def calculate_race():
correct = 0
answers = []
input = []
count = 0
for d in data:
answers.append(question2b_race_truth.truth[count])
input.append(d)
if count == 49:
break
count += 1
prob = svmutil.svm_problem(answers, input)
param = svmutil.svm_parameter('-t 2 -c 4')
param.cross_validation = 1
param.nr_fold = 10
cv = svmutil.svm_train(prob, param)
param = svmutil.svm_parameter('-t 2 -c 4')
m = svmutil.svm_train(prob, param)
count = 0
for d in data:
if count < 50:
count += 1
continue
else:
x0, max_idx = gen_svm_nodearray(d)
p = libsvm.svm_predict(m, x0)
if p == question2b_race_truth.truth[count]:
correct += 1
count += 1
return cv, correct / float(50) * 100
def TrainSvmRbf2(Y, X, sweep_c=range(-5,5), sweep_g=range(-5,5)):
num_negatives = float(Y.count(-1))
num_positives = float(Y.count(1))
best_c = -1
best_g = -1
best_acc = -1
for c_pow in sweep_c:
for g_pow in sweep_g:
current_c = np.power(2.0,c_pow)
current_g = np.power(2.0,g_pow)
prob = svm.svm_problem(Y,X)
param = svm.svm_parameter('-t 2 -c %f -g %f -w-1 %f -w1 %f -q' % (current_c,
current_g,
100/num_negatives,
100/num_positives))
current_pos_acc, current_neg_acc = CrossValidate(Y, X, param)
current_acc = current_pos_acc
print 'c = %f, g = %f, cv acc = %f, neg acc = %f' % (current_c, current_g, current_acc,
current_neg_acc)
if best_acc < current_acc:
best_acc = current_acc
best_c = current_c
best_g = current_g
prob = svm.svm_problem(Y,X)
param = svm.svm_parameter('-t 2 -c %f -g %f -w-1 %f -w1 %f -q' % (best_c, best_g,
100/num_negatives,
100/num_positives))
svm_model = svm.svm_train(prob, param)
p_labs, p_acc, p_vals = svm.svm_predict(Y, X, svm_model, '-q')
pdb.set_trace()
return svm_model
def TrainSvmLinear2(Y, X, sweep_c=range(-2,18)):
num_positives = float(Y.count(1))
num_negatives = float(Y.count(-1))
best_c = -1
best_acc = -1
for c_pow in sweep_c:
current_c = np.power(2.0,c_pow)
param = svm.svm_parameter('-t 0 -c %f -w-1 %f -w1 %f -q' % (current_c,
100/num_negatives,
100/num_positives))
current_pos_acc, current_neg_acc = CrossValidate(Y, X, param)
current_acc = current_pos_acc
print '%f, %f, %f' % (current_c, current_acc, current_neg_acc)
if best_acc < current_acc:
best_acc = current_acc
best_c = current_c
prob = svm.svm_problem(Y,X)
param = svm.svm_parameter('-t 0 -c %f -w-1 %f -w1 %f -q' % (best_c,
100/num_negatives,
100/num_positives))
svm_model = svm.svm_train(prob, param)
p_labs, p_acc, p_vals = svm.svm_predict(Y, X, svm_model, '-q')
return svm_model
def TrainSvmLinear(Y, X, sweep_c=range(-2,8)):
num_positives = float(Y.count(1))
num_negatives = float(Y.count(-1))
best_c = -1
best_acc = -1
for c_pow in sweep_c:
current_c = np.power(2.0,c_pow)
prob = svm.svm_problem(Y,X)
param = svm.svm_parameter('-v 5 -t 0 -c %f -w-1 %f -w1 %f -q' % (current_c,
100/num_negatives,
100/num_positives))
current_acc = svm.svm_train(prob, param)
print '%f, %f' % (current_c, current_acc)
if best_acc < current_acc:
best_acc = current_acc
best_c = current_c
# recompute accuracy
param = svm.svm_parameter('-t 0 -c %f -w-1 %f -w1 %f -q' % (best_c,
100/num_negatives,
100/num_positives))
svm_model = svm.svm_train(prob, param)
