def execute(trainfile, sampler):
print("--- Executing")
print("Using trainfile: ", trainfile)
print("--- Loading (transformed) data")
data = Data.Data()
train_df = data.load(trainfile)
y = train_df["is_attributed"]
X = train_df.drop(["is_attributed"], axis=1)
columns = X.columns.values
before_class_weight = dict(
zip([0, 1], compute_class_weight('balanced', [0, 1], y)))
print("Original weights: ", before_class_weight)
X_resampled = None
y_resampled = None
if sampler == "RANDOM":
oversampler = RandomOverSampler(random_state=0)
oversampler.fit(X, y)
X_resampled, y_resampled = oversampler.sample(X, y)
elif sampler == "ADASYN":
oversampler = ADASYN(random_state=0)
oversampler.fit(X, y)
X_resampled, y_resampled = oversampler.sample(X, y)
elif sampler == "SMOTE":
oversampler = SMOTE(random_state=0)
oversampler.fit(X, y)
X_resampled, y_resampled = oversampler.sample(X, y)
else:
print("Invalid sampler: ", sampler)
after_class_weight = dict(
zip([0, 1], compute_class_weight('balanced', [0, 1], y_resampled)))
print("Sampler: ", sampler, ", weights: ", after_class_weight)
X_resampled = X_resampled.astype(int)
y_resampled = y_resampled.astype(int)
# print("X_resampled: ", X_resampled)
# print("y_resampled: ", y_resampled)
df = pd.DataFrame(data=X_resampled, columns=columns)
df["is_attributed"] = y_resampled
# df["is_attributed"] = df["is_attributed"].astype(int)
compressor = "blosc"
outfilename = trainfile + "." + sampler
print("Output file (over-sampled): ", outfilename)
df.to_hdf(outfilename,
"table",
mode="w",
append=True,
complevel=9,
complib=compressor)
def classify(self):
y_data = self.get_result(self.task.label)
X_data = self.get_result(self.task.features)
y = np.array(y_data.data).ravel()
X = np.array(pd.get_dummies(X_data.data))
#X = MinMaxScaler().fit_transform(X)
X_train = X[:-TILE_SIZE]
y_train = y[:-TILE_SIZE]
X_test = X[-TILE_SIZE:]
y_test = y[-TILE_SIZE:]
cw = compute_class_weight('auto', np.array([0,1]), y)
cw = {0:cw[0],1:cw[1]}
b = get_classifier(self.task.classifier, cw)
b.partial_fit(X_train, y_train, classes=np.array([0,1]))
y_prob = None
y_pred = None
if self.task.classifier in ['perceptron','svm']:
y_pred = b.predict(X_test)
y_prob = np.array([[0,y] for y in y_pred])
else:
y_prob = b.predict_proba(X_test)
y_pred = [1 if t[0] >= 0.5 else 0 for t in y_prob]
cm = confusion_matrix(y_test, y_pred)
stats = classify_stats(cm, y_test, y_prob, TILE_SIZE)
result = ClassifyResult(self.task, 1.0, b, stats)
self.results[self.task.uuid] = result
def class_weight(labels):
y_train = [np.array(x[1].split()).astype(np.int) for x in labels]
y_count = []
for y in y_train:
y_count.extend(y)
cw = compute_class_weight('balanced', np.arange(28), y_count)
return cw
def __init__(self, pattern1train, pattern2train, pattern3train, pattern1test, pattern2test,
pattern3test, y_train, y_test, batch_size=args.batch_size, lr=args.lr, epochs=args.epochs):
input_dim = tf.keras.Input(shape=(1, 250), name='input')
self.pattern1train = pattern1train
self.pattern2train = pattern2train
self.pattern3train = pattern3train
self.pattern1test = pattern1test
self.pattern2test = pattern2test
self.pattern3test = pattern3test
self.y_train = y_train
self.y_test = y_test
self.batch_size = batch_size
self.epochs = epochs
self.class_weight = compute_class_weight(class_weight='balanced', classes=[0, 1], y=y_train)
pattern1vec = tf.keras.layers.Dense(250, activation='relu', name='outputpattern1vec')(input_dim)
pattern2vec = tf.keras.layers.Dense(250, activation='relu', name='outputpattern2vec')(input_dim)
pattern3vec = tf.keras.layers.Dense(250, activation='relu', name='outputpattern3vec')(input_dim)
mergevec = tf.keras.layers.Concatenate(axis=1, name='mergevec')(
[pattern1vec, pattern2vec, pattern3vec]) # concatenate patterns
flattenvec = tf.keras.layers.Flatten(name='flattenvec')(mergevec) # flatten pattern vectors into one vec
finalmergevec = tf.keras.layers.Dense(100, activation='relu', name='outputmergevec')(flattenvec)
prediction = tf.keras.layers.Dense(1, activation='sigmoid', name='output')(finalmergevec)
model = tf.keras.Model(inputs=[input_dim], outputs=[prediction])
adama = tf.keras.optimizers.Adam(lr)
loss = tf.keras.losses.binary_crossentropy
model.compile(optimizer=adama, loss=loss, metrics=['accuracy'])
model.summary()
self.model = model
self.finalmergevec = finalmergevec
def test_auto_weight():
