# Passing model=None to use the default model for evaluating the committees' disagreement
select_ind = QBCStrategy.select(label_ind,
unlab_ind,
model=None,
batch_size=1)
label_ind.update(select_ind)
unlab_ind.difference_update(select_ind)
# Update model and calc performance according to the model you are using
model.fit(X=X[label_ind.index, :], y=y[label_ind.index])
pred = model.predict(X[test_idx, :])
accuracy = alibox.calc_performance_metric(
y_true=y[test_idx],
y_pred=pred,
performance_metric='accuracy_score')
# Save intermediate results to file
st = alibox.State(select_index=select_ind, performance=accuracy)
saver.add_state(st)
saver.save()
# Passing the current progress to stopping criterion object
stopping_criterion.update_information(saver)
# Reset the progress in stopping criterion object
stopping_criterion.reset()
QBC_result.append(copy.deepcopy(saver))
analyser = alibox.get_experiment_analyser(x_axis='num_of_queries')
analyser.add_method(method_name='QBC', method_results=QBC_result)
print(analyser)
analyser.plot_learning_curves(title='Example of AL', std_area=True)
def create_and_implement_strategy(strategy_name, data, labels, queries):
# Keep only the values of data and labels dataframe (Later, we use the global split based on idxs)
X = data.values
y = np.asarray(labels)
toolbox = ToolBox(X=X, y=y, query_type='AllLabels', saving_path='.')
# Create Logistic Regression model ( Default Setting with liblinear solver)
model = toolbox.get_default_model()
# Implement query strategy
uncertainty_strategy = toolbox.get_query_strategy(strategy_name=strategy_name)
# Create array to save the results
examples = []
# Set stopping criterion, we will stop in 1000 labeled examples
stopping_criterion = toolbox.get_stopping_criterion('num_of_queries', queries)
# Get the indexes of the global split
with open("dataset_al", "rb") as f:
train_idx, test_idx, labeled_idx, unlabeled_idx = pickle.load(f)
# Create saver to save the results
saver = StateIO(round=0, train_idx=train_idx,
test_idx=test_idx, init_L=labeled_idx,
init_U=unlabeled_idx, saving_path='.')
# print(train_idx.shape, test_idx.shape)
# Starting with some labeled examples
model.fit(X=X[labeled_idx.index, :], y=y[labeled_idx.index])
y_pred = model.predict(X[test_idx, :])
# Calculate the accuracy of the prediction
accuracy = toolbox.calc_performance_metric(y_true=y[test_idx], y_pred=y_pred, performance_metric='accuracy_score')
# Save accuracy of the prediction
saver.set_initial_point(accuracy)
while not stopping_criterion.is_stop():
# Select example of the unlabeled dataset
example = uncertainty_strategy.select(labeled_idx, unlabeled_idx, model=model, batch_size=1)
# Update the label idxs
labeled_idx.update(example)
unlabeled_idx.difference_update(example)
# Train model for the added example
model.fit(X=X[labeled_idx.index, :], y=y[labeled_idx.index])
y_pred = model.predict(X[test_idx, :])
# Calculate accuracy
accuracy = toolbox.calc_performance_metric(y_true=y[test_idx], y_pred=y_pred,
performance_metric='accuracy_score')
# f1 = alibox.calc_performance_metric(y_true=y[test_idx], y_pred=y_pred, performance_metric='f1_score')
# Save update results
state = toolbox.State(select_index=example, performance=accuracy)
saver.add_state(state)
saver.save()
# Update progress for stopping criterion
stopping_criterion.update_information(saver)
stopping_criterion.reset()
examples.append(copy.deepcopy(saver))
# Uncomment and return in order to save the new active learning dataset
# Save selected x_train examples
X_train = X[labeled_idx, :]
# Save labels for the examples
y_train = y[labeled_idx, :]
# Reshape target
y_train = np.array(y_train).reshape(-1)
# Save to pickle
# with open('qbc_dataset','wb') as f:
# pickle.dump((X_train, y_train), f)
return examples
query_y[select_ins, select_y1] = 0.5