p_labs, p_acc, p_vals = svm.svm_predict(Y, X, svm_model, '-q')
prob = svm.svm_problem(Y,X)
param = svm.svm_parameter('-t 0 -c %f -w-1 %f -w1 %f -q' % (best_c,
100/num_negatives,
100/num_positives))
svm_model = svm.svm_train(prob, param)
p_labs, p_acc, p_vals = svm.svm_predict(Y, X, svm_model, '-q')
pdb.set_trace()
return svm_model
def train(self):
for i in range(4):
self.convert()
#rbf
param1 = svmutil.svm_parameter("-t 2 -b 1 -c 1 -g 0.001")
param2 = svmutil.svm_parameter("-t 2 -b 1 -c 0.1 -g 0.001")
self.mr.append(svmutil.svm_train(self.problem[0], param1))#hist
self.mr.append(svmutil.svm_train(self.problem[1], param2))#vector
#linear
param3 = svmutil.svm_parameter("-t 0 -b 1 -c 0.1")
param4 = svmutil.svm_parameter("-t 0 -b 1 -c 0.01")
self.ml.append(svmutil.svm_train(self.problem[0], param3))#hist
self.ml.append(svmutil.svm_train(self.problem[1], param4))#vector
self.images = self.images[1:]+self.images[:1]
def main(path, k):
prabs = []
lns = []
for kk in range(0,k-1):
testLabel = []
trainPoint = []
trainLabel = []
testPoint = []
wcCount = 0
for u in os.listdir(path):
if u[-2:] == 'WC':r
wcCount += 1
filePath = path+u
WC = pickle.load(open(filePath, 'rb'))
if wcCount % k == 0 + kk:
testLabel.append(int(u[1]))
testPoint.append(WC)
else:
trainLabel.append(int(u[1]))
trainPoint.append(WC)
lns.append(len(testLabel))
prob = svmutil.svm_problem(trainLabel, trainPoint)
param = svmutil.svm_parameter('-t 0 -c 4 -b 1 -q')
m = svmutil.svm_train(prob, param)
svmutil.svm_save_model('n.model', m)
p_label, p_acc, p_val = svmutil.svm_predict(testLabel, testPoint, m, '-b 1')
prabs.append(p_acc[0])
def kfold(data, labels, k):
try:
import svmutil
except:
return 0
prabs = []
for xxx in range(0, 10):
picks = np.random.choice(len(data), len(data) / k, replace=False)
testLabel = labels[picks]
testPoint = data[picks]
trainPoint = data[np.setdiff1d(range(0, len(data)), picks)]
trainLabel = labels[np.setdiff1d(range(0, len(data)), picks)]
trainLabel = trainLabel.tolist()
trainPoint = trainPoint.tolist()
prob = svmutil.svm_problem(trainLabel, trainPoint)
param = svmutil.svm_parameter('-t 3 -c 4 -b 1 -q')
testLabel = testLabel.tolist()
testPoint = testPoint.tolist()
m = svmutil.svm_train(prob, param)
svmutil.svm_save_model('n.model', m)
p_label, p_acc, p_val = svmutil.svm_predict(testLabel, testPoint, m, '-b 1')
prabs.append(p_acc[0])
print sum(prabs) / float(len(prabs))
print 'std' + str(np.std(prabs))
return sum(prabs) / float(len(prabs))
def train(self,x,y):
"""
training using y=list,x=dict
parameter = string of parameters
"""
prob=su.svm_problem(y,x)
para=""
para+= "-s %d -t %d -d %d -g %f -r %f -c %f -n %f -p %f -e %f -b %d" %\
(
self.type,
self.kernel,
self.degree,
self.gamma,
self.coef0,
self.c,
self.nu,
self.p,
self.eps,
self.prob
)
if(self.v!=0):
para+=" -v %d" % self.v
if(self.q!=0):
para+= " -q"
print para
para1=su.svm_parameter(para)
self.model=su.svm_train(prob,para1)
return True
def train(cls, featuresets, params="-t 0 -q"):
"""Train a classifier using the given featuresets.
Args:
featuresets: List of featuresets.
params: Parameter string to pass to svmutil.svm_parameter.
