def train_test_split(self, feature_matrix, target_matrix, test_size=0.2):
'''
Stratified Shuffle split technique is used to split train and test set,
to have the equal proportion of classes in train and test.
Input:
feature_matrix : Feature matrix with rare classes filtered out
target_matrix : Target matrix with rare classes filtered out
test_size: default is 20%
Output:
train_x, train_y, test_x, test_y
'''
lp = LabelPowerset()
sss_level_1 = StratifiedShuffleSplit(lp.transform(target_matrix),
n_iter=1,
test_size=0.2,
random_state=123)
for train_ix, test_ix in sss_level_1:
train_x = feature_matrix.iloc[train_ix, :]
train_y = target_matrix.iloc[train_ix, :]
test_x = feature_matrix.iloc[test_ix, :]
test_y = target_matrix.iloc[test_ix, :]
return train_x, train_y, test_x, test_y
class MyLabelPowerSetFeatureSelect():
def fit(self, X, y):
# I'm using a gaussian naive bayes base classifier
self.LabelPowerSetObject = LabelPowerset(GaussianNB())
# fitting the data
self.LabelPowerSetObject.fit(X, y)
# transformed y
y_transformed = self.LabelPowerSetObject.transform(y)
# instanciating with SelectKBest object
self.X_new = SelectKBest(chi2, k=2)
# the feature selecting
self.X_transformed = self.X_new.fit_transform(X, y_transformed)
# save indices of the saved attributes
self.selected_attributes_indices = self.X_new.get_support(indices = True)
#print(self.attributes_indices,'the indices of the selected atributes')
return self
def transform(self, X):
return X[:,self.selected_attributes_indices]
def predict(self, X):
return self.LabelPowerSetObject.predict(X)
def predict_proba(self, X):
return self.LabelPowerSetObject.predict_proba(X)
def filter_rare_classes(self, feature_matrix, target_matrix):
'''
In order to perform stratified split between train and test,there
should be atleast 2 instances present in the data. Hence, filter
label combinations that occurs only once in the entire dataset.
Input :
Feature Matrix : matrix of features
Target Matrix : matrix containing the the target labels
Output :
Feature Matrix : Filtered
Target Matrix : Filtered
'''
lp = LabelPowerset()
multi_class_target_labels = lp.transform(target_matrix)
classes_vc = np.asarray(
np.unique(multi_class_target_labels,
return_counts=True)).T # 1635 unique classes
class_to_keep = classes_vc[np.where(classes_vc[:, 1] > 1)][:, 0]
mask = [
True if
(multi_class_target_labels[i] in (class_to_keep)) else False
for i in range(len(multi_class_target_labels))
]
feature_matrix = feature_matrix[mask]
target_matrix = target_matrix[mask]
return feature_matrix, target_matrix
def multi_labelTo_multi_class(Y, model):
num_of_labels = Y.ndim
if (num_of_labels == 1):
print("This is not a multi-label problem!!!!!!")
return Y
#LabelPowerset is used here as it contains the transform function
#that actuall do the multi_label to muti_class transformation.
transclf = LabelPowerset(classifier=model, require_dense=[False, True])
return [transclf, transclf.transform(Y)]
def resampling_data(self, X, y):
# Import a dataset with X and multi-label y
lp = LabelPowerset()
ros = RandomOverSampler(random_state=42)
