def learnPhase():
if os.path.isfile("Doc2VecSVMNauceni.pkl"):
return None
tablecolrow = loadData("train.csv")
tablecolrow[3] = FilterQuestions(tablecolrow[3])
tablecolrow[4] = FilterQuestions(tablecolrow[4])
model = prepareDoc2Vec(tablecolrow[3], tablecolrow[4])
for i in range(len(tablecolrow[3])):
tablecolrow[3][i] = model.infer_vector(tablecolrow[3][i].split(" "))
tablecolrow[4][i] = model.infer_vector(tablecolrow[4][i].split(" "))
traindataX = [None] * len(tablecolrow[3])
traindataY = [None] * len(tablecolrow[3])
for i in range(len(traindataX)):
traindataX[i] = tablecolrow[3][i] + tablecolrow[4][i]
traindataY[i] = int(tablecolrow[5][i])
svmKlasifikator = SVC(kernel='rbf',
verbose=True,
probability=True,
max_iter=1000000)
print("Learning started")
tmStart = timer()
svmKlasifikator.fit(traindataX, traindataY)
tmEnd = timer()
print("Predicting lasted", tmEnd - tmStart)
joblib.dump(svmKlasifikator, 'Doc2VecSVMNauceni.pkl')
print("Spremljen je napredak ucenja")
class SVMLearner(RLRLearner):
def __init__(self, data_model):
super().__init__(data_model)
self.svm = SVC(kernel='linear', probability=True, tol=0.0001)
def fit_transform(self, pairs, y):
y = numpy.array(y)
if not y.any() and self.y.any():
random_pair = random.choice(self.candidates)
exact_match = (random_pair[0], random_pair[0])
pairs = pairs + [exact_match]
y = numpy.concatenate([y, [1]])
elif numpy.count_nonzero(y) == len(y) and numpy.count_nonzero(
self.y) == len(self.y):
random_pair = random.choice(self.candidates)
pairs = pairs + [random_pair]
y = numpy.concatenate([y, [0]])
super().fit_transform(pairs, y)
def fit(self, X, y):
self.y = y
self.X = X
self.svm.fit(X, y)
def predict_proba(self, examples):
return self.svm.predict_proba(examples)[:, 1].reshape(-1, 1)
def learnModel(data):
if os.path.isfile("BagOfWordsSVMNauceni.pkl"):
return None
data[0] = FilterQuestions(data[0])
data[1] = FilterQuestions(data[1])
# Initialize the "CountVectorizer" object, which is scikit-learn's
# bag of words tool.
vectorizer = CountVectorizer(analyzer = "word", \
tokenizer = None, \
preprocessor = None, \
stop_words = None, \
max_features = 20000)
allQuestions = data[0] + data[1]
vectorizer.fit(allQuestions)
joblib.dump(vectorizer, 'BagOfWordsVectorizerNauceni.pkl')
znacajkePitanja = [vectorizer.transform(data[0]), vectorizer.transform(data[1])]
for i, r in enumerate(data[2]):
data[2][i] = int(r)
znacajkePitanja = hstack(znacajkePitanja).tocsr()
svmKlasifikator = SVC(kernel='rbf', verbose=True, probability=True, max_iter=1000000)
print("Learning started")
tmStart = timer()
svmKlasifikator.fit(znacajkePitanja, data[2])
tmEnd = timer()
print("Learning ended")
print("Learning lasted", tmEnd - tmStart)
joblib.dump(svmKlasifikator, 'BagOfWordsSVMNauceni.pkl')
print("Spremljen je napredak ucenja")
class SVCImpl():
def __init__(self, C=1.0, kernel='rbf', degree=3, gamma='auto_deprecated', coef0=0.0, shrinking=True, probability=False, tol=0.001, cache_size=200, class_weight='balanced', verbose=False, max_iter=(- 1), decision_function_shape='ovr', random_state=None):
self._hyperparams = {
'C': C,
'kernel': kernel,
'degree': degree,
'gamma': gamma,
'coef0': coef0,
'shrinking': shrinking,
'probability': probability,
'tol': tol,
'cache_size': cache_size,
'class_weight': class_weight,
'verbose': verbose,
'max_iter': max_iter,
'decision_function_shape': decision_function_shape,
'random_state': random_state}
def fit(self, X, y=None):
self._sklearn_model = SKLModel(**self._hyperparams)
if (y is not None):
self._sklearn_model.fit(X, y)
else:
self._sklearn_model.fit(X)
return self
def predict(self, X):
return self._sklearn_model.predict(X)
def predict_proba(self, X):
return self._sklearn_model.predict_proba(X)
class SVMClassifier(ClassifierI):
"""Wrapper for scikit-learn svm classifier."""