Returns:
SvmClassifier object.
"""
all_features = set()
all_labels = set()
for featuredict, label in featuresets:
all_features.update(set(featuredict.keys()))
all_labels.add(label)
all_labels = sorted(all_labels)
all_features = sorted(all_features)
featureindex = dict(zip(all_features, range(1, len(all_features) + 1)))
labelindex = dict(zip(all_labels, range(1, len(all_labels) + 1)))
vectors, labels = cls.featuresets_to_svm(featureindex, labelindex,
featuresets)
prob = svmutil.svm_problem(labels, vectors)
param = svmutil.svm_parameter(params)
model = svmutil.svm_train(prob, param)
return cls(featureindex, labelindex, model)
def _lib_train_libsvm(user_tfidf, num_pos, num_neg, ignore):
sparse_user_tfidf, num_pos, num_neg = _convert_to_sparse_matrix(user_tfidf, num_pos, num_neg, ignore)
labels = ([1] * num_pos) + ([-1] * num_neg)
param = svm_parameter("-t %d" % KERNEL_NUMBER)
prob = svm_problem(labels, sparse_user_tfidf)
modellog = svm_train(prob, param)
return modellog
def trainSVM(trainMatrix, trainCategory):
svm.svm_model.predict = lambda self, x: svm.svm_predict([0], [x], self)[0][0]
prob = svm.svm_problem(trainCategory, trainMatrix)
param = svm.svm_parameter()
param.kernel_type = svm.LINEAR
param.C = 10
model = svm.svm_train(prob, param)
return model
def get_cross_val(x, y, x_val, y_val, gamma_c):
prob = svmutil.svm_problem(y, x)
param = svmutil.svm_parameter('-t 2 -q -c {0} -g {1}'.format(gamma_c.C, gamma_c.gamma))
m = svmutil.svm_train(prob, param)
svmutil.svm_save_model("model", m)
p_label_validation, p_acc_validation, p_val_validation = svmutil.svm_predict(y_val, x_val, m)
return p_acc_validation[0]
def getSvmParam(cross_validation_only = False):
param = svmutil.svm_parameter()
param.parse_options('-q') #quiet
if cross_validation_only:
param.cross_validation = True
param.nr_fold = 10
param.kernel_type = svmutil.LINEAR
param.C = 10.0 ** 2.0
return param
def __init__(self, data=[], labels=[], kernel=svmutil.RBF, c=10):
self.__svmparam__ = svmutil.svm_parameter('-q')
self.__svmparam__.kernel_type = kernel
self.__svmparam__.C = c
self.c = c
self.data = data
self.labels = labels
if len(data) > 0:
self.problem = svmutil.svm_problem(labels,data)
self.model = svmutil.svm_train(self.problem,self.__svmparam__,'-q')
else:
self.problem = None
self.model = None
def multiclass_train(valid_labels, labels, data, svm_parameters=None):
if svm_parameters == None:
# make default empty parameters
svm_parameters = []
for i in valid_labels:
svm_parameters.append(svmutil.svm_parameter())
models = []
for i in valid_labels:
oaa_labels = relabel_one_against_all(labels, i)
prob = svmutil.svm_problem(oaa_labels, data)
model = svmutil.svm_train(prob, svm_parameters[i])
models.append(model)
return models
def compute_auc_kern(data, labels, k=10, C=1.0, kern=6, gamma=None):
''' This is an interface to the extended libsvm implementation
with new kernels '''
kv = cross_validation.StratifiedKFold(labels, n_folds=k, random_state=1)
s = 0.0
for train_index, test_index in kv:
data_train = data[train_index]
labels_train = labels[train_index]
data_test = data[test_index]
labels_test = labels[test_index]
y_train = labels_train.tolist()
x_train = data_train.tolist()
prob = svmutil.svm_problem(y_train, x_train)
if gamma != None:
param = svmutil.svm_parameter("-t %i -c %.410f -q -g %i" % (kern, C, gamma))
else:
param = svmutil.svm_parameter("-t %i -c %.410f -q" % (kern, C))
model = svmutil.svm_train(prob, param)
y_test = labels_test.tolist()
x_test = data_test.tolist()
p_label, p_acc, p_val = svmutil.svm_predict(y_test, x_test, m=model)
fpr, tpr, thresholds = roc_curve(labels_test, p_val, pos_label=-1.0)
AUC = roc_auc_score(labels_test, p_val)
s += AUC
return s / k
def cSvmTrainSet(self):