# Test class weights for imbalanced data
from sklearn.linear_model import LogisticRegression
# We take as dataset the two-dimensional projection of iris so
# that it is not separable and remove half of predictors from
# class 1.
# We add one to the targets as a non-regression test: class_weight="balanced"
# used to work only when the labels where a range [0..K).
from sklearn.utils import compute_class_weight
X, y = iris.data[:, :2], iris.target + 1
unbalanced = np.delete(np.arange(y.size), np.where(y > 2)[0][::2])
classes = np.unique(y[unbalanced])
class_weights = compute_class_weight('balanced', classes, y[unbalanced])
assert_true(np.argmax(class_weights) == 2)
for clf in (svm.SVC(kernel='linear'), svm.LinearSVC(random_state=0),
LogisticRegression()):
# check that score is better when class='balanced' is set.
y_pred = clf.fit(X[unbalanced], y[unbalanced]).predict(X)
clf.set_params(class_weight='balanced')
y_pred_balanced = clf.fit(X[unbalanced], y[unbalanced],).predict(X)
assert_true(metrics.f1_score(y, y_pred, average='weighted')
<= metrics.f1_score(y, y_pred_balanced,
average='weighted'))
def classifier(X, Y, clusters):
X_train = X.sample(frac=0.8)
Y_train = Y.loc[X_train.index]
X_val = X.drop(X_train.index)
Y_val = Y.drop(X_train.index)
c_w = compute_class_weight('balanced', np.unique(clusters), clusters)
c_w = dict(enumerate(c_w))
METRICS = [Recall(name='recall'), AUC(name='auc', multi_label=False)]
es = EarlyStopping(monitor='weighted_recall',
mode='max',
verbose=0,
patience=6)
model = Sequential()
model.add(Dense(32, input_dim=X_train.shape[1], activation='relu'))
model.add(Dense(16, activation='relu'))
model.add(Dense(Y_train.shape[1], activation='softmax'))
# Compile model
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=METRICS,
weighted_metrics=METRICS)
model.fit(X_train,
Y_train,
epochs=500,
batch_size=128,
validation_data=(X_val, Y_val),
shuffle=False,
verbose=1,
callbacks=[es],
class_weight=c_w)
return model
def trainModel(self, train, train_labels):
# Set start time
e1 = cv2.getTickCount()
print 'Training Keras DNN ...'
print train.shape
print train_labels.shape
# distribute all the class the same way
numeric_Y = np.dot(train_labels, range(0, self.numberOutput))
classes = np.unique(numeric_Y)
class_weight = compute_class_weight("balanced", classes, numeric_Y)
print 'class weigtht =', class_weight
early_stopping = EarlyStopping(monitor='acc',
patience=5,
verbose=0,
mode='auto')
self.model.fit(train,
train_labels,
nb_epoch=100,
batch_size=32,
shuffle=True,
class_weight=class_weight,
callbacks=[early_stopping])
#self.model.fit(train, train_labels,nb_epoch=50, batch_size=32,shuffle=True,class_weight=class_weight)
# Set end time
e2 = cv2.getTickCount()
time = (e2 - e1) / cv2.getTickFrequency()
print 'Training duration:', time
def classify(self):
y_data = self.get_result(self.task.label)
X_data = self.get_result(self.task.features)
y = np.array(y_data.data).ravel()
X = np.array(pd.get_dummies(X_data.data))
#X = MinMaxScaler().fit_transform(X)
X_train = X[:-TILE_SIZE]
y_train = y[:-TILE_SIZE]
X_test = X[-TILE_SIZE:]
y_test = y[-TILE_SIZE:]
cw = compute_class_weight('auto', np.array([0, 1]), y)
cw = {0: cw[0], 1: cw[1]}
b = get_classifier(self.task.classifier, cw)
b.partial_fit(X_train, y_train, classes=np.array([0, 1]))
y_prob = None
y_pred = None
if self.task.classifier in ['perceptron', 'svm']:
y_pred = b.predict(X_test)
y_prob = np.array([[0, y] for y in y_pred])
else:
y_prob = b.predict_proba(X_test)
y_pred = [1 if t[0] >= 0.5 else 0 for t in y_prob]
cm = confusion_matrix(y_test, y_pred)
stats = classify_stats(cm, y_test, y_prob, TILE_SIZE)
result = ClassifyResult(self.task, 1.0, b, stats)
self.results[self.task.uuid] = result
def weigh_and_show(dataframe):
from sklearn.utils import compute_class_weight, compute_sample_weight
x = compute_class_weight("balanced",sorted(dataframe.class_id.unique()),dataframe.class_id)
# show class weight
for i, proposed_weights in enumerate(x):
# print(f"{df.class_name.unique()[i]}: {x[i]}")
print('{:<25s}: {:<}'.format(disease[i], x[i]))
# build dataframe with list of class_weights, class_name, sum of class names and product of class weights and sum of individual classes
temp_weights = list(x)
# temp_sample_weights = list(x2)
temp_class = list(sorted(dataframe.class_id.unique()))
temp_class_name = disease
temp_sum = list(dataframe.class_id.value_counts()[temp_class])
temp_weight_products = [temp_weights[i] * temp_sum[i] for i in range(len(temp_weights))]
# array check
print(f"temp_weights: {len(temp_weights)}")
# print(f"sample_weights: {len(temp_sample_weights)}")
print(f"class_name: {len(df.class_name.unique())}")
print(f"sum_of_class_names: {len(temp_sum)}")
#build dataframe
temp_dataframe = pd.DataFrame({'class_weights': temp_weights,
# 'sample_weights': temp_sample_weights,
'class_name': temp_class,
'sum_of_class_names': temp_sum,
'weight_products': temp_weight_products,})
return temp_dataframe
def make_predictions(x_train, x_test, y_train, y_test):
poly = PolynomialFeatures(interaction_only=False, include_bias=False)
x_train = poly.fit_transform(x_train)
# Let's target class imbalance problem.
class_weights = compute_class_weight('balanced', np.unique(y_train),
y_train)
class_weights_dict = {
np.unique(y_train)[i]: w
for i, w in enumerate(class_weights)
}
print(class_weights_dict)
clf = LogisticRegression(penalty='l2',
random_state=222,
solver='newton-cg',
C=999999,
class_weight=class_weights_dict,
multi_class='multinomial').fit(x_train, y_train)
IDs = x_test.index
x_test = poly.fit_transform(x_test)
y_pred = clf.predict(x_test)
print('Accuracy of logistic regression classifier on test set: {:.2f}'.