def __init__(self, C=1.0, kernel='rbf', degree=3, gamma='auto', coef0=0.0,
shrinking=True, probability=False, tol=1e-3, cache_size=200,
class_weight=None, verbose=False, max_iter=-1,
decision_function_shape=None, random_state=None):
"""Init. See scikit-learn."""
self._clf = SVC(C=1, kernel=kernel, degree=degree, gamma=gamma,
coef0=coef0, shrinking=shrinking,
probability=probability, tol=tol, cache_size=cache_size,
class_weight=class_weight, verbose=verbose,
max_iter=max_iter,
decision_function_shape=decision_function_shape,
random_state=random_state)
self.classes_ = None
def __repr__(self):
return "<SVMClassifier(%r)>" % self._clf
def classify_many(self, vectors):
"""Classify a batch of verbs.
:param vectors: An doc term array of vectors
:return: The predicted class label for each input sample.
:rtype: list
"""
classes = self.classes_
return [classes[i] for i in self._clf.predict(vectors)]
def prob_classify_many(self, vectors):
"""Compute per-class probabilities for a batch of samples.
:param vectors: A doc term array of vectors
:rtype: list of ``ProbDistI``
"""
y_proba_list = self._clf.predict_proba(vectors)
return [self._make_probdist(y_proba) for y_proba in y_proba_list]
def labels(self):
"""The class labels learned by this classifier.
:rtype: list
"""
return list(self.classes_)
def train(self, vectors, labels):
"""
Train (fit) the scikit-learn svm classifier.
:param vectors: a doc-term array of vectors to learn from
:param labels: a list of labels corresponding to the rows
of the doc term array.
"""
self.classes_, labels = np.unique(labels, return_inverse=True)
self._clf.fit(vectors, labels)
return self
def _make_probdist(self, y_proba):
classes = self.classes_
return dict((classes[i], p) for i, p in enumerate(y_proba))
class CreateSVC(CreateLinearSVC):
def fit(self, data, args):
self.model = SVC(probability=True)
with Timer() as t:
self.model.fit(data.X_train, data.y_train)
return t.interval
def svm_train(X, y, model_path):
model = SVC()
model.fit(X, y)
expected = y
predicted = model.predict(X)
print(metrics.classification_report(expected, predicted))
print(metrics.confusion_matrix(expected, predicted))
joblib.dump(model, model_path)
def train_svm(params, suffix, train_X, train_Y, test_X, test_Y):
C = params['C']
kernel = params['kernel']
model = SVC(gamma='scale', probability=True, C=C, kernel=kernel)
print("Params C:", C, "kernel:", kernel)
model.fit(train_X, train_Y)
print("Train score", model.score(train_X, train_Y))
test_score = model.score(test_X, test_Y)
print("Test score", test_score)
return test_score, None
def phoneAccelerometerISVM():
print("Loading data...")
data = pd.read_csv("./Train_Phone-Acc-nexus4_1-a.csv")
print("Done!")
# Parse data and make bike vs not-biking classification using an SVM.
# Note: I'm assuming a window width of 500
print("Finding time series windows indexes for each class kind...")
previousClassLabel = str(data.get_value(data.index[0], 'gt'))
pos = 0
y = []
X = []
window = 500
while pos < data.shape[0]:
# Make y label.
if str(data.iloc[pos]['gt']) == 'sit':
y.append(1)
else:
y.append(-1)
# Make X row.
X.append(data.iloc[pos:pos + window]['y'])
# Move to the next window
pos += window
print("Done!")
# Build and fit the SVM.
print("Training SVM on all data accelerometer data...")
X = np.array(X)
y = np.array(y)
#clfs = LinearSVC()
clfs = SVC()
clfs.fit(X, y)
print("Done!")