dataMat = []
labelMat = []
file_pattern = re.compile('^%s-\d.rec' % self.legalName)
for fdata in os.listdir('data'):
if file_pattern.match(fdata):
data,label = loadDataSet('data/'+fdata,1)
else:
data,label = loadDataSet('data/'+fdata,-1)
dataMat+=data
labelMat+=label
libSvmFormatSaveInFile(dataMat,labelMat,'data_format/%s.mat' % self.legalName) # todo: duoxiancheng
y,x = svmutil.svm_read_problem('data_format/%s.mat' % self.legalName)
prob = svmutil.svm_problem(y,x,isKernel = True)
param = svmutil.svm_parameter('-t 0 ')
self.model = svmutil.svm_train(prob,param)
print self.model
def __init__(self, c=None, gamma=None, filename=None, neighbours=3, verbose=0):
self.neighbours = neighbours
self.verbose = verbose
if filename:
# If a filename is given, load a model from the given filename
if verbose:
print 'Loading classifier from "%s"...' % filename
self.model = svm_load_model(filename)
elif c == None or gamma == None:
raise Exception("Please specify both C and gamma.")
else:
self.param = svm_parameter()
self.param.C = c # Soft margin
self.param.kernel_type = RBF # Radial kernel type
self.param.gamma = gamma # Parameter for radial kernel
self.model = None
def main():
parser = argparse.ArgumentParser()
parser.add_argument('-r', '--resdir', type=str, required=True,
help="Results directory")
parser.add_argument('-f', '--feature', type=str, required=True,
help='feature to use to learn')
args = parser.parse_args()
SCORES_FPATH = os.path.join(args.resdir, 'scores.txt')
FEAT_DIR = os.path.join(args.resdir, 'features', args.feature)
scores = np.fromfile(SCORES_FPATH, sep='\n')
feats = []
for i in range(1, len(scores) + 1):
feats.append(np.fromfile(os.path.join(FEAT_DIR, str(i) + '.txt'), sep='\n').tolist())
# feats = np.array(feats)
print('Read all features')
params = svmutil.svm_parameter('-s 4 -t 2')
model = svmutil.svm_train(svmutil.svm_problem(scores, feats), params)
svmutil.svm_save_model(os.path.join(args.resdir, 'svr.model'), model)
print svmutil.svm_predict(scores, feats, model)
def main(path):
label = []
points = []
for u in os.listdir(path):
if u[-2:] == 'WC':
filePath = path+u
WC = pickle.load(open(filePath, 'rb'))
label.append(u[1])
points.append(WC)
label = [int(i) for i in label]
prob = svmutil.svm_problem(label, points)
param = svmutil.svm_parameter('-t 0 -c 4 -b 1')
m = svmutil.svm_train(prob, param)
svmutil.svm_save_model('n.model', m)
p_label, p_acc, p_val = svmutil.svm_predict(label, points, m, '-b 1')
return p_acc
def prob2_to_4():
x = np.array([
[ 1, 0],
[ 0, 1],
[ 0, -1],
[-1, 0],
[ 0, 2],
[ 0, -2],
[-2, 0]
])
y = np.array([-1,-1,-1,1,1,1,1])
print "===prob 2==="
xf = np.fliplr(x.copy())
print xf*xf - 2*x + np.array([[3,-3]])
print "===prob 3==="
prob = SVM.svm_problem(y.tolist(), x.tolist())
param = SVM.svm_parameter('-t 1 -c 100 -d 2 -r 1 -g 1') # very large C for hard margin
m = SVM.svm_train(prob, param)
sumA = 0
poly = [0] * 6 # xx, xy, yy, x, y, 1
for i in xrange(m.l):
idx = m.sv_indices[i]
alphay = m.sv_coef[0][i]
alpha = abs(m.sv_coef[0][i])
print "{:d} {:+1.2f}".format(idx, alpha)
sumA += alpha
v = x[idx-1]
poly[0] += alphay*v[0]*v[0]
poly[1] += alphay*v[1]*v[0]*2
poly[2] += alphay*v[1]*v[1]
poly[3] += alphay*v[0]*2
poly[4] += alphay*v[1]*2
# poly[5] += alphay*1 # no need because Sum(alphay) = 0
poly[5] -= m.rho[0]
print "Sum of alpha is {:1.3f}\nb = {}".format(sumA, m.rho[0])
print "{:+2.2f}xx {:+2.2f}xy {:+2.2f}yy {:+2.2f}x {:+2.2f}y {:+2.2f}".format(*poly)
def bench_svm(X, Y):
"""
bench with swig-generated wrappers that come with libsvm
"""
import svmutil
X1 = X.tolist()
Y1 = Y.tolist()
gc.collect()
# start time
tstart = datetime.now()
problem = svmutil.svm_problem(Y1, X1)
param = svmutil.svm_parameter()
param.svm_type=0
param.kernel_type=2