format(clf.score(x_test, y_test)))
confusion_matrix_ = confusion_matrix(y_test, y_pred)
print('Confusion matrix:')
print(confusion_matrix_)
testres = pd.DataFrame(np.column_stack([IDs.values, y_pred]))
testres.columns = ['ID', 'conicSection']
return (testres)
def run(self):
# Define generators
data = EMNIST('../data/emnist/')
train = SimnetGenerator(data.get_training_batch, data.get_sizes()[0])
val = SimnetGenerator(data.get_validation_batch, data.get_sizes()[1])
#class weights
class_weights = compute_class_weight('balanced',
np.unique(data._test_labels),
data._test_labels)
with tf.Session() as sess:
# Define models
simnet = Simnet()
# Do fitting
train_history, val_history = simnet.fit(sess, train, val, 100,
1024, 100)
acc, avg_acc, weighted_acc = simnet.evaluate_special(
sess,
data.get_test_batch,
1024,
data.get_classification_samples,
data.get_sizes()[2],
class_weights=class_weights)
print("Test ACC: ", acc, " TEST AVG ACC: ", avg_acc,
"Weightes AVG ACC:", weighted_acc)
def __init__(self, x_train, y_train, x_test, y_test, batch_size, epochs,
dropout, lr, name):
self.x_train = x_train
self.x_test = x_test
self.y_train = y_train
self.y_test = y_test
self.name = name
self.lr = lr
self.batch_size = batch_size
self.epochs = epochs
self.class_weight = compute_class_weight(class_weight='balanced',
classes=np.arange(10),
y=y_train.argmax(axis=1))
model = Sequential()
model.add(
Bidirectional(LSTM(300, return_sequences=True),
input_shape=(n_steps, dim_input)))
model.add(AttentionWithContext())
model.add(Addition())
model.add(Dense(300))
model.add(LeakyReLU())
model.add(Dropout(dropout))
model.add(Dense(300))
model.add(LeakyReLU())
model.add(Dropout(dropout))
model.add(Dense(10, activation='softmax'))
# Lower learning rate to prevent divergence
adamax = Adamax(self.lr)
model.compile(adamax, 'categorical_crossentropy', metrics=['accuracy'])
self.model = model
def test_multiclass_classifier_class_weight():
"""tests multiclass with classweights for each class"""
alpha = .1
n_samples = 20
tol = .00001
max_iter = 50
class_weight = {0: .45, 1: .55, 2: .75}
fit_intercept = True
X, y = make_blobs(n_samples=n_samples, centers=3, random_state=0,
cluster_std=0.1)
step_size = get_step_size(X, alpha, fit_intercept, classification=True)
classes = np.unique(y)
clf1 = LogisticRegression(solver='sag', C=1. / alpha / n_samples,
max_iter=max_iter, tol=tol, random_state=77,
fit_intercept=fit_intercept,
class_weight=class_weight)
clf2 = clone(clf1)
clf1.fit(X, y)
clf2.fit(sp.csr_matrix(X), y)
le = LabelEncoder()
class_weight_ = compute_class_weight(class_weight, np.unique(y), y)
sample_weight = class_weight_[le.fit_transform(y)]
coef1 = []
intercept1 = []
coef2 = []
intercept2 = []
for cl in classes:
y_encoded = np.ones(n_samples)
y_encoded[y != cl] = -1
spweights1, spintercept1 = sag_sparse(X, y_encoded, step_size, alpha,
n_iter=max_iter, dloss=log_dloss,
sample_weight=sample_weight)
spweights2, spintercept2 = sag_sparse(X, y_encoded, step_size, alpha,
n_iter=max_iter, dloss=log_dloss,
sample_weight=sample_weight,
sparse=True)
coef1.append(spweights1)
intercept1.append(spintercept1)
coef2.append(spweights2)
intercept2.append(spintercept2)
coef1 = np.vstack(coef1)
intercept1 = np.array(intercept1)
coef2 = np.vstack(coef2)
intercept2 = np.array(intercept2)
for i, cl in enumerate(classes):
assert_array_almost_equal(clf1.coef_[i].ravel(),
coef1[i].ravel(),
decimal=2)
assert_almost_equal(clf1.intercept_[i], intercept1[i], decimal=1)
assert_array_almost_equal(clf2.coef_[i].ravel(),
coef2[i].ravel(),
decimal=2)
assert_almost_equal(clf2.intercept_[i], intercept2[i], decimal=1)
def __init__(self, data, name="", batch_size=args.batch_size, lr=args.lr, epochs=args.epochs, dropout=args.dropout):
vectors = np.stack(data.iloc[:, 1].values)
labels = data.iloc[:, 0].values
positive_idxs = np.where(labels == 1)[0]
negative_idxs = np.where(labels == 0)[0]
undersampled_negative_idxs = np.random.choice(negative_idxs, len(positive_idxs), replace=False)
resampled_idxs = np.concatenate([positive_idxs, undersampled_negative_idxs])
x_train, x_test, y_train, y_test = train_test_split(vectors[resampled_idxs], labels[resampled_idxs],
test_size=0.2, stratify=labels[resampled_idxs])
self.x_train = x_train
self.x_test = x_test
self.y_train = to_categorical(y_train)
self.y_test = to_categorical(y_test)
self.name = name
self.batch_size = batch_size
self.epochs = epochs
self.class_weight = compute_class_weight(class_weight='balanced', classes=[0, 1], y=labels)
model = Sequential()
model.add(LSTM(300, input_shape=(vectors.shape[1], vectors.shape[2])))
model.add(Dense(300))
model.add(LeakyReLU())
model.add(Dropout(dropout))
model.add(Dense(300))
model.add(LeakyReLU())
model.add(Dropout(dropout))
model.add(Dense(2, activation='softmax'))
# Lower learning rate to prevent divergence
adamax = Adamax(lr)
model.compile(adamax, 'categorical_crossentropy', metrics=['accuracy'])
self.model = model
def initialize_labels(self, Y):
y_nodes_flat = [y_val for y in Y for y_val in y.nodes]
y_links_flat = [y_val for y in Y for y_val in y.links]
self.prop_encoder_ = LabelEncoder().fit(y_nodes_flat)
self.link_encoder_ = LabelEncoder().fit(y_links_flat)
self.n_prop_states = len(self.prop_encoder_.classes_)
self.n_link_states = len(self.link_encoder_.classes_)
self.prop_cw_ = np.ones_like(self.prop_encoder_.classes_,
dtype=np.double)
self.link_cw_ = compute_class_weight(self.class_weight,
self.link_encoder_.classes_,
y_links_flat)
self.link_cw_ /= self.link_cw_.min()
logging.info('Setting node class weights {}'.format(", ".join(
"{}: {}".format(lbl, cw)
for lbl, cw in zip(self.prop_encoder_.classes_, self.prop_cw_))))
logging.info('Setting link class weights {}'.format(", ".join(
"{}: {}".format(lbl, cw)
for lbl, cw in zip(self.link_encoder_.classes_, self.link_cw_))))
def test_auto_weight():
