, 42]))
def test_query_by_committee_kl_divergence(self):
trn_ds = Dataset(
self.X, np.concatenate([self.y[:10], [None] * (len(self.y) - 10)]))
qs = QueryByCommittee(trn_ds,
disagreement='kl_divergence',
models=[
LogisticRegression(C=1.0),
LogisticRegression(C=0.01),
LogisticRegression(C=100)
],
random_state=1126)
qseq = run_qs(trn_ds, qs, self.y, self.quota)
assert_array_equal(
qseq, np.array([228, 111, 162, 243, 213, 122, 110, 108, 156, 37]))
def split_train_test(X, y, test_size, n_class):
target = np.unique(y)
# mapping the targets to 0 to n_classes-1
# y = np.array([np.where(target == i)[0][0] for i in data['target']])
X_trn, X_tst, y_trn, y_tst = train_test_split(X, y, test_size=test_size)
X_trn = X_trn[:, 1:]
y_tst = X_tst[:, 0]
X_tst = X_tst[:, 1:]
y_tst = y_tst.astype('int32')
# making sure each class appears ones initially
init_y_ind = np.array(
[np.where(y_trn == i)[0][0] for i in range(len(target))])
y_ind = np.array([i for i in range(len(X_trn)) if i not in init_y_ind])
trn_ds = Dataset(
np.vstack((X_trn[init_y_ind], X_trn[y_ind])),
np.concatenate((y_trn[init_y_ind], [None] * (len(y_ind)))))
tst_ds = Dataset(X_tst, y_tst)
fully_labeled_trn_ds = Dataset(
np.vstack((X_trn[init_y_ind], X_trn[y_ind])),
np.concatenate((y_trn[init_y_ind], y_trn[y_ind])))
cost_matrix = 1.0 * np.ones([len(target), len(target)])
for ii in range(0, len(target)):
for jj in range(0, len(target)):
# cost_matrix[ii, jj] = (ii - jj) * (ii - jj)
cost_matrix[ii, jj] = abs(ii - jj) / n_class
np.fill_diagonal(cost_matrix, 0)
return trn_ds, tst_ds, fully_labeled_trn_ds, cost_matrix
def test_binary_relevance_lr(self):
br = BinaryRelevance(base_clf=LogisticRegression(
solver='liblinear', multi_class="ovr", random_state=1126))
br.train(Dataset(self.X_train, self.Y_train))
br_pred_train = br.predict(self.X_train).astype(int)
br_pred_test = br.predict(self.X_test).astype(int)
br_pred_proba_train = br.predict_proba(self.X_train).astype(float)
br_pred_proba_test = br.predict_proba(self.X_test).astype(float)
for i in range(np.shape(self.Y_train)[1]):
clf = sklearn.linear_model.LogisticRegression(solver='liblinear',
multi_class="ovr",
random_state=1126)
clf.fit(self.X_train, self.Y_train[:, i])
assert_array_almost_equal(
clf.predict(self.X_train).astype(int), br_pred_train[:, i])
assert_array_almost_equal(
clf.predict(self.X_test).astype(int), br_pred_test[:, i])
assert_array_almost_equal(
clf.predict_proba(self.X_train)[:, 1].astype(float),
br_pred_proba_train[:, i].astype(float))
assert_array_almost_equal(
clf.predict_proba(self.X_test)[:, 1].astype(float),
br_pred_proba_test[:, i].astype(float))
self.assertEqual(
np.mean(np.abs(self.Y_test - br_pred_test).mean(axis=1)),
br.score(Dataset(self.X_test, self.Y_test), 'hamming'))
self.assertRaises(
NotImplementedError,
lambda: br.score(Dataset(self.X_test, self.Y_test),
criterion='not_exist'))
def test_query_by_committee_vote(self):
#self.skipTest("In this version we randomize make queries")
trn_ds = Dataset(
self.X, np.concatenate([self.y[:10], [None] * (len(self.y) - 10)]))
qs = QueryByCommittee(trn_ds,
disagreement='vote',
models=[
LogisticRegression(C=1.0),
LogisticRegression(C=0.01),
LogisticRegression(C=100)
],
random_state=1126)
qseq = run_qs(trn_ds, qs, self.y, self.quota)
assert_array_equal(qseq,
np.array([10, 12, 11, 13, 16, 14, 17, 18, 19, 21]))
def test_ActiveLearningByLearning(self):
np.random.seed(1126)
trn_ds = Dataset(
self.X, np.concatenate([self.y[:10], [None] * (len(self.y) - 10)]))
qs = ActiveLearningByLearning(trn_ds,
T=self.quota,
query_strategies=[
UncertaintySampling(
trn_ds,
model=LogisticRegression()),
HintSVM(trn_ds)
],
model=LogisticRegression())
qseq = run_qs(trn_ds, qs, self.y, self.quota)
assert_array_equal(
qseq, np.array([103, 220, 118, 75, 176, 50, 247, 199, 46, 55]))
def test_ActiveLearningByLearning(self):
trn_ds = Dataset(
self.X, np.concatenate([self.y[:10], [None] * (len(self.y) - 10)]))
qs = ActiveLearningByLearning(trn_ds,
T=self.quota,
query_strategies=[
UncertaintySampling(
trn_ds,
model=LogisticRegression()),
HintSVM(trn_ds, random_state=1126)
],
model=LogisticRegression(),
random_state=1126)
qseq = run_qs(trn_ds, qs, self.y, self.quota)
assert_array_equal(
qseq, np.array([173, 103, 133, 184, 187, 147, 251, 83, 93, 33]))
def test_cost_sensitive_random_pair_encoding(self):
trn_ds = Dataset(self.X, self.y[:5] + [None] * (len(self.y) - 5))
model = BinaryRelevance(LogisticRegression(solver='liblinear',
multi_class="ovr"))
base_model = LogisticRegression(
solver='liblinear', multi_class="ovr", random_state=1126)
qs = CostSensitiveReferencePairEncoding(
trn_ds,
scoring_fn=pairwise_f1_score,
model=model,
base_model=base_model,
n_models=10,
n_jobs=1,
random_state=1126)
qseq = run_qs(trn_ds, qs, self.y, self.quota)
assert_array_equal(qseq,
np.array([149, 434, 1126, 719, 983, 564, 816, 732, 101, 1242]))
def test_ALBLTestCase(self):
trn_ds = Dataset(
self.X, np.concatenate([self.y[:10], [None] * (len(self.y) - 10)]))
qs = ActiveLearningByLearning(
trn_ds,
T=self.quota,
query_strategies=[
UncertaintySampling(trn_ds,
model=SVM(kernel="linear",
decision_function_shape="ovr")),
QUIRE(trn_ds),
RandomSampling(trn_ds)
],
model=SVM(kernel="linear", decision_function_shape="ovr"),
random_state=1126)
qseq = run_qs(trn_ds, qs, self.y, self.quota)
assert_array_equal(
qseq, np.array([173, 103, 133, 184, 187, 147, 251, 83, 93, 33]))
def make_query(self):
Xlabeled, y = self.dataset.get_labeled_entries()
Xlabeled = np.array(Xlabeled)
y = list(y)
unlabeled_entry_ids, X_pool = self.dataset.get_unlabeled_entries()
label_count = self.dataset.get_num_of_labels()
clf = copy.copy(self.model)
clf.train(Dataset(Xlabeled, y))
p = Pool(self.n_jobs)
errors = p.map(
_E, [(Xlabeled, y, x, clf, label_count, self.sigma, self.model)
for x in X_pool])
p.terminate()
return unlabeled_entry_ids[errors.index(min(errors))]
def test_density_weighted_meta_uncertainty_lc(self):
trn_ds = Dataset(self.X[:20], np.concatenate([self.y[:6],
[None] * 14]))
base_qs = UncertaintySampling(trn_ds,
method='lc',
model=LogisticRegression(
solver='liblinear',
multi_class="ovr"))
similarity_metric = cosine_similarity
clustering_method = KMeans(n_clusters=3, random_state=1126)
qs = DensityWeightedMeta(dataset=trn_ds,
base_query_strategy=base_qs,
similarity_metric=similarity_metric,
clustering_method=clustering_method,
beta=1.0,
random_state=1126)
model = LogisticRegression(solver='liblinear', multi_class="ovr")
qseq = run_qs(trn_ds, qs, self.y, self.quota)
assert_array_equal(qseq,
np.array([13, 18, 9, 12, 8, 16, 10, 19, 15, 17]))
def train(self, X, y):
self.n_samples = np.shape(X)[0]
self.n_labels = np.shape(y)[1]
score0 = self.scoring_fn(
y, np.tile(self.rep_label[0], (self.n_samples, 1)))
score1 = self.scoring_fn(
y, np.tile(self.rep_label[1], (self.n_samples, 1)))
lbl = (((score1 - score0) > 0) + 0.0)
weight = np.abs(score1 - score0)
if np.sum(weight) > 0:
weight = weight / np.sum(weight) * len(X)
if len(np.unique(lbl)) == 1:
self.label = np.unique(lbl)[0]
self.base_clf_ = None
else:
self.base_clf_ = copy.deepcopy(self.base_clf)
self.base_clf_.train(Dataset(X, lbl), sample_weight=weight)
def make_query(self, n_queries=1, n_jobs=20):
labeled_entries = self.dataset.get_labeled_entries()
Xlabeled, y = zip(*labeled_entries)
Xlabeled = np.array(Xlabeled)
y = list(y)
unlabeled_entries = self.dataset.get_unlabeled_entries()
unlabeled_entry_ids, X_pool = zip(*unlabeled_entries)
label_count = self.dataset.get_num_of_labels()
clf = copy.copy(self.model)
clf.train(Dataset(Xlabeled, y))
p = Pool(n_jobs)
errors = p.map(self.E,
[(Xlabeled, y, x, clf, label_count) for x in X_pool])
p.terminate()
return unlabeled_entry_ids[errors.index(min(errors))]
def test_query_by_committee_vote(self):
trn_ds = Dataset(
self.X, np.concatenate([self.y[:10], [None] * (len(self.y) - 10)]))
qs = QueryByCommittee(trn_ds,
disagreement='vote',
models=[
LogisticRegression(C=1.0,
solver="liblinear",
multi_class="ovr"),
LogisticRegression(C=0.01,
solver="liblinear",
multi_class="ovr"),
LogisticRegression(C=100,
solver="liblinear",
multi_class="ovr")
],
random_state=1126)
qseq = run_qs(trn_ds, qs, self.y, self.quota)
assert_array_equal(
qseq, np.array([267, 210, 229, 220, 134, 252, 222, 142, 245, 228]))
def make_query(self):
dataset = self.dataset
X, Y = dataset.get_labeled_entries()
Y = np.array(Y)
unlabeled_entry_ids, X_pool = dataset.get_unlabeled_entries()
X_pool = np.array(X_pool)
clfs = []
boundaries = []
for i in range(self.n_labels):
if len(np.unique(Y[:, i])) == 1:
clf = DummyClf()
else:
clf = copy.deepcopy(self.base_clf)
clf.train(Dataset(X, Y[:, i]))
boundaries.append(np.abs(clf.predict_real(X_pool)[:, 1]))
clfs.append(clf)
choices = np.where(np.array(boundaries) == np.min(boundaries))[1]
ask_id = self.random_state_.choice(choices)
return unlabeled_entry_ids[ask_id]
def train(self, dataset):
r"""Train model with given feature.
Parameters
----------
X : array-like, shape=(n_samples, n_features)
Train feature vector.
Y : array-like, shape=(n_samples, n_labels)
Target labels.
Attributes
----------
clfs\_ : list of :py:mod:`libact.models` object instances
Classifier instances.
Returns
-------
self : object
Retuen self.
"""
X, Y = dataset.format_sklearn()
X = np.array(X)
Y = np.array(Y)
self.n_labels_ = np.shape(Y)[1]
self.n_features_ = np.shape(X)[1]
self.clfs_ = []
for i in range(self.n_labels_):
# TODO should we handle it here or we should handle it before calling
if len(np.unique(Y[:, i])) == 1:
clf = DummyClf()
else:
clf = copy.deepcopy(self.base_clf)
clf.train(Dataset(X, Y[:, i]))
self.clfs_.append(clf)
return self
def make_query(self):
dataset = self.dataset
X, y = zip(*dataset.get_labeled_entries())
unlabled_entries = dataset.get_unlabeled_entries()
if isinstance(self.random_sampling, int):
unlabled_entries = random.sample(unlabled_entries,
k=self.random_sampling)
elif isinstance(self.random_sampling, float):
unlabled_entries = random.sample(
unlabled_entries,
k=int(len(unlabled_entries) * self.random_sampling))
unlabeled_entry_ids, X_pool = zip(*unlabled_entries)
classes = np.unique(y)
n_classes = len(classes)
self.model.train(dataset)
proba = self.model.predict_proba(X_pool)
scores = []
for i, x in enumerate(X_pool):
score = []
for yi in range(n_classes):
m = copy.deepcopy(self.model)
m.train(Dataset(np.vstack((X, [x])), y + (yi, )))
p = m.predict_proba(X_pool)
if self.loss == '01': # 0/1 loss
score.append(proba[i, yi] * np.sum(1 - np.max(p, axis=1)))
elif self.loss == 'log': # log loss
score.append(proba[i, yi] * -np.sum(p * np.log(p)))
scores.append(np.sum(score))
choices = np.where(np.array(scores) == np.min(scores))[0]
ask_idx = self.random_state_.choice(choices)
return unlabeled_entry_ids[ask_idx]