model = svmutil.svm_train(problem, param)
svmutil.svm_predict([0]*len(X1), X1, model)
delta = (datetime.now() - tstart)
# stop time
svm_results.append(delta.seconds + delta.microseconds/mu_second)
def svm_learning_curve(x, y):
m = len(y)
n = len(x)
steep = m / 100;
training_examples = []
train_accuracy = []
validation_accuracy = []
for i in range(steep, m, steep):
prob = svmutil.svm_problem(y[:i], x[:i])
param = svmutil.svm_parameter('-t 2 -q -c 0.01')
m = svmutil.svm_train(prob, param)
p_label_train, p_acc_train, p_val_train = svmutil.svm_predict(y[:i], x[:i], m)
p_label_validation, p_acc_validation, p_val_validation = svmutil.svm_predict(y[i:], x[i:], m)
print p_acc_train[0], "\t", p_acc_validation[0], "\n"
training_examples.append(i)
train_accuracy.append(p_acc_train[0])
validation_accuracy.append(p_acc_validation[0])
return training_examples, train_accuracy, validation_accuracy
def __init__(self, kernel=None, classifier=None, probability=True, params=None,
input_dim=None, output_dim=None, dtype=None):
"""
kernel -- The kernel to use
classifier -- The type of the SVM
params -- a dict of parameters to be passed to the svm_parameter
probability -- Must be set to True, if algorithms based on probability
shall be used.
"""
if not params:
params = {}
# initialise the parameter and be quiet
self.parameter = libsvmutil.svm_parameter("-q")
if probability:
# allow for probability estimates
self.parameter.probability = 1
super(LibSVMClassifier, self).__init__(input_dim=input_dim,
output_dim=output_dim,
dtype=dtype)
if kernel:
self.set_kernel(kernel)
if classifier:
self.set_classifier(classifier)
# set all other parameters
for k, v in params.iteritems():
if not k in self.parameter._names:
# check that the name is a valid parameter
msg = "'{}' is not a valid parameter for libsvm".format(k)
raise mdp.NodeException(msg)
if hasattr(self.parameter, k):
setattr(self.parameter, k, v)
else:
msg = "'svm_parameter' has no attribute {}".format(k)
raise AttributeError(msg)
p_label_train, p_acc_train, p_val_train = svmutil.svm_predict(y[:i], x[:i], m)
p_label_validation, p_acc_validation, p_val_validation = svmutil.svm_predict(y[i:], x[i:], m)
print p_acc_train[0], "\t", p_acc_validation[0], "\n"
training_examples.append(i)
train_accuracy.append(p_acc_train[0])
validation_accuracy.append(p_acc_validation[0])
return training_examples, train_accuracy, validation_accuracy
def get_cross_val(x, y, x_val, y_val, gamma_c):
prob = svmutil.svm_problem(y, x)
param = svmutil.svm_parameter('-t 2 -q -c {0} -g {1}'.format(gamma_c.C, gamma_c.gamma))
m = svmutil.svm_train(prob, param)
svmutil.svm_save_model("model", m)
p_label_validation, p_acc_validation, p_val_validation = svmutil.svm_predict(y_val, x_val, m)
return p_acc_validation[0]
if __name__ == '__main__':
y, x = svmutil.svm_read_problem("char_recon_shuffled.db")
gamma = 1.0 / (2.0 * (3.0 ** 7) ** 2)
C = 3.0 ** 3.0
prob = svmutil.svm_problem(y, x)
param = svmutil.svm_parameter('-t 2 -q -c {0} -g {1}'.format(C, gamma))
m = svmutil.svm_train(prob, param)
svmutil.svm_save_model("model", m)
def _complete_training(self, debug=False):
""" Iterate over the complete data to get the initial model """
########## read complexities file if given ##########
if self.complexities_path is not None:
import yaml
complexities_file=open(self.complexities_path, 'r')
complexities = yaml.load(complexities_file)
# nr of channels = nr of features (==dim) / features_per_channel
if not 'features_per_channel' in complexities:
complexities['features_per_channel'] = 1
self.complexity = complexities[
round(self.dim/complexities['features_per_channel'])]
self._log("Read complexity %s from file. Dimension is %s" %
(self.complexity, self.dim), level=logging.INFO)