query_y[select_ins, select_y2] = 1
# record results
label_ind.update([(select_ins, select_y1), (select_ins, select_y2)])
unlab_ind.difference_update([(select_ins, select_y1),
(select_ins, select_y2)])
if iter % 5 == 0:
# train/test
X_tr, y_tr, _ = get_Xy_in_multilabel(label_ind, X=X, y=query_y)
model.fit(X=X_tr, y=y_tr)
pres, pred = model.predict(X[test_idx])
perf = alibox.calc_performance_metric(
y_true=mult_y[test_idx],
y_pred=pred,
performance_metric='hamming_loss')
# save
st = alibox.State(select_index=[(select_ins, select_y1),
(select_ins, select_y2)],
performance=perf)
saver.add_state(st)
AURO_results.append(copy.copy(saver))
analyser = alibox.get_experiment_analyser()
analyser.add_method(method_name='AURO', method_results=AURO_results)
analyser.plot_learning_curves()
from sklearn.datasets import load_iris from alipy import ToolBox X, y = load_iris(return_X_y=True) alibox = ToolBox(X=X, y=y, query_type='AllLabels', saving_path='.') # get tools tr, te, lab, unlab = alibox.split_AL() tr0, te0, lab0, unlab0 = alibox.get_split(round=0) oracle = alibox.get_clean_oracle() saver = alibox.get_stateio(round=0) repo = alibox.get_repository(round=0) rand_strategy = alibox.get_query_strategy(strategy_name="QueryInstanceRandom") perf = alibox.calc_performance_metric(y_true=[1], y_pred=[1], performance_metric='accuracy_score') model = alibox.get_default_model() sc = alibox.get_stopping_criterion(stopping_criteria='num_of_queries', value=50) analyser = alibox.get_experiment_analyser(x_axis='num_of_queries') acethread = alibox.get_ace_threading() # data struct defined in alipy ind = alibox.IndexCollection([1, 2, 3]) m_ind = alibox.MultiLabelIndexCollection([(1, 0), (2, )]) st = alibox.State(select_index=[1], performance=perf) # io alibox.save() # al_settings.pkl is the default name. To use another name, please pass a specific file name # to 'saving_path' parameter when initializing the ToolBox object. (e.g., saving_path='./my_file.pkl') alibox = ToolBox.load(path='./al_settings.pkl')
class TorchFold:
def __init__(self, dataset, labels, testset, testlab, model, phase, path,
stopping):
self.dataset = dataset
self.labels = labels
self.testset = testset
self.testlab = testlab
self.model = model
self.phase = phase
self.classes = int(max(labels))
self.alibox = ToolBox(X=dataset,
y=labels,
query_type='AllLabels',
saving_path='./%s' % path)
self.alibox.split_AL(test_ratio=0,
initial_label_rate=0.05,
split_count=1)
self.stopping_criterion = self.alibox.get_stopping_criterion(
stopping[0], value=stopping[1])
self.query_strategy = QueryInstanceUncertainty(
X=dataset, y=labels, measure='least_confident')
# self.query_strategy = QueryInstanceQBC(disagreement='KL_divergence')
self.random = QueryRandom()
self.unc_result = []
self.title = ''
self.acc = []
self.gmeans = []
self.recall = []
self.precision = []
self.specificity = []
self.auc = []
self.f1 = []
self.pos = []
self.neg = []
self.ratio = []
self.loss = []
self.mcc = []
self.path = path
def train(self):
for round in range(1):
try:
os.mkdir('%s/%d' % (self.path, round))
except FileExistsError:
pass
# get data split of one fold
train_idx, test_idx, label_ind, unlab_ind = self.alibox.get_split(
round)
# get intermediate results saver for one fold experiment
saver = self.alibox.get_stateio(round)
# set initial performance point
model = self.model
# print(torch.cuda.current_device())
# print(torch.cuda.device_count(), torch.cuda.is_available())
net = NN.NeuralNetwork(model=model,
num_classes=2,
batch_size=500,
device_ids=[0],
epochs=50)
net.lr_fc = 0.0001
net.initiate(self.dataset[label_ind.index],
self.labels[label_ind.index])
net.predict(self.testset)
pred = net.preds
weight = []
conf_mat = confusion_matrix(y_true=self.testlab, y_pred=pred)
precision = conf_mat[1, 1] / (conf_mat[1, 1] + conf_mat[0, 1])