# Test class weights for imbalanced data
from sklearn.linear_model import LogisticRegression
# We take as dataset the two-dimensional projection of iris so
# that it is not separable and remove half of predictors from
# class 1.
# We add one to the targets as a non-regression test:
# class_weight="balanced"
# used to work only when the labels where a range [0..K).
from sklearn.utils import compute_class_weight
X, y = iris.data[:, :2], iris.target + 1
unbalanced = np.delete(np.arange(y.size), np.where(y > 2)[0][::2])
classes = np.unique(y[unbalanced])
class_weights = compute_class_weight('balanced', classes=classes,
y=y[unbalanced])
assert np.argmax(class_weights) == 2
for clf in (svm.SVC(kernel='linear'), svm.LinearSVC(random_state=0),
LogisticRegression()):
# check that score is better when class='balanced' is set.
y_pred = clf.fit(X[unbalanced], y[unbalanced]).predict(X)
clf.set_params(class_weight='balanced')
y_pred_balanced = clf.fit(X[unbalanced], y[unbalanced],).predict(X)
assert (metrics.f1_score(y, y_pred, average='macro')
<= metrics.f1_score(y, y_pred_balanced,
average='macro'))
def classifier_training_example(mels_path: str) -> None:
batch_size = 64
epochs = 50
model_name = 'class_cnn3'
patience = 10
val_split = 0.25
classifier = build_cnn3_classifier()
classifier.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['acc'])
classifier.summary()
mels_npz = np.load(mels_path)
samples_x = mels_npz['samples_x']
samples_y = mels_npz['samples_y']
log.info(f'Training samples x shape = {samples_x.shape}')
log.info(f'Training samples y shape = {samples_y.shape}')
class_weights = compute_class_weight('balanced',
classes=np.unique(
np.argmax(samples_y, axis=-1)),
y=np.argmax(samples_y, axis=-1))
log.info(f'Training samples class weights = {class_weights}')
class_weights = {idx: w for idx, w in enumerate(class_weights)}
train_drum_classifier(classifier,
x=samples_x,
y=samples_y,
model_name=model_name,
batch_size=batch_size,
epochs=epochs,
val_split=val_split,
patience=patience,
class_weights=class_weights)
def test_auto_weight():
"""Test class weights for imbalanced data"""
from sklearn.linear_model import LogisticRegression
# we take as dataset a the two-dimensional projection of iris so
# that it is not separable and remove half of predictors from
# class 1
from sklearn.utils import compute_class_weight
X, y = iris.data[:, :2], iris.target
unbalanced = np.delete(np.arange(y.size), np.where(y > 1)[0][::2])
classes = np.unique(y[unbalanced])
class_weights = compute_class_weight('auto', classes, y[unbalanced])
assert_true(np.argmax(class_weights) == 2)
for clf in (svm.SVC(kernel='linear'), svm.LinearSVC(random_state=0),
LogisticRegression()):
# check that score is better when class='auto' is set.
y_pred = clf.fit(X[unbalanced], y[unbalanced]).predict(X)
clf.set_params(class_weight='auto')
y_pred_balanced = clf.fit(
X[unbalanced],
y[unbalanced],
).predict(X)
assert_true(
metrics.f1_score(y, y_pred) <= metrics.f1_score(
y, y_pred_balanced))
def test_multiclass_classifier_class_weight():
"""tests multiclass with classweights for each class"""
alpha = .1
n_samples = 20
tol = .00001
max_iter = 50
class_weight = {0: .45, 1: .55, 2: .75}
fit_intercept = True
X, y = make_blobs(n_samples=n_samples, centers=3, random_state=0,
cluster_std=0.1)
step_size = get_step_size(X, alpha, fit_intercept, classification=True)
classes = np.unique(y)
clf1 = LogisticRegression(solver='sag', C=1. / alpha / n_samples,
max_iter=max_iter, tol=tol, random_state=77,
fit_intercept=fit_intercept,
class_weight=class_weight)
clf2 = clone(clf1)
clf1.fit(X, y)
clf2.fit(sp.csr_matrix(X), y)
le = LabelEncoder()
class_weight_ = compute_class_weight(class_weight, np.unique(y), y)
sample_weight = class_weight_[le.fit_transform(y)]
coef1 = []
intercept1 = []
coef2 = []
intercept2 = []
for cl in classes:
y_encoded = np.ones(n_samples)
y_encoded[y != cl] = -1
spweights1, spintercept1 = sag_sparse(X, y_encoded, step_size, alpha,
n_iter=max_iter, dloss=log_dloss,
sample_weight=sample_weight)
spweights2, spintercept2 = sag_sparse(X, y_encoded, step_size, alpha,
n_iter=max_iter, dloss=log_dloss,
sample_weight=sample_weight,
sparse=True)
coef1.append(spweights1)
intercept1.append(spintercept1)
coef2.append(spweights2)
intercept2.append(spintercept2)
coef1 = np.vstack(coef1)
intercept1 = np.array(intercept1)
coef2 = np.vstack(coef2)
intercept2 = np.array(intercept2)
for i, cl in enumerate(classes):
assert_array_almost_equal(clf1.coef_[i].ravel(),
coef1[i].ravel(),
decimal=2)
assert_almost_equal(clf1.intercept_[i], intercept1[i], decimal=1)
assert_array_almost_equal(clf2.coef_[i].ravel(),
coef2[i].ravel(),
decimal=2)
assert_almost_equal(clf2.intercept_[i], intercept2[i], decimal=1)
def create_criterion(device, weight_with=None):
"""
Creates a `torch.nn.BCEWithLogitsLoss`.