# print("Predicting accelerometer classes on all data using SVM...")
# ypred = predict(X, clfs.coef_.reshape(len(clfs.coef_.ravel()), 1))
# print("Done!")
# error = calculateTotalAbsoluteError(y, ypred) / y.shape[0]
# print("Accelerometer training error (Means kind of nothing): %f"%error)
# Cross validation
print("Training SVM on accelerometer training only data...")
X_train, X_test, y_train, y_test = train_test_split(X, y, stratify=y, test_size = 0.1) #, random_state = 0
clfs = SVC()
clfs.fit(X_train, y_train)
yhat = clfs.predict(X_test)
print("Abs Error = %f"%( calculateTotalAbsoluteError(yhat, y_test)/len(yhat)))
print("Test data mean accuracy SVM score: %f"%clfs.score(X_test, y_test))
f1_c0 = f1_score(y_test, clfs.predict(X_test), pos_label=1, average='binary')
#print("Test data f1 score for class -1: %f"%(f1_c0))
print("Test data f1 score for class +1: %f" % (f1_c0))
print("Done!")
def SVCClassify(self, x_train, y_train):
'''
Basic Support Vector Machine Classifier
'''
# the parameter can be set
kernel = 'rbf'
# init classifier and train it
# if need the proba-predict result, parameter probability must be =True
clf = SVC(kernel=kernel, probability=True)
clf.fit(x_train, y_train)
return clf
def SVCClassify(self, x_train, y_train):
'''
Basic Support Vector Machine Classifier
'''
# the parameter can be set
kernel = 'rbf'
# init classifier and train it
# if need the proba-predict result, parameter probability must be =True
clf = SVC(kernel=kernel, probability=True)
clf.fit(x_train, y_train)
return clf
def cross_validate(samples, labels, outputDir):
'''
Function to perform K-fold cross validation
'''
# K(=10) FOLD CROSS VALIDATION
K = 10
fold_samples, fold_labels = cv_split(samples, np.array(labels), K)
log_loss = [['Log Loss'],[]]
total_ll = 0.0
for fold in range(K):
samples_chunk = fold_samples[:fold] + fold_samples[fold+1:]
labels_chunk = fold_labels[:fold] + fold_labels[fold+1:]
#Training L1 logistic regression
logRegrL1 = linear_model.LogisticRegression(C=1, penalty='l1')
logRegrL1.fit( np.concatenate(samples_chunk, axis=0), np.concatenate(labels_chunk, axis = 0) )
#Training SVM with linear kernel
svmLin = SVC(kernel='linear', probability=True)
svmLin.fit( np.concatenate(samples_chunk, axis=0), np.concatenate(labels_chunk, axis = 0) )
#Training Random Forest Classifier
rfc = RandomForestClassifier(n_estimators=100)
rfc.fit( np.concatenate(samples_chunk, axis=0), np.concatenate(labels_chunk, axis = 0) )
#TEST ON CROSS VALIDATION HOLD OUT SET
val = [i for i in range(len(fold_labels[fold]))]
id = 0
for item in fold_samples[fold]:
predictionL1 = logRegrL1.predict_proba(item)#first component is probability of 0 class, second is of class 1
predictionSvmLin = svmLin.predict_proba(item)
predictionRfc = rfc.predict_proba(item)
#Taking the average of each of the model predictions as final health status prediction
val[id] = (predictionL1[0][1] + predictionSvmLin[0][1] + predictionRfc[0][1])/3.0
id = id + 1
for i in range(len(fold_labels[fold])):
total_ll += logloss(fold_labels[fold][i], val[i])
log_loss[1] = total_ll/len(samples)
#Save csv file in the output directory with name Dota2Val.csv
np.savetxt(outputDir + "\\Dota2Val.csv",
log_loss,
delimiter=',',
fmt='%s'
)
def train_and_predict(samples, labels, feature_selector, inputDir, outputDir):
#Training L1 logistic regression
logRegrL1 = linear_model.LogisticRegression(C=1, penalty='l1')
logRegrL1.fit(samples, labels)
#Training SVM with linear kernel
svmLin = SVC(kernel='linear', probability=True)
svmLin.fit(samples, labels)
#Training Random Forest Classifier
rfc = RandomForestClassifier(n_estimators=100)
rfc.fit(samples, labels)
#test set
testDir = inputDir + "/set_test"
testFiles = sorted([
join(testDir, f) for f in listdir(testDir) if isfile(join(testDir, f))
],
key=numericalSort)
#Read feature vectors of test images
testSamples = cubeVoxelsVar(testFiles)
testSamples = feature_selector.transform(testSamples)
print(len(testSamples))
#2D array to report final prediction in format (ID,Prediction)