recall = conf_mat[1, 1] / (conf_mat[1, 1] + conf_mat[1, 0])
specificity = conf_mat[0, 0] / (conf_mat[0, 0] + conf_mat[0, 1])
gmeans = sqrt(recall * specificity)
f1 = metrics.f1_score(y_true=self.testlab, y_pred=pred)
auc = metrics.roc_auc_score(y_true=self.testlab, y_score=pred)
accuracy = self.alibox.calc_performance_metric(
y_true=self.testlab,
y_pred=pred.reshape(list(self.testlab.shape)),
performance_metric='accuracy_score')
self.auc.append(auc)
self.acc.append(accuracy)
self.f1.append(f1)
self.gmeans.append(gmeans)
self.recall.append(recall)
self.precision.append(precision)
self.specificity.append(specificity)
all = len(label_ind) + len(unlab_ind)
lab_init = len(label_ind)
lab = list(self.labels[label_ind.index])
self.pos.append(lab.count(1))
self.neg.append(lab.count(0))
self.ratio.append(lab.count(0) / lab.count(1))
tn, tp, fp, fn = conf_mat[0,
0], conf_mat[1,
1], conf_mat[0,
1], conf_mat[1,
0]
mcc = ((tn * tp) - (fn * fp)) / sqrt(
(tn + fp) * (tn + fn) * (tp + fp) * (tp + fn))
self.mcc.append(mcc)
saver.set_initial_point(gmeans)
iteration = 0
while not self.stopping_criterion.is_stop():
# select subsets of Uind samples according to query strategy
iteration += 1
if self.phase == 'active':
net.predict(self.dataset[unlab_ind.index])
prob_pred = net.probablistic_matrix()
if len(label_ind) < all * 0.3:
if iteration % 10:
select_ind = self.query_strategy.select_by_prediction_mat(
unlabel_index=unlab_ind,
predict=prob_pred,
batch_size=int(lab_init * 0.4))
# batch_size=1)
else:
select_ind = self.random.select(
label_ind,
unlab_ind,
batch_size=int(lab_init * 0.4))
# select_ind = self.random.select(label_ind, unlab_ind, batch_size=1)
else:
select_ind = self.query_strategy.select_by_prediction_mat(
unlabel_index=unlab_ind,
predict=prob_pred,
batch_size=int(len(label_ind) * 0.4))
# batch_size=1)
elif self.phase == 'passive':
if len(label_ind) < all * 0.3:
select_ind = self.random.select(label_ind,
unlab_ind,
batch_size=int(
lab_init * 0.4))
# select_ind = self.random.select(label_ind, unlab_ind, batch_size=1)
else:
select_ind = self.random.select(
label_ind,
unlab_ind,
batch_size=int(len(label_ind) * 0.4))
# select_ind = self.random.select(label_ind, unlab_ind, batch_size=1)
# print(select_ind)
label_ind.update(select_ind)
unlab_ind.difference_update(select_ind)
# update model and calc performance accoding to the updated model
loss = net.train(self.dataset[label_ind.index],
self.labels[label_ind.index])
# if not iteration%2:
net.predict(self.testset)
pred = net.preds
conf_mat = confusion_matrix(y_true=self.testlab, y_pred=pred)
precision = conf_mat[1, 1] / (conf_mat[1, 1] + conf_mat[0, 1])
recall = conf_mat[1, 1] / (conf_mat[1, 1] + conf_mat[1, 0])
specificity = conf_mat[0,
0] / (conf_mat[0, 0] + conf_mat[0, 1])
gmeans = sqrt(recall * specificity)
f1 = metrics.f1_score(y_true=self.testlab, y_pred=pred)
auc = metrics.roc_auc_score(y_true=self.testlab, y_score=pred)
accuracy = self.alibox.calc_performance_metric(
y_true=self.testlab,
y_pred=pred.reshape(list(self.testlab.shape)),
performance_metric='accuracy_score')
self.auc.append(auc)
self.acc.append(accuracy)
self.f1.append(f1)
self.gmeans.append(gmeans)
self.recall.append(recall)
self.precision.append(precision)
self.specificity.append(specificity)
lab = list(self.labels[label_ind.index])
self.pos.append(lab.count(1))
self.neg.append((lab.count(0)))
self.ratio.append(lab.count(0) / lab.count(1))
self.loss.append(loss)
tn, tp, fp, fn = conf_mat[0, 0], conf_mat[1, 1], conf_mat[
0, 1], conf_mat[1, 0]
mcc = ((tn * tp) - (fn * fp)) / sqrt(
(tn + fp) * (tn + fn) * (tp + fp) * (tp + fn))
self.mcc.append(mcc)
# save the results
st = self.alibox.State(select_ind, gmeans)
saver.add_state(st)
saver.save()
self.stopping_criterion.update_information(saver)
lab = list(self.labels[label_ind.index])
print('\n class \n0 and 1\n', lab.count(0), lab.count(1))