If weight_with is not None, uses class weight for the positive class
Arguments:
----------
device: "cuda" or "cpu"
weight_with: data.Dataset
"""
if weight_with:
y = [row.subtask_a for row in weight_with]
class_weights = compute_class_weight('balanced', ['NOT', 'OFF'], y)
# normalize it
class_weights = class_weights / class_weights[0]
criterion = nn.BCEWithLogitsLoss(pos_weight=torch.Tensor([class_weights[1]]))
else:
criterion = nn.BCEWithLogitsLoss()
criterion = criterion.to(device)
return criterion
def get_synthetic_data():
x, y = make_classification(n_samples=10000,
n_features=200,
n_informative=200,
n_redundant=0,
n_repeated=0,
n_classes=3,
n_clusters_per_class=2,
weights=None,
flip_y=0.02,
class_sep=0.4,
hypercube=True,
shift=0.0,
scale=1.0,
shuffle=True,
random_state=2)
log("class weights", compute_class_weight('balanced', np.unique(y), y))
encoder = LabelEncoder()
encoder.fit(y)
encoded_y = encoder.transform(y)
y = np_utils.to_categorical(encoded_y)
mm_scaler = MinMaxScaler()
x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.3)
x_train = mm_scaler.fit_transform(x_train)
x_test = mm_scaler.transform(x_test)
return x_train, y_train, x_test, y_test
def fit(self, dataset):
"""Fits the intent classifier with a valid Snips dataset
Returns:
:class:`LogRegIntentClassifier`: The same instance, trained
"""
from sklearn.linear_model import SGDClassifier
from sklearn.utils import compute_class_weight
logger.info("Fitting LogRegIntentClassifier...")
dataset = validate_and_format_dataset(dataset)
self.load_resources_if_needed(dataset[LANGUAGE])
self.fit_builtin_entity_parser_if_needed(dataset)
self.fit_custom_entity_parser_if_needed(dataset)
language = dataset[LANGUAGE]
data_augmentation_config = self.config.data_augmentation_config
utterances, classes, intent_list = build_training_data(
dataset, language, data_augmentation_config, self.resources,
self.random_state)
self.intent_list = intent_list
if len(self.intent_list) <= 1:
return self
self.featurizer = Featurizer(
config=self.config.featurizer_config,
builtin_entity_parser=self.builtin_entity_parser,
custom_entity_parser=self.custom_entity_parser,
resources=self.resources,
random_state=self.random_state,
)
self.featurizer.language = language
none_class = max(classes)
try:
x = self.featurizer.fit_transform(dataset, utterances, classes,
none_class)
except _EmptyDatasetUtterancesError:
logger.warning("No (non-empty) utterances found in dataset")
self.featurizer = None
return self
alpha = get_regularization_factor(dataset)
class_weights_arr = compute_class_weight("balanced",
range(none_class + 1),
classes)
# Re-weight the noise class
class_weights_arr[-1] *= self.config.noise_reweight_factor
class_weight = {idx: w for idx, w in enumerate(class_weights_arr)}
self.classifier = SGDClassifier(random_state=self.random_state,
alpha=alpha,
class_weight=class_weight,
**LOG_REG_ARGS)
self.classifier.fit(x, classes)
logger.debug("%s", DifferedLoggingMessage(self.log_best_features))
return self
def test_binary_classifier_class_weight():
"""tests binary classifier with classweights for each class"""
alpha = .1
n_samples = 50
n_iter = 20
tol = .00001
fit_intercept = True
X, y = make_blobs(n_samples=n_samples,
centers=2,
random_state=10,
cluster_std=0.1)
step_size = get_step_size(X, alpha, fit_intercept, classification=True)
classes = np.unique(y)
y_tmp = np.ones(n_samples)
y_tmp[y != classes[1]] = -1
y = y_tmp
class_weight = {1: .45, -1: .55}
clf1 = LogisticRegression(solver='sag',
C=1. / alpha / n_samples,
max_iter=n_iter,
tol=tol,
random_state=77,
fit_intercept=fit_intercept,
multi_class='ovr',
class_weight=class_weight)
clf2 = clone(clf1)
clf1.fit(X, y)
clf2.fit(sp.csr_matrix(X), y)
le = LabelEncoder()
class_weight_ = compute_class_weight(class_weight, np.unique(y), y)
sample_weight = class_weight_[le.fit_transform(y)]
spweights, spintercept = sag_sparse(X,
y,
step_size,
alpha,
n_iter=n_iter,
dloss=log_dloss,
sample_weight=sample_weight,
fit_intercept=fit_intercept)
spweights2, spintercept2 = sag_sparse(X,
y,
step_size,
alpha,
n_iter=n_iter,
dloss=log_dloss,
sparse=True,
sample_weight=sample_weight,
fit_intercept=fit_intercept)
assert_array_almost_equal(clf1.coef_.ravel(), spweights.ravel(), decimal=2)
assert_almost_equal(clf1.intercept_, spintercept, decimal=1)
assert_array_almost_equal(clf2.coef_.ravel(),
spweights2.ravel(),
decimal=2)
assert_almost_equal(clf2.intercept_, spintercept2, decimal=1)
def pre_init(self, data: LoadedKata):