# not compatible with regression!
# self._log("Instances of Class %s: %s, %s: %s" \
# % (self.classes[0],
# self.labels.count(self.classes.index(self.classes[0])),
# self.classes[1],
# self.labels.count(self.classes.index(self.classes[1]))))
# instead this?:
self._log("Performing training of SVM.")
########## Calculation of default gamma ##########
self.calculate_gamma()
self.num_samples = len(self.samples)
# nr_weight is the number of elements in the array weight_label and
# weight. Each weight[i] corresponds to weight_label[i], meaning that
# the penalty of class weight_label[i] is scaled by a factor of
# weight[i]. If you do not want to change penalty for any of the
# classes, just set nr_weight to 0.
########## preparation of the libsvm command ##########
# for probability output add "-b 1" to options
options = \
"-c %.42f -d %d -g %.42f -r %.42f -n %.42f -p %.42f -e %.20f -m %.42f" % \
(self.complexity, self.exponent, self.gamma,
self.offset, self.nu, self.epsilon, self.tolerance, 1000)
# use 1000MB instead of 100MB (default)
# options += " -b 1" un-comment this for probabilistic output!
if self.multinomial:
options += " -b 1"
for i,w in enumerate(self.weight):
options += " -w%d %.42f" % (i, w)
if self.kernel_type == 'LINEAR':
options += " -t 0"
elif self.kernel_type == 'POLY':
options += " -t 1"
elif self.kernel_type == 'RBF':
options += " -t 2"
elif self.kernel_type == 'SIGMOID':
options += " -t 3"
else:
self.kernel_type = 'LINEAR'
options += " -t 0"
warnings.warn("Kernel unknown! Precomputed Kernels are not " +
"yet implemented. Linear Kernel used.")
# PRECOMPUTED: kernel values in training_set_file
# (not yet implemented)
if self.svm_type == 'C-SVC':
options += " -s 0"
elif self.svm_type == 'nu-SVR':
options += " -s 1"
elif self.svm_type == 'one-class SVM':
options += " -s 2"
elif self.svm_type == 'epsilon-SVR':
options += " -s 3"
else:
options += " -s 0"
self.svm_type = 'C-SVC'
warnings.warn("SVM-type unknown. C-SVC will be used!")
if not self.debug:
options += " -q"
self._log("Libsvm is now quiet!")
old_libsvm_options = options
if self.max_iterations != 0:
options += " -i %d" % self.max_iterations
try:
param = svmutil.svm_parameter(options)
except ValueError:
param = svmutil.svm_parameter(old_libsvm_options)
self._log(
"Using max_iterations is not supported by the standard " +
"LIBSVM. Change your Python path to our customized version!",
level=logging.CRITICAL)
# transform labels with *label_function*
if self.str_label_function is not None:
self.label_function = eval(self.str_label_function)
self.labels = self.label_function(self.labels)
# build the classifier
# h = [map(float,list(data)) for data in self.samples]
problem = svmutil.svm_problem(self.labels, [
map(float, list(data)) for data in self.samples])
model = svmutil.svm_train(problem, param)
if not self.multinomial:
if (self.svm_type == 'C-SVC' or self.svm_type == 'one-class SVM') \
and self.kernel_type == 'LINEAR':
self.calculate_classification_vector(model)
if self.debug:
# This calculation is needed for further analysis
self.calculate_slack_variables(model)
print "LIBSVM Parameter:"
self.print_variables()
else:
# Slack variables are the same no matter which kernel is used
# This method is mainly used to reduce the number of samples
# being stored later on.
if self.debug:
self.calculate_slack_variables(model)
self.model = model
else:
self.model = model
# Slack variables are the same no matter which kernel is used
# This method is mainly used to reduce the number of samples
# being stored later on.
# read number of iterations needed to solve the problem
if self.max_iterations != 0:
try:
predictor_iterations = model.get_num_iterations()
self.classifier_information["~~Solver_Iterations~~"] = \
predictor_iterations
if predictor_iterations == 0 or \
predictor_iterations == numpy.Inf:
self.classifier_information["~~SVM_Converged~~"] = False
else:
self.classifier_information["~~SVM_Converged~~"] = True
except:
warnings.warn("Could not read state of the LibSVM Solver " +
"from the C-Library!")
try:
self.classifier_information["~~offset~~"] = self.b
self.classifier_information["~~w0~~"] = self.w[0]
self.classifier_information["~~w1~~"] = self.w[1]
except:
pass
self.delete_training_data()
# use targets to train one svm for each hidden neuron
print "Training SVMs..."
probs = []
params = []
svms = []
ws = []
bs = []
werrs = 0
for n in range(hl.n_out):
print "Hidden neuron: %d" % n,
print " Problem...",
if n == 0:
probs.append(svmutil.svm_problem(svm_targets[n], svm_inputs))
else:
probs.append(svmutil.svm_problem(svm_targets[n], None, tmpl=probs[0]))
params.append(svmutil.svm_parameter("-q -s 0 -t 0 -c 100"))
print " Training...",
svms.append(svmutil.svm_train(probs[n], params[n]))
print " Saving...",
svmutil.svm_save_model("hidden%04d.svm" % n, svms[n])
print " Testing..."
# get weights from SVM
w, b = get_svm_weights(svms[n], hl.n_in)
ws.append(w)
bs.append(b)
# test model
predv = numpy.dot(w, trsx.T) + b
pred = numpy.sign(predv)
pos = 0
#gamma_vals = np.linspace(0.4,1,3)
#gamma_vals = [0.05,0.03,0.01]
#c_vals = [math.pow(2, x) for x in range(100,104, 1)]
#gamma_vals = [math.pow(2, x) for x in range(100,104,1)]
#c_vals = [0.1,1]
#gamma_vals = [0.1,2]
print "c", c_vals
print "gamma", gamma_vals
redo = False
if redo == True:
accs = np.zeros((len(c_vals), len(gamma_vals)))
for c_index in range(len(c_vals)):
for gamma_index in range(len(gamma_vals)):
param = svmutil.svm_parameter('-s 0 -t 2 -c {0} -g {1}'.format(c_vals[c_index], gamma_vals[gamma_index]))
#accuracy = svmutil.svm_train(prob, param)
model = svmutil.svm_train(prob, param)
predicted_labels, accuracy, decision_vals = svmutil.svm_predict(paramsellabels, paramseldata, model)
accs[gamma_index, c_index] = accuracy[0]
gr = open("grid_results", 'w')
pickle.dump(accs, gr)
gr.close()
else:
gr = open("grid_results", 'r')
accs = pickle.load(gr)
gr.close()
caxis, gaxis = np.meshgrid(np.log(c_vals), np.log(gamma_vals))
#
# print prediction_accuracy(predicted_labels, testinglabels)
#===========================================================================
ttlabels = []
ttlabels.extend(traininglabels)
ttlabels.extend(testinglabels)
ttdata = []
ttdata.extend(trainingdata)
ttdata.extend(testingdata)
prob = svmutil.svm_problem(traininglabels, trainingdata)
#prob = svmutil.svm_problem(traininglabels.extend(testinglabels), trainingdata.extend(testingdata))
param = svmutil.svm_parameter('-s 0 -t 2 -c {0} -g {1}'.format(np.exp(0.5), np.exp(-3.56)))
#accuracy = svmutil.svm_train(prob, param)
model = svmutil.svm_train(prob, param)
predicted_labels, accuracy, decision_vals = svmutil.svm_predict(testinglabels, testingdata, model)
# constructing a confusion matrix
svm_confusion = make_confusion(testinglabels, predicted_labels)
print svm_confusion
svm_conf_file = open('svm_conf', 'w')
write_confusion(svm_confusion, svm_conf_file)
svm_conf_file.close()
def train(self, positive_classes, negatives):
"""
Train the supervised SVM classifier model.