final = [[0 for j in range(2)] for i in range(139)]
final[0][0] = 'ID'
final[0][1] = 'Prediction'
id = 1
#Predict health status of test image using each of the 3 models trained above
for item in testSamples:
predictionL1 = logRegrL1.predict_proba(
item
) #first component is probability of 0 class, second is of class 1
predictionSvmLin = svmLin.predict_proba(item)
predictionRfc = rfc.predict_proba(item)
final[id][0] = id
#Taking the average of each of the model predictions as final health status prediction
final[id][1] = (predictionL1[0][1] + predictionSvmLin[0][1] +
predictionRfc[0][1]) / 3.0
id = id + 1
#Save csv file in the output directory with name final_sub.csv
np.savetxt(outputDir + "/final_sub.csv", final, delimiter=',', fmt='%s')
def learnModel(train):
data = []
for duplicate in train["is_duplicate"]:
data.append(int(duplicate))
znacajkePitanja = get_avg(train)
svmKlasifikator = SVC(kernel='rbf',
verbose=True,
probability=True,
max_iter=10000)
print("Learning started")
tmStart = timer()
svmKlasifikator.fit(znacajkePitanja, data)
tmEnd = timer()
print("Learning ended")
print("Learning lasted", tmEnd - tmStart)
joblib.dump(svmKlasifikator, 'Word2VecSVMNauceni.pkl')
print("Spremljen je napredak ucenja")
def train_all(self, g):
X = np.concatenate([self.train_X, self.val_X], axis=0)
if self.use_scale:
self.scale.fit(X)
X = self.scale.transform(X)
for i in range(3):
y = np.concatenate([self.train_y, self.val_y], axis=0)
y[y!=i+1]=0
y[y!=0]=1
clf = SVC()
clf.set_params(**g)
self.model_a.append(clf.fit(X, y))
def classifier_panchenko2016(X_train,
y_train,
X_test,
y_test,
separateClassifier=False):
train_or_test_labels = ["train"
for i in y_train] + ["test" for i in y_test]
y_train, X_train, y_test, X_test = outlier_removal(train_or_test_labels,
X_train + X_test,
y_train + y_test)
y_train, X_train = features_extraction(
y_train,
X_train,
separateClassifier=separateClassifier,
featuresCount=100)
y_test, X_test = features_extraction(y_test,
X_test,
separateClassifier=separateClassifier,
featuresCount=100)
scaler = MinMaxScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)
classifier = SVC(kernel="rbf",
C=2e11,
gamma=2e-1,
max_iter=5000,
class_weight="balanced",
verbose=1)
print("fitting")
classifier.fit(X_train, y_train)
print("testing")
y_predictions = classifier.predict(X_test) #, y_test)
return y_test, y_predictions
def train(self, g):
self.model = []
X = self.train_X.copy()
if self.use_scale:
self.scale.fit(X)
X = self.scale.transform(X)
for i in range(3):
y = self.train_y.copy()
y[y!=i+1]=0
y[y!=0]=1
clf = SVC()
clf.set_params(**g)
self.model.append(clf.fit(X, y))
#Read age labels of training images
labels = []
for t in targets:
labels.append(t[0])
#Training LASSO regressor, alpha value tuned to produce best result when used alone on the test set
regrL = linear_model.Lasso(alpha=15.0)
regrL.fit(samples, labels)
#Training Ridge regressor, alpha value tuned to produce best result when used alone on the test set
regrR = linear_model.Ridge(alpha=1e-13, normalize=True)
regrR.fit(samples, labels)
#Training SVM with linear kernel
regrS = SVC(kernel='linear')
regrS.fit(samples, labels)
#Training Random Forest Classifier
rfc = RandomForestClassifier(n_estimators=100)
rfc.fit(samples, labels)
#test set
testDir = inputDir + "\\set_test"
testFiles = sorted([
join(testDir, f) for f in listdir(testDir) if isfile(join(testDir, f))
],
key=numericalSort)
#Read feature vectors of test images
testSamples = readVoxels(testFiles)
print(len(testSamples))
X_train, Y_train = train_data_set.convert_2_binary_format()
test_data_set = DataSet()
test_data_set.load(config.get_value('test'), class_index, has_header=False)
Xtest, Ytest = test_data_set.convert_2_binary_format_with(
X_train.item_dict, Y_train.item_dict)
Ytest = Ytest.flatten()
class_count = train_data_set.number_of_classes()
unexpected_rules = IOHelper.load_json_object(config.get_value('rules'))
refined_unexpected_rules = filter_association_rules(unexpected_rules)
print('svm testing...')