# Applies the above stated multi-label (ML) to multi-class (MC) transformation.
yt = lp.transform(y)
X_resampled, y_resampled = ros.fit_sample(X, yt)
# Inverts the ML-MC transformation to recreate the ML set
y_resampled = lp.inverse_transform(y_resampled)
return X_resampled, y_resampled
def load_ucmerced_dataset():
dataset = scipy.io.loadmat(
'/home/rishabh/Downloads/multi-label-analysis-master/src/graphcnn/setup/dataset/dataset1.mat'
)
dataset = dataset['dataset1']
edges = np.squeeze(dataset['edges']) #adjacecny matrix
index = np.squeeze(dataset['index']) # image index to keep track
classes = np.squeeze(dataset['class']) #image class number to keep track
#loading features in which NaN values have been replaced
features = scipy.io.loadmat(
'/home/rishabh/Downloads/multi-label-analysis-master/src/graphcnn/setup/dataset/features.mat'
)
features = features['features']
features = features['val']
features = features[0]
for i in range(0, len(features)):
if np.isnan(features[i]).any() == True:
print('features %d have NaN:' % i, np.isnan(features[i]).any())
# loading multi-labels
labels = scipy.io.loadmat(
'/home/rishabh/Downloads/multi-label-analysis-master/src/graphcnn/setup/dataset/LandUse_multilabels.mat'
)
labels = labels['labels']
labels = np.transpose(labels, (1, 0))
# Calculating class weights
lp = LabelPowerset()
trans_labels = lp.transform(labels)
unique, counts = np.unique(trans_labels, return_counts=True)
class_freq = 1.0 / counts
weight_mat = np.zeros((np.shape(trans_labels)))
for i in range(len(weight_mat)):
weight_mat[i] = class_freq[np.where(trans_labels[i] == unique)]
# Calculating label weights
sum_labels = np.sum(labels, axis=0, dtype=np.float32)
sum_tot = np.sum(sum_labels, dtype=np.float32)
label_freq = np.true_divide(sum_labels, sum_tot)
return features, edges, labels, weight_mat, label_freq, index, classes
def multiple_smote(X, y):
"""
为multi-label样本过采样
"""
# Import a dataset with X and multi-label y
y = np.array(y)
lp = LabelPowerset()
# oversampler = ADASYN(random_state=1994, n_neighbors=2)
oversampler = SMOTE(k_neighbors=2)
# Applies the above stated multi-label (ML) to multi-class (MC) transformation.
yt = lp.transform(y)
X_resampled, y_resampled = oversampler.fit_resample(X, yt)
# Inverts the ML-MC transformation to recreate the ML set
y_resampled = lp.inverse_transform(y_resampled) # return a sparse matrix
return X_resampled, y_resampled.toarray()
def label_fscore(self, subset=False):
if subset:
true_list, pred_list = self.true_list, self.pred_list
if self._is_binarized:
# transform multilabel to multiclass for subset measurement
lp = LabelPowerset()
transformed = lp.transform(np.concatenate((true_list,
pred_list)))
true_list, pred_list = np.split(transformed, 2)
else:
true_list, pred_list = self._binarized_labels()
prec, rec, fscore, count = skm.precision_recall_fscore_support(
true_list, pred_list)
fscores_dict = {}
for c, f in zip(count, fscore):
# label = class_by_count[c]
# for when remove_multi_labeled is used
if c == 0:
continue
fscores_dict[c] = f
return fscores_dict
def resample_multilabel(data, target):
"""
Apply LP-transformation to create balanced classes, then convert back to multilabel targets
"""
target = target.astype(int)
def convert_hads_to_str(hads_data, hads_type):
hads_strs = []
for val in hads_data:
if val == 0:
str_convert = '%s_normal' % hads_type
elif val == 1:
str_convert = '%s_borderline' % hads_type
elif val == 2:
str_convert = '%s_abnormal' % hads_type
hads_strs.append(str_convert)
return hads_strs
def convert_str_to_hads(hads_tuples):
hads_array = np.ndarray(shape=(len(hads_tuples), 2))
for t, tup in enumerate(hads_tuples):