The class label ``negative`` is reserved for the negative class.
If a model is already loaded, we will raise an exception in order to
prevent accidental overwrite.
NOTE:
This abstract method provides generalized error checking and
should be called via ``super`` in implementing methods.
:param positive_classes: Dictionary mapping positive class labels to
iterables of DescriptorElement training examples.
:type positive_classes:
dict[collections.Hashable,
collections.Iterable[smqtk.representation.DescriptorElement]]
:param negatives: Iterable of negative DescriptorElement examples.
:type negatives: collections.Iterable[smqtk.representation.DescriptorElement]
:raises ValueError: The ``negative`` label was found in the
``positive_classes`` dictionary. This is reserved for the negative
example class.
:raises ValueError: There were no positive or negative examples.
:raises RuntimeError: A model already exists in this instance.Following
through with training would overwrite this model. Throwing an
exception for information protection.
"""
super(LibSvmClassifier, self).train(positive_classes, negatives)
# Offset from 0 for positive class labels to use
# - not using label of 0 because we think libSVM wants positive labels
CLASS_LABEL_OFFSET = 1
# Stuff for debug reporting
etm_ri = None
param_debug = {"-q": ""}
if self._log.getEffectiveLevel() <= logging.DEBUG:
etm_ri = 1.0
param_debug = {}
# Form libSVM problem input values
self._log.debug("Formatting problem input")
train_labels = []
train_vectors = []
train_group_sizes = []
self.svm_label_map = {}
# Making SVM label assignment deterministic to alphabetic order
for i, l in enumerate(sorted(positive_classes), CLASS_LABEL_OFFSET):
# Map integer SVM label to semantic label
self.svm_label_map[i] = l
self._log.debug("-- class %d (%s)", i, l)
# requires a sequence, so making the iterable ``g`` a tuple
g = positive_classes[l]
if not isinstance(g, collections.Sequence):
g = tuple(g)
train_group_sizes.append(float(len(g)))
x = elements_to_matrix(g, report_interval=etm_ri)
x = self._norm_vector(x)
train_labels.extend([i] * x.shape[0])
train_vectors.extend(x.tolist())
del g, x
self._log.debug("-- negatives (-1)")
# Map integer SVM label to semantic label
self.svm_label_map[-1] = self.NEGATIVE_LABEL
# requires a sequence, so making the iterable ``negatives`` a tuple
if not isinstance(negatives, collections.Sequence):
negatives = tuple(negatives)
negatives_size = float(len(negatives))
x = elements_to_matrix(negatives, report_interval=etm_ri)
x = self._norm_vector(x)
train_labels.extend([-1] * x.shape[0])
train_vectors.extend(x.tolist())
del negatives, x
self._log.debug(
"Training elements: %d labels, %d vectors " "(should be the same)", len(train_labels), len(train_vectors)
)
self._log.debug("Forming train params")
#: :type: dict
params = deepcopy(self.train_params)
params.update(param_debug)
# Only need to calculate positive class weights when C-SVC type
if "-s" not in params or int(params["-s"]) == 0:
for i, n in enumerate(train_group_sizes, CLASS_LABEL_OFFSET):
params["-w" + str(i)] = max(1.0, negatives_size / float(n))
self._log.debug("Making parameters obj")
svm_params = svmutil.svm_parameter(self._gen_param_string(params))
self._log.debug("Creating