print('\n', conf_mat)
torch.save(self.model,
'./%s/%d/model%d' % (self.path, round, iteration))
self.stopping_criterion.reset()
self.unc_result.append(copy.deepcopy(saver))
joblib.dump(self.auc, './%s/%d/auc' % (self.path, round))
joblib.dump(self.acc, './%s/%d/acc' % (self.path, round))
joblib.dump(self.f1, './%s/%d/f1' % (self.path, round))
joblib.dump(self.gmeans, './%s/%d/gmeans' % (self.path, round))
joblib.dump(self.recall, './%s/%d/recall' % (self.path, round))
joblib.dump(self.precision,
'./%s/%d/precision' % (self.path, round))
joblib.dump(self.specificity,
'./%s/%d/specificity' % (self.path, round))
joblib.dump(self.pos, './%s/%d/pos' % (self.path, round))
joblib.dump(self.neg, './%s/%d/neg' % (self.path, round))
joblib.dump(self.ratio, './%s/%d/ratio' % (self.path, round))
joblib.dump(self.mcc, './%s/%d/mcc' % (self.path, round))
self.analyser = self.alibox.get_experiment_analyser(
x_axis='num_of_queries')
self.analyser.add_method(method_name='QueryInstanceUncertaity-lc',
method_results=self.unc_result)
print(self.analyser)
class TorchRegressionFold:
def __init__(self,
dataset,
labels,
testset,
testlab,
model,
phase,
path,
stopping,
measure='nearest_neighbor',
distance='linear'):
self.dataset = dataset
self.labels = labels
self.testset = testset
self.testlab = testlab
self.model = model
self.phase = phase
self.classes = int(max(labels))
self.alibox = ToolBox(X=dataset,
y=np.asarray([0] * len(labels), dtype=np.int),
query_type='AllLabels',
saving_path='./%s' % path)
self.alibox.split_AL(test_ratio=0,
initial_label_rate=0.05,
split_count=1)
self.stopping_criterion = self.alibox.get_stopping_criterion(
stopping[0], value=stopping[1])
self.measure = measure
if measure == 'residue':
self.query_strategy = QueryInstanceResidueRegressor(
X=self.dataset, y=self.labels, distance=distance)
else:
self.query_strategy = QueryInstanceDistribution(measure=measure)
self.random = QueryRandom()
self.unc_result = []
self.title = ''
self.loss = []
self.path = path
self.one = self.two = self.three = self.four = self.five = self.six = None
self.max, self.mae, self.mse, self.evs, self.r2 = [], [], [], [], []
self.sample = []
def train(self):
from sklearn.metrics import (mean_squared_log_error as msle, max_error
as max, mean_absolute_error as mae,
mean_squared_error as mse,
explained_variance_score as evs, r2_score
as r2, mean_tweedie_deviance as tweedie)
for round in range(1):
try:
os.mkdir('%s/%d' % (self.path, round))
except FileExistsError:
pass
# get data split of one fold
train_idx, test_idx, label_ind, unlab_ind = self.alibox.get_split(
round)
# get intermediate results saver for one fold experiment
saver = self.alibox.get_stateio(round)
# set initial performance point
model = self.model
net = NN.NeuralNetworkRegressor(model=model,
batch_size=1,
device_ids=[0],
epochs=50)
net.lr_fc = 0.01
net.initiate(self.dataset[label_ind.index],
self.labels[label_ind.index])
net.predict(self.testset)
pred = net.preds
# evaluation
all = len(label_ind) + len(unlab_ind)
lab_init = len(label_ind)
self.mse.append(mse(self.testlab, pred))
self.mae.append(mae(self.testlab, pred))
self.max.append(max(self.testlab, pred))
self.evs.append(evs(self.testlab, pred))
self.r2.append(r2(self.testlab, pred))
self.sample.append(len(label_ind.index))
saver.set_initial_point(mse(self.testlab, pred))
iteration = 0
while not self.stopping_criterion.is_stop():
# select subsets of Uind samples according to query strategy
iteration += 1
lr_fc = net.lr_fc * (1 - len(label_ind.index) / (all * 1.001))
for p in net.optimizer.param_groups:
p['lr'] = lr_fc
print('learning rate is',
net.optimizer.state_dict()['param_groups'][0]['lr'])
if self.phase == 'active':
if self.measure != 'residue':
net.predict(self.dataset[unlab_ind.index])
else:
net.predict(self.dataset[label_ind])