# initialize the array with the labels to avoid zero division
observed_labels = copy.deepcopy(self.labels)
for x in data.data_loader:
observed_labels += [self.labels[label_id] for label_id in x[self.label_name]]
weights = compute_class_weight("balanced", classes=np.asarray(self.labels), y=observed_labels)
self.register_buffer("balance", torch.tensor(weights.astype(np.float32)))
def get_sample_weights(self):
labels = self.get_dataframe()[self.get_label_column()]
categories = labels.cat.categories
class_weights = compute_class_weight('balanced',
classes=categories,
y=labels)
sample_weights = class_weights[labels.cat.codes]
return sample_weights
def __init__(self, tree_model,
X_train,
y_train,
feature_names : List[str],
class_names : (List[str],Mapping[int,str])=None):
self.tree_model = tree_model
self.feature_names = feature_names
self.class_names = class_names
self.class_weight = tree_model.class_weight
if getattr(tree_model, 'tree_') is None: # make sure model is fit
tree_model.fit(X_train, y_train)
if tree_model.tree_.n_classes > 1:
if isinstance(self.class_names, dict):
self.class_names = self.class_names
elif isinstance(self.class_names, Sequence):
self.class_names = {i:n for i, n in enumerate(self.class_names)}
else:
raise Exception(f"class_names must be dict or sequence, not {self.class_names.__class__.__name__}")
if isinstance(X_train, pd.DataFrame):
X_train = X_train.values
self.X_train = X_train
if isinstance(y_train, pd.Series):
y_train = y_train.values
self.y_train = y_train
self.node_to_samples = ShadowDecTree.node_samples(tree_model, X_train)
if self.isclassifier():
self.unique_target_values = np.unique(y_train)
self.class_weights = compute_class_weight(tree_model.class_weight, self.unique_target_values, self.y_train)
tree = tree_model.tree_
children_left = tree.children_left
children_right = tree.children_right
# use locals not args to walk() for recursion speed in python
leaves = []
internal = [] # non-leaf nodes
def walk(node_id):
if (children_left[node_id] == -1 and children_right[node_id] == -1): # leaf
t = ShadowDecTreeNode(self, node_id)
leaves.append(t)
return t
else: # decision node
left = walk(children_left[node_id])
right = walk(children_right[node_id])
t = ShadowDecTreeNode(self, node_id, left, right)
internal.append(t)
return t
root_node_id = 0
# record root to actual shadow nodes
self.root = walk(root_node_id)
self.leaves = leaves
self.internal = internal
def calculate_class_weights(self, labels, occured_labels):
from sklearn.utils import compute_class_weight
label_indexes = [labels.index(i) for i in occured_labels]
class_weight_list = compute_class_weight('balanced', np.unique(label_indexes) ,label_indexes )
class_weight = dict(zip(np.unique(label_indexes), class_weight_list))
return class_weight
def get_class_weight(y_true):
classes = np.arange(y_true.shape[-1])
class_counts = y_true.sum(0).astype(np.int64)
pos = 0
y_weight = np.zeros(int(class_counts.sum()))
for i, count in enumerate(class_counts):
y_weight[pos:pos + count] = i
pos += count
weights = np.sqrt(compute_class_weight("balanced", classes, y_weight))
return {i: weight for i, weight in enumerate(weights)}
def cw_to_dict(y_class):
"""
input: 1D array, labels
output: balanced class weight dictionary
"""
cw = utils.compute_class_weight('balanced', [0, 1],
y_class) #compute class weight
cw_dict = {}
for idx in range(len(cw)):
cw_dict[idx] = cw[idx]
return cw_dict
def getData():
X = []
Y = []
void_label = -1
train_path = '/content/input'
label_path = '/content/highway200'
train_files = sorted(os.listdir(train_path))
label_files = sorted(os.listdir(label_path))
for i in range(len(train_files)):
img = load_img(os.path.join(train_path, train_files[i]))
img = img_to_array(img)
X.append(img)
img = load_img(os.path.join(label_path, label_files[i]), color_mode = 'grayscale')
img = img_to_array(img)
shape = img.shape
img /= 255.0
img = img.reshape(-1)
idx = np.where(np.logical_and(img > 0.25, img < 0.8))[0] # find non-ROI
if len(idx) > 0:
img[idx] = -1
img = img.reshape(shape)
img = np.floor(img)
Y.append(img)
X = np.asarray(X)
Y = np.asarray(Y)
idx = list(range(X.shape[0]))
np.random.shuffle(idx)
np.random.shuffle(idx)
X = X[idx]
Y = Y[idx]
cls_weight_list = []
for i in range(Y.shape[0]):
y = Y[i].reshape(-1)
idx = np.where(y!=void_label)[0]
if(len(idx)>0):
y = y[idx]
lb = np.unique(y) # 0., 1
cls_weight = compute_class_weight('balanced', lb , y)
class_0 = cls_weight[0]
class_1 = cls_weight[1] if len(lb)>1 else 1.0
cls_weight_dict = {0:class_0, 1: class_1}
cls_weight_list.append(cls_weight_dict)