svc_model = SVC(kernel='poly', degree=3, coef0=0.1, random_state=1)
svc_model.fit(X_train.relation_matrix, Y_train.values.flatten())
svc_y_pred = svc_model.predict(Xtest)
print(f1_score(Ytest, svc_y_pred, average=None))
if (class_count <= 2):
fpr, tpr, _ = roc_curve(Ytest, svc_y_pred.flatten())
print(auc(fpr, tpr))
refine_with_unexpectedness(test_data_set, Y_train.item_dict, svc_y_pred,
Ytest, refined_unexpected_rules)
print('Random forest testing...')
rf_model = RandomForestClassifier(n_estimators=20, random_state=1)
rf_model.fit(X_train.relation_matrix, Y_train.values.flatten())
rf_y_pred = rf_model.predict(Xtest)
train_labels = []
test_arrays = []
test_labels = []
for email in emails:
email_id = email.id
prefix_train_pos = 'email_' + str(email_id)
if email_id % 5 != 0:
train_arrays.append(model.docvecs[prefix_train_pos])
train_labels.append(int(email.label))
else:
test_arrays.append(model.docvecs[prefix_train_pos])
test_labels.append(int(email.label))
classifier = SVC()
classifier.fit(numpy.array(train_arrays), numpy.array(train_labels))
print("Overall score is %f." % classifier.score(numpy.array(test_arrays), numpy.array(test_labels)))
corrects = []
wrongs = []
for email in emails:
email_id = email.id
prefix_train_pos = 'email_' + str(email_id)
if email_id % 5 == 0:
prediction = classifier.predict([model.docvecs[prefix_train_pos]])[0]
actual = int(email.label)
if prediction != actual:
wrongs.append((email.id, prediction, actual))
else:
# print(max(classifier.predict_proba([model.docvecs[prefix_train_pos]])[0]), actual)
def train_and_predict(samples, labels, feature_selector, inputDir, outputDir):
#test set
testDir = inputDir + "\\set_test"
testFiles = sorted([
join(testDir, f) for f in listdir(testDir) if isfile(join(testDir, f))
],
key=numericalSort)
# Different features for gender
testSamples_gender = cubeVoxelsVar_gender(testFiles)
# Same features for age and health
testSamples_age = cubeVoxelsVar_age(testFiles)
testSamples_health = testSamples_age
testSamples = [testSamples_gender, testSamples_age, testSamples_health]
#2D array to report final prediction in format (ID,Prediction)
final = [[0 for j in range(4)] for i in range(1 + 138 * 3)]
final[0][0] = 'ID'
final[0][1] = 'Sample'
final[0][2] = 'Label'
final[0][3] = 'Predicted'
total_labels = ['gender', 'age', 'health']
for label in range(3):
print("Prediction label 1 started!")
id_count = label
#Training logistic regression
logRegrL1 = linear_model.LogisticRegression()
logRegrL1.fit(samples[label], labels[label])
#Training SVM with linear kernel
svmLin = SVC(kernel='linear')
svmLin.fit(samples[label], labels[label])
#Training Random Forest Classifier
rfc = RandomForestClassifier(n_estimators=100)
rfc.fit(samples[label], labels[label])
print("Training complete!")