for s, str in enumerate(tup):
if '_normal' in str:
hads_array[t, s] = 0
elif '_borderline' in str:
hads_array[t, s] = 1
elif '_abnormal' in str:
hads_array[t, s] = 2
return hads_array
anx_strings = convert_hads_to_str(target[:, 0], 'anxiety')
dep_strings = convert_hads_to_str(target[:, 1], 'depression')
multilabel_hads = [(anx_strings[n], dep_strings[n])
for n in range(len(anx_strings))]
mlb = preprocessing.MultiLabelBinarizer()
binary_matrix = mlb.fit_transform(multilabel_hads)
from skmultilearn.problem_transform import LabelPowerset
lp = LabelPowerset()
target_lp_transformed = lp.transform(binary_matrix)
resampler = RandomOverSampler(sampling_strategy='not majority',
random_state=seed)
data_resampled, target_lp_transformed_resampled = resampler.fit_sample(
data, target_lp_transformed)
binary_matrix_resampled = lp.inverse_transform(
target_lp_transformed_resampled)
target_resampled_multilabel = mlb.inverse_transform(
binary_matrix_resampled)
target_resampled_multilabel_array = convert_str_to_hads(
target_resampled_multilabel)
anx_resampled_to_str = convert_hads_to_str(
target_resampled_multilabel_array[:, 0], 'anxiety')
dep_resampled_to_str = convert_hads_to_str(
target_resampled_multilabel_array[:, 1], 'depression')
target_resampled_multilabel_df = pd.DataFrame()
target_resampled_multilabel_df['anxiety'] = anx_resampled_to_str
target_resampled_multilabel_df['depression'] = dep_resampled_to_str
return data_resampled, target_resampled_multilabel_df.values, target_lp_transformed_resampled
def lp_framework(self):
clusters = []
for i in range(len(np.unique(self.clustering_labels))):
cluster = []
for j in range(len(self.clustering_labels)):
if i == self.clustering_labels[j]:
cluster.append(self.labels[j])
clusters.append(cluster)
print(self.labels)
print(clusters)
transformed_labels = []
LabelPowersetTramsformer = LabelPowerset()
for cluster in clusters:
print(cluster)
print(self.universe[:, cluster].shape)
# print(self.universe[:, cluster])
print(len(cluster))
if len(cluster) == 1:
transformed_labels.append(
list(self.universe[:, cluster].reshape(-1)))
else:
# print(self.universe[:, cluster])
print(
LabelPowersetTramsformer.transform(self.universe[:,
cluster]))
print(
type(
LabelPowersetTramsformer.transform(
self.universe[:, cluster]).shape))
print(
list(
LabelPowersetTramsformer.transform(
self.universe[:, cluster])))
tmp = list(
LabelPowersetTramsformer.transform(self.universe[:,
cluster]))
print(
type(
list(
LabelPowersetTramsformer.transform(
self.universe[:, cluster]))))
transformed_labels.append(tmp)
# print(len(transformed_labels))
# print(len(transformed_labels[0]))
# print(transformed_labels[0])
# print(transformed_labels[0][0])
# print(transformed_labels[0][0][0])
print(np.array(transformed_labels).shape)
print("universe shape:" + str(self.universe.shape))
print(self.universe[:, self.attributes].shape)
print(np.array(transformed_labels).T.shape)
self.universe = np.append(self.universe[:, self.attributes],
np.array(transformed_labels).T,
axis=1)
print(self.universe.shape)
print(self.attributes)
self.labels = [
i for i in range(len(self.attributes), self.universe.shape[1])
]
print(self.labels)
pass
train_text_ori = train_df['alltext'].tolist()
train_text_ori = [' '.join(t.split()) for t in train_text_ori]
train_text_ori = np.array(train_text_ori, dtype=object)[:, np.newaxis]
train_label_ori = train_df.values[:,3:-1]
test_text = test_df['alltext'].tolist()
test_text = [' '.join(t.split()) for t in test_text]
test_text = np.array(test_text, dtype=object)[:, np.newaxis]
test_label = test_df.values[:,3:-1]
## Upsampling of data for each label
print('Before upsampling: ',train_text_ori.shape,train_label_ori.shape,test_text.shape,test_label.shape)
yt = lp.transform(train_label_ori.astype('int'))
train_text, y_resampled = ros.fit_sample(train_text_ori.astype('str'), yt)
train_label = lp.inverse_transform(y_resampled).toarray()
# train_text=train_text_ori
# train_label = train_label_ori
print('After Up-sampling',train_text.shape,train_label.shape,test_text.shape,test_label.shape)
# Instantiate tokenizer
tokenizer = create_tokenizer_from_hub_module()
# Convert data to InputExample format
train_examples = convert_text_to_examples(train_text, train_label)
def load_ucmerced_dataset():
dataset = scipy.io.loadmat(
'dataset/newtotpatt.mat') #dataset1.mat, totgra.mat
dataset = dataset['newtotpatt'] #dataset1, totgra
#dataset = np.transpose(dataset);
edges = np.squeeze(dataset['edges']) #adjacecny matrix
index = np.squeeze(dataset['index']) # image index to keep track
classes = np.squeeze(dataset['class']) #image class number to keep track
#'''
#loading features in which NaN values have been replaced
features1 = scipy.io.loadmat('dataset/features1n.mat') #normfeatures1.mat
features1 = features1['features1n']
features1 = np.squeeze(features1['val'])
features2 = scipy.io.loadmat('dataset/features2n.mat') #normfeatures1.mat
features2 = features2['features2n']
features2 = np.squeeze(features2['val'])
features3 = scipy.io.loadmat('dataset/features3n.mat') #normfeatures1.mat
features3 = features3['features3n']
features3 = np.squeeze(features3['val'])
features = np.concatenate((features1, features2))
features = np.concatenate((features, features3))
#features = np.squeeze(dataset(features))
#features = features['val']
#features = features[0]
print type(features[0])
print features.shape
for i in range(0, len(features)):
if np.isnan(features[i]).any() == True:
print('features %d have NaN:' % i, np.isnan(features[i]).any())
'''
f = h5py.File('dataset/pattfeatures.mat','r')#dataset1.mat
test = f['features'] #dataset1
#dataset = np.transpose(dataset);
test = np.squeeze(test['val']) #adjacecny matrix
st=test
for i in range(30400):
st[i] = test[i]
features = f[st]
#index = np.squeeze(dataset['index'][:]) # image index to keep track
#val = np.array(val)
#index = np.array(index)
#classes = np.squeeze(dataset['class']) #image class number to keep track
#features = features['val']
#features = test
#features = f[features]
#str1 = ''.join(features(i) for i in features[:])
#features = np.array([features])
#features[0] = np.array(features[0])
print type(features[0])
print features[0].shape
#for i in range(0,len(features)):
# if np.isnan(features[i]).any() == True:
# print('features %d have NaN:'% i,np.isnan(features[i]).any())
'''
#loading multi-labels
labels = scipy.io.loadmat('dataset/pattlabels.mat') #LandUse_multilabels
labels = labels['labels']
#labels = np.transpose(labels)
#loading positive-labels
#pos = scipy.io.loadmat('dataset/Indexes.mat') ##labels
#pos = pos['ind']
#pos = np.transpose(pos)
#neg = scipy.io.loadmat('dataset/Index-.mat') ##labels
#neg = labels['neg_ind']
#neg = np.transpose(neg)
#load pairs
#pairs = scipy.io.loadmat('dataset/pair.mat')
# Calculating class weights
lp = LabelPowerset()
trans_labels = lp.transform(labels)
unique, counts = np.unique(trans_labels, return_counts=True)
class_freq = 1.0 / counts
weight_mat = np.zeros((np.shape(trans_labels)))
for i in range(len(weight_mat)):
weight_mat[i] = class_freq[np.where(trans_labels[i] == unique)]
# Calculating label weights
sum_labels = np.sum(labels, axis=0, dtype=np.float32)
sum_tot = np.sum(sum_labels, dtype=np.float32)
label_freq = np.true_divide(sum_labels, sum_tot)
return features, edges, labels, weight_mat, label_freq, index, classes