cls_weight_list = np.asarray(cls_weight_list)
return [X, Y, cls_weight_list]
def get_class_weight(y_output):
#news = pd.read_csv(file,encoding='utf-8',header=None)
#tags = news.ix[:,1:].as_matrix()
class_weight = []
tags_num = y_output.shape[1]
for i in range(tags_num):
sample_weight = compute_class_weight('balanced', [0, 1], y_output[:,
i])
class_weight.append(sample_weight)
class_weight = np.vstack(class_weight).T
return class_weight
def _get_class_weights(loader: DataLoader) -> np.ndarray:
"""
compute class weights
:param loader:
:return:
"""
labels = loader.dataset.metadata["dx"].cat.codes # type: ignore
classes = np.unique(labels)
return compute_class_weight(class_weight="balanced",
classes=classes,
y=labels)
def train():
#X_lst, y_lst = load_dataset_folder(DATASET_DIR)
X_lst, y_lst = load(DATASET_DIR)
X, y = stack_data(X_lst, y_lst, onehot=True, num_classes=NUM_CLASSES)
normalized_X = normalize(X, axis=1, norm='l2')
print(X.shape, y.shape)
logging.debug(f'X shape: {X.shape}\ty shape: {y.shape}')
X_train, X_test, y_train, y_test = train_test_split(normalized_X,
y,
test_size=0.33,
random_state=42)
# Convert to tf.data.Dataset
train_data = tf.data.Dataset.from_tensor_slices((X_train, y_train))
test_data = tf.data.Dataset.from_tensor_slices((X_test, y_test))
train_data = train_data.batch(BATCH_SIZE).cache().repeat()
test_data = test_data.batch(BATCH_SIZE)
model = generate_model_01(BATCH_SIZE, num_classes=NUM_CLASSES)
# Compute class weights
class_weights = compute_class_weight('balanced',
np.unique(np.argmax(y_train, axis=1)),
np.argmax(y_train, axis=1))
# TensorBoard callback
log_dir = SAVE_LOGS_DIR + 'fit/' + datetime.datetime.now().strftime(
'%Y%m%d-%H%M%S')
cb_tb = tf.keras.callbacks.TensorBoard(log_dir=log_dir)
cb_es = tf.keras.callbacks.EarlyStopping(monitor='val_loss',
patience=200,
restore_best_weights=True,
verbose=True)
training_history = model.fit(train_data,
epochs=EPOCHS,
validation_data=test_data,
steps_per_epoch=15,
class_weight={
0: class_weights[0],
1: class_weights[1],
2: class_weights[2]
},
callbacks=[cb_tb, cb_es])
# Store training history
np.savetxt(SAVE_LOGS_DIR + 'acc.txt', training_history.history['acc'])
np.savetxt(SAVE_LOGS_DIR + 'val_acc.txt',
training_history.history['val_acc'])
np.savetxt(SAVE_LOGS_DIR + 'loss.txt', training_history.history['loss'])
np.savetxt(SAVE_LOGS_DIR + 'val_loss.txt',
training_history.history['val_loss'])
model.save_weights(SAVE_LOGS_DIR + 'weights.h5')
def load_training_data():
raw_training_data = pd.read_csv('train.csv')
# convert types to ints
raw_training_data['target'] = raw_training_data['target'].apply(class_to_int)
raw_training_data = raw_training_data.astype('float32')
raw_training_data['target'] = raw_training_data['target'].astype('int32')
raw_training_data = raw_training_data.iloc[np.random.permutation(len(raw_training_data))] #shuffle data
# Get the features and the classes
features = np.log(raw_training_data.iloc[:, 1:94] + 1).values # apply log function
classes = raw_training_data['target'].values
print np.unique(classes)
#split train/validate
feat_train, feat_test, class_train, class_test = cross_validation.train_test_split(features, classes,
test_size=0.3,
random_state=1232)
feat_train, feat_val, class_train, class_val = cross_validation.train_test_split(feat_train, class_train,
test_size=0.3,
random_state=1232)
#scale the features
std_scale = preprocessing.StandardScaler().fit(feat_train)
feat_train = std_scale.transform(feat_train)
feat_val = std_scale.transform(feat_val)
feat_test = std_scale.transform(feat_test)
#class weights
weights = compute_class_weight('auto', np.unique(classes), class_train)
weights = weights.astype('float32')
print weights
train_weights = []
val_weights = []
for i in class_train:
train_weights.append(weights[i])
for i in list(class_val):
val_weights.append(weights[i])
#convert to np array for theanets
training_data = [feat_train, class_train, np.array(train_weights)]
validation_data = [feat_val, class_val, np.array(val_weights)]
test_data = [feat_test, class_test]
return training_data, validation_data, test_data, std_scale
def fit(self, X, y):
from sklearn.preprocessing import LabelEncoder
from sklearn.utils import compute_class_weight
label_encoder = LabelEncoder().fit(y)
classes = label_encoder.classes_
class_weight = compute_class_weight(self.class_weight, classes, y)
# Intentionally modify the balanced class_weight
# to simulate a bug and raise an exception
if self.class_weight == "balanced":
class_weight += 1.