# Do feature selection only for age and health
if label == 0:
testSamples_curr = testSamples[label]
else:
testSamples_curr = feature_selector[label].transform(
testSamples[label])
print(len(testSamples_curr))
id = label + 1
#Predict gender, age and health status of test image using each of the 3 models trained above
for sampleNum, sample in enumerate(testSamples_curr):
predictionL1 = logRegrL1.predict(sample)
predictionSvmLin = svmLin.predict(sample)
predictionRfc = rfc.predict(sample)
final[id][0] = id_count
final[id][1] = sampleNum
final[id][2] = total_labels[label]
votes = predictionL1[0] + predictionSvmLin[0] + predictionRfc[0]
final[id][3] = 'TRUE' if votes >= 2.0 else 'FALSE'
id = id + 3
id_count = id_count + 3
print('Prediction done!')
#Save csv file in the output directory with name final_sub.csv
np.savetxt(outputDir + "\\final_sub.csv", final, delimiter=',', fmt='%s')
"Sidorovbigramsdeprel"
]
print("train: ", len(tweets_training))
print("test: ", len(tweets_test))
X, X_test, feature_name, feature_index = feature_manager.create_feature_space(
tweets_training, feature_type, tweets_test)
print(feature_name)
print("feature space dimension X:", X.shape)
print("feature space dimension X_test:", X_test.shape)
clf = SVC(kernel="linear")
clf.fit(X, labels_training)
test_predict = clf.predict(X_test)
"""prec, recall, f, support = precision_recall_fscore_support(
labels_test,
test_predict,
beta=1)
accuracy = accuracy_score(
test_predict,
labels_test
)
print(prec, recall, f, support )
print(accuracy)"""
for i in range(0, len(tweets_test)):
label_string += '0 '
fopen.close()
# generate the
label_vec = numpy.fromstring(label_string.strip(), dtype=int, sep=' ')
print 'Totally we get %s labels' % (label_vec.shape[0]) # for debugging
# create the feature matrix, in which each row represents a video
video_num = len(video_list)
feat_mat = numpy.zeros([video_num, feat_dim])
for i in xrange(video_num):
# BOW features of this video
feat_vec = numpy.genfromtxt(feat_dir + video_list[i],
dtype=numpy.float32,
delimiter=";")
assert (feat_vec.shape[0] == feat_dim)
# fill the feature vector to the matrix
feat_mat[i, :] = feat_vec
# initialize svm
svm = SVC(kernel=chi2_kernel)
# svm = SVC(probability=True)
# train the svm models
svm.fit(feat_mat, label_vec)
# finally save the k-means model
cPickle.dump(svm, open(output_file, "wb"), cPickle.HIGHEST_PROTOCOL)
print 'SVM trained successfully for event %s!' % (event_name)
audio_name=line.split(" ")[0]
#print count
count=count+1
# if (count%100==0):
# print count
label=line.split(" ")[1].split("\n")[0]
if "imtraj" in feat_dir:
feat_vec=import_imtraj_txt(feat_dir+audio_name+".spbof")
else:
feat_vec=np.genfromtxt(feat_dir+audio_name,delimiter=";")
if (label==event_name):
label=1
pos_count+=1
else:
label=0
neg_count+=1
if len(X)==0:
X=[feat_vec]
else:
X=np.append(X,[feat_vec],axis=0)
Y=Y+[label]
print "Data loading finished positive "+str(pos_count)+" negative "+str(neg_count)
#pipe_lrSVC=SVC(C=10,gamma=0.0001,probability=True)
pipe_lrSVC=SVC(probability=True)
#svm=LinearSVC(C=10)
#pipe_lrSVC=CalibratedClassifierCV(svm)
pipe_lrSVC.fit(preprocessing.scale(X),Y)
pickle.dump(pipe_lrSVC,open(output_file+'.pickle','wb'))
print 'SVM trained successfully for event %s!' % (event_name)+" round num %s" % (round_num)
X_val, Y_val = get_data("val_esea_real.lst", ranks)
trn_embedding = embed(X_trn, triplet_model)
val_embedding = embed(X_val, triplet_model)