pred = net.preds
if self.measure == 'distance':
if iteration == 1:
self._update_previous_prediction(pred)
else:
self._update_previous_prediction(
pred, select_ind, unlab_ind_save)
previous = self._get_previous_prediction()
else:
previous = None
if len(label_ind) < all * 0.6:
if iteration % 10:
select_ind = self.query_strategy.select_by_prediction(
unlabel_index=unlab_ind,
predict=pred,
labels=self.labels[label_ind.index],
batch_size=int(lab_init * 1),
X_lab=self.dataset[label_ind.index],
X_unlab=self.dataset[unlab_ind.index],
previous=previous)
else:
select_ind = self.random.select(label_ind,
unlab_ind,
batch_size=int(
lab_init * 1))
else:
select_ind = self.query_strategy.select_by_prediction(
unlabel_index=unlab_ind,
predict=pred,
labels=self.labels[label_ind.index],
batch_size=int(len(label_ind) * 0.3),
X_lab=self.dataset[label_ind.index],
X_unlab=self.dataset[unlab_ind.index],
previous=previous)
elif self.phase == 'passive':
if len(label_ind) < all * 0.6:
select_ind = self.random.select(label_ind,
unlab_ind,
batch_size=int(
lab_init * 1))
# select_ind = self.random.select(label_ind, unlab_ind, batch_size=1)
else:
select_ind = self.random.select(
label_ind,
unlab_ind,
batch_size=int(len(label_ind) * 0.3))
# select_ind = self.random.select(label_ind, unlab_ind, batch_size=1)
# update the datasets and previous prediction
unlab_ind_save = unlab_ind.index
label_ind.update(select_ind)
unlab_ind.difference_update(select_ind)
# update model and calc performance accoding to the updated model
loss = net.train(self.dataset[label_ind.index],
self.labels[label_ind.index])
# if not iteration%2:
net.predict(self.testset)
pred = net.preds
# evaluation
self.mse.append(mse(self.testlab, pred))
self.mae.append(mae(self.testlab, pred))
self.max.append(max(self.testlab, pred))
self.evs.append(evs(self.testlab, pred))
self.r2.append(r2(self.testlab, pred))
self.sample.append(len(label_ind.index))
self.loss.append(loss)
# save the results
st = self.alibox.State(select_ind, mse(self.testlab, pred))
saver.add_state(st)
saver.save()
self.stopping_criterion.update_information(saver)
torch.save(self.model,
'./%s/%d/model%d' % (self.path, round, iteration))
self.stopping_criterion.reset()
self.unc_result.append(copy.deepcopy(saver))
joblib.dump(self.mse, './%s/%d/mse' % (self.path, round))
joblib.dump(self.mae, './%s/%d/mae' % (self.path, round))
joblib.dump(self.max, './%s/%d/max' % (self.path, round))
joblib.dump(self.evs, './%s/%d/evs' % (self.path, round))
joblib.dump(self.r2, './%s/%d/r2' % (self.path, round))
joblib.dump(self.sample, './%s/%d/sample' % (self.path, round))
joblib.dump(self.loss, './%s/%d/loss' % (self.path, round))
joblib.dump(self.testlab, './%s/%d/testlab' % (self.path, round))
joblib.dump(pred, './%s/%d/pred' % (self.path, round))
self.analyser = self.alibox.get_experiment_analyser(
x_axis='num_of_queries')
self.analyser.add_method(
method_name='QueryInstanceDistribution-distance',
method_results=self.unc_result)
print(self.analyser)
def _update_previous_prediction(self, new, selected=None, unlab=None):
if self.six is not None: del_ind = [unlab.index(i) for i in selected]
if self.two is not None: self.one = np.delete(self.two, del_ind)
if self.three is not None: self.two = np.delete(self.three, del_ind)
if self.four is not None: self.three = np.delete(self.four, del_ind)
if self.five is not None: self.four = np.delete(self.five, del_ind)
if self.six is not None: self.five = np.delete(self.six, del_ind)
self.six = new
def _get_previous_prediction(self):
if self.one is not None:
return np.vstack((self.one, self.two, self.three, self.four,
self.five, self.six))
elif self.two is not None:
return np.vstack(
(self.two, self.three, self.four, self.five, self.six))
elif self.three is not None:
return np.vstack((self.three, self.four, self.five, self.six))