# Simply assigning coef_ to the class_weight
self.coef_ = class_weight
return self
def test_binary_classifier_class_weight():
"""tests binary classifier with classweights for each class"""
alpha = .1
n_samples = 50
n_iter = 20
tol = .00001
fit_intercept = True
X, y = make_blobs(n_samples=n_samples, centers=2, random_state=10,
cluster_std=0.1)
step_size = get_step_size(X, alpha, fit_intercept, classification=True)
classes = np.unique(y)
y_tmp = np.ones(n_samples)
y_tmp[y != classes[1]] = -1
y = y_tmp
class_weight = {1: .45, -1: .55}
clf1 = LogisticRegression(solver='sag', C=1. / alpha / n_samples,
max_iter=n_iter, tol=tol, random_state=77,
fit_intercept=fit_intercept,
class_weight=class_weight)
clf2 = clone(clf1)
clf1.fit(X, y)
clf2.fit(sp.csr_matrix(X), y)
le = LabelEncoder()
class_weight_ = compute_class_weight(class_weight, np.unique(y), y)
sample_weight = class_weight_[le.fit_transform(y)]
spweights, spintercept = sag_sparse(X, y, step_size, alpha, n_iter=n_iter,
dloss=log_dloss,
sample_weight=sample_weight,
fit_intercept=fit_intercept)
spweights2, spintercept2 = sag_sparse(X, y, step_size, alpha,
n_iter=n_iter,
dloss=log_dloss, sparse=True,
sample_weight=sample_weight,
fit_intercept=fit_intercept)
assert_array_almost_equal(clf1.coef_.ravel(),
spweights.ravel(),
decimal=2)
assert_almost_equal(clf1.intercept_, spintercept, decimal=1)
assert_array_almost_equal(clf2.coef_.ravel(),
spweights2.ravel(),
decimal=2)
assert_almost_equal(clf2.intercept_, spintercept2, decimal=1)
def test_auto_weight():
"""Test class weights for imbalanced data"""
from sklearn.linear_model import LogisticRegression
# we take as dataset a the two-dimensional projection of iris so
# that it is not separable and remove half of predictors from
# class 1
from sklearn.utils import compute_class_weight
X, y = iris.data[:, :2], iris.target
unbalanced = np.delete(np.arange(y.size), np.where(y > 1)[0][::2])
classes = np.unique(y[unbalanced])
class_weights = compute_class_weight('auto', classes, y[unbalanced])
assert_true(np.argmax(class_weights) == 2)
for clf in (svm.SVC(kernel='linear'), svm.LinearSVC(random_state=0),
LogisticRegression()):
# check that score is better when class='auto' is set.
y_pred = clf.fit(X[unbalanced], y[unbalanced]).predict(X)
clf.set_params(class_weight='auto')
y_pred_balanced = clf.fit(X[unbalanced], y[unbalanced],).predict(X)
assert_true(metrics.f1_score(y, y_pred)
<= metrics.f1_score(y, y_pred_balanced))
def classify(self):
y_data = self.results[self.task.label].data
X_data = self.results[self.task.features].data
y = np.array(y_data)
X = np.array(X_data)
cw = compute_class_weight('auto', np.array([0,1]), y)
cw = {0:cw[0],1:cw[1]}
print cw
b = get_classifier(self.task.classifier, cw)
tile_size = 1000
num_tiles = y.size / tile_size
for i in range(num_tiles):
pos = i * tile_size
X_sub = X[pos : pos + tile_size]
y_sub = y[pos : pos + tile_size]
y_prob = None
y_pred = None
if self.task.classifier == 'svm':
y_pred = b.predict(X_sub)
y_prob = np.array([[0,y] for y in y_pred])
else:
y_prob = b.predict_proba(X_sub)
y_pred = [1 if y[1] >= 0.5 else 0 for y in y_prob]
cm = confusion_matrix(y_sub, y_pred)
stats = classify_stats(cm, y_test, y_prob)
y_pred = pd.DataFrame(y_pred, columns=y_data.columns)
result = ClassifyResult(self.task, 1.0, b, stats)
self.results[self.task.uuid] = result
b.partial_fit(X_sub, y_sub)
def _compute_class_weight_dictionary(y):
# helper for returning a dictionary instead of an array
classes = np.unique(y)
class_weight = compute_class_weight("balanced", classes, y)
class_weight_dict = dict(zip(classes, class_weight))
return class_weight_dict
# path to image folder
base_path = os.path.join(base_path, caltech101.config.tar_inner_dirname)
# X_test contain only paths to images
(X_test, y_test) = util.load_paths_from_files(base_path, 'X_test.txt', 'y_test.txt')
for cv_fold in [0]: # on which cross val folds to run; cant loop over several folds due to some bug
print("fold {}".format(cv_fold))
experiment_name = '_bn_triangular_cv{}_e{}'.format(cv_fold, nb_epoch)
# load cross val split
(X_train, y_train), (X_val, y_val) = util.load_cv_split_paths(base_path, cv_fold)
# compute class weights, since classes are highly imbalanced
class_weight = compute_class_weight('auto', range(nb_classes), y_train)
if normalize_data:
print("Load mean and std...")
X_mean, X_std = util.load_cv_stats(base_path, cv_fold)
normalize_data = (X_mean, X_std)
nb_train_sample = X_train.shape[0]
nb_val_sample = X_val.shape[0]
nb_test_sample = X_test.shape[0]
print('X_train shape:', X_train.shape)
print(nb_train_sample, 'train samples')
if X_val is not None:
print(nb_val_sample, 'validation samples')
print(nb_test_sample, 'test samples')
from sklearn.svm import LinearSVC
from sklearn.metrics import average_precision_score
from sklearn.utils import compute_class_weight
import numpy as np
import logging
logging.basicConfig(level=logging.DEBUG)
iris = load_iris()
X = iris.data
y = iris.target
y[y != 1] = -1
y[y == 1] = 1
weights = compute_class_weight("auto", np.unique(y), y)
sample_weight = np.zeros(y.shape, dtype=np.float)
sample_weight[y==1] = weights[0]
sample_weight[y==-1] = weights[1]
# n_iter = int(1e6 / X.shape[0])
vw_clf = VWClassifier(quiet=False, loss_function="hinge", passes=500)
vw_clf.fit(X, y.astype(np.double), sample_weight)
scores = vw_clf.decision_function(X)
print "VW AP: %.3f" % average_precision_score(y, scores)
vw_clf.set_params(l2=0.1)
vw_clf.fit(X, y.astype(np.double), sample_weight)
print "VW AP: %.3f" % average_precision_score(y, scores)
# vw_clf.fit(X, y.astype(np.double), sample_weight)