# print (X_trn, X_val)
# print (trn_embedding, val_embedding)
# print (triplet_model.get_weights())
clf = SVC(
# class_weight='balanced',
probability=True,
# tol=1e-4,
)
clf.fit(trn_embedding, Y_trn)
print(clf.score(val_embedding, Y_val))
print(clf.predict_proba(val_embedding))
print(roc_auc_score(Y_val, clf.predict(val_embedding)))
print(classification_report(Y_val, clf.predict(val_embedding), digits=4))
all_files = [x[:-8] for x in os.listdir(ALL_FILES)]
X = [
pickle.load(open(os.path.join(FEATURE_PATH, x + ".fkmeans"), "rb"),
encoding='latin1') for x in all_files
]
# Y = [ranks[x.split()[0].strip()] for x in all_files]
proba = clf.predict_proba(embed(np.array(X), triplet_model))
# read in features
features = []
for video_id in video_ids:
feat_path = feat_dir + video_id + "." + feat_suffix
feature = [0]*feat_dim
if os.path.exists(feat_path) is True:
if feat_type == 'dense':
feature = numpy.genfromtxt(feat_path, delimiter=';')
else:
line = numpy.genfromtxt(feat_path, delimiter=' ', dtype=str)
if len(line.shape) == 0:
line = numpy.array([line])
for item in line:
if len(item) == 0:
continue
tokens = item.split(':')
key = int(tokens[0])-1
value = float(tokens[1])
if key < feat_dim:
feature[key] = value
features.append(feature)
# train svm
clf = SVC(probability=True)
clf.fit(features, labels)
# Dump model
with open(output_file, 'wb') as f:
cPickle.dump(clf, f)
print 'SVM trained successfully for event %s!' % (event_name)
[list(feature_names).index(f) for f in feature_filtered])
feature_index_filtered = numpy.concatenate(
feature_index_global[list(feature_index_filtered)])
#print(feature_name_global[feature_index_filtered])
X_filter = X[:, feature_index_filtered]
#print(feature_filtered,X.shape,X_filter.shape)
predict = []
golden = []
for index_train, index_test in kf:
X_train = X_filter[index_train]
X_test = X_filter[index_test]
clf = SVC(kernel='linear')
clf.fit(X_train, stance[index_train])
test_predict = clf.predict(X_test)
predict = numpy.concatenate((predict, test_predict))
golden = numpy.concatenate((golden, stance[index_test]))
prec, recall, f, support = precision_recall_fscore_support(golden,
predict,
beta=1)
accuracy = accuracy_score(golden, predict)
print('"' + (' '.join(feature_filtered)) + '"' + '\t' +
str(((f[0] + f[1] + f[2]) / 3)) + '\t' +
str(((f[1] + f[2]) / 2)) + '\t' + str(prec) + '\t' +
str(recall) + '\t' + str(f) + '\n')
feat_dir = "kmeans/"
feat_dim = 50
output_file = "mfcc_pred/svm.%s.model" % event_name
fread = open("list/train", "r")
clf = SVC(probability=True)
X, Y = [], []
for i in fread.readlines():
i = i.split(" ")
line = i[0]
label = i[1].replace('\n', '')
kmeans_path = "kmeans/" + line + ".kmeans.txt"
if os.path.exists(kmeans_path):
kmeans_feat = numpy.genfromtxt(kmeans_path, delimiter=";")
else:
kmeans_feat = numpy.zeros(feat_dim)
label = "NULL"
if label != event_name:
label = "NULL"
X.append(kmeans_feat)
Y.append(label)
X = numpy.array(X)
Y = numpy.array(Y)
clf.fit(X, Y)
cPickle.dump(clf, open(output_file, "wb"))
print " "
t2 = time.time() - t1
print "Time taken for training %s SVM : %f seconds" % (event_name, t2)
print 'SVM trained successfully for event %s!' % event_name
Y_all = Y
i = 1
else:
X_all = np.vstack((X_all, X))
Y_all = np.append(Y_all, Y)
i += 1
clf = SVC(kernel=laplacian_kernel)
features = ['.cnn.', '.mfcc.', '.asr.']
X_score = np.zeros((len(X_all), 3))
j = 0
for j, feature in enumerate(features):
X = X_all[:, dimension_i[j]:dimension_i[j] + dimension_i[j + 1]]
clf.fit(X, Y_all)
print('saving scores...')
X_score[:, j] = clf.decision_function(X)
# print (X_score[:, j])
np.save(event_name + feature + 'score', X_score[:, j])
cPickle.dump(clf, open(output_file + feature + 'score', "wb"))
clf = SVC(kernel='linear')
clf.fit(X_score, Y_all)
fread.close()
cPickle.dump(clf, open(output_file, "wb"))
print 'SVM trained successfully for event %s!' % (event_name)
test_size=0.2,
random_state=42)
#:# preprocessing
transform_pipeline = Pipeline([('scaler', StandardScaler())])
X_train = pd.DataFrame(transform_pipeline.fit_transform(X_train),
columns=X_train.columns)
#:# model
params = {'gamma': 5, 'kernel': 'sigmoid', 'probability': True}
classifier = SVC(**params)
classifier.fit(X_train, y_train)
#:# hash
#:# aad366f6d5961bc98783c2ad9fb3918d
md5 = hashlib.md5(str(classifier).encode('utf-8')).hexdigest()
print(f'md5: {md5}')
#:# audit
y_pred = classifier.predict(transform_pipeline.transform(X_test))
y_pred_proba = classifier.predict_proba(
transform_pipeline.transform(X_test))[:, 1]
tn, fp, fn, tp = confusion_matrix(y_test, y_pred).ravel()
print(f'acc: {accuracy_score(y_test, y_pred)}')
print(f'auc: {roc_auc_score(y_test, y_pred_proba)}')
file_read.close()
# Highly imbalanced data for training
X_df = np.array(X_df)
y_df = np.array(y_df)
#print(Counter(y_df))
#print(X_df.shape)
# Add SMOTE resampling for dealing with imbalanced issue
smote = SMOTE()
X_df_res, y_df_res = smote.fit_sample(X_df, y_df)
#print(Counter(y_df_res))
# Train svm
clf = SVC(kernel=chi2_kernel,
class_weight='balanced',
C=2.0,
gamma='scale',
probability=True,
random_state=0)
# clf = SVC(kernel = 'rbf', class_weight='balanced',
# C = 1.0, gamma = 'scale', probability=True, random_state=0)
# clf.fit(X_df, y_df)
clf.fit(X_df_res, y_df_res)
# Save the SVC model
with open(output_file, 'w') as f:
cPickle.dump(clf, f)
print 'SVM trained successfully for event %s!' % (event_name)
if Y_label != 'NULL' or random.random() > 0:
if Y_label == event_name:
Y = 1
else:
Y = 0
if i == 0:
X_all = X
Y_all = Y
i = 1
else:
X_all = np.vstack((X_all, X))
Y_all = np.append(Y_all, Y)
i += 1
# print (i)
# print (np.sum(X_all, axis = 1))
# print(X_all, Y_all)
clf = SVC(kernel=chi2_kernel)
# clf = SVC()
clf.fit(X_all, Y_all)
print(clf.score(X_all, Y_all))
print(clf.predict(X_all))
fread.close()
cPickle.dump(clf, open(output_file, "wb"))
print 'SVM trained successfully for event %s!' % (event_name)
i += 1
clf = SVC(kernel=laplacian_kernel)
features = ['.cnn.', '.mfcc.', '.asr.', '.cnn+mfcc.', '.cnn+mfcc+asr.']
X_score = np.zeros((len(X_all), len(features)))
j = 0
for j, feature in enumerate(features):
if j <= 2:
X = X_all[:, dimension_i[j]:dimension_i[j] + dimension_i[j + 1]]
elif j == 3:
X = X_all[:, 0:150]
elif j == 4:
X = X_all
clf.fit(X, Y_all)
print('saving scores...')
X_score[:, j] = clf.decision_function(X)
# print (X_score[:, j])
np.save(event_name + feature + 'score', X_score[:, j])
cPickle.dump(clf, open(output_file + feature + 'score', "wb"))
clf = SVC(kernel='linear')
clf.fit(X_score[:, [0, 1, 2, 4]], Y_all)
fread.close()
cPickle.dump(clf, open(output_file, "wb"))
print 'SVM trained successfully for event %s!' % (event_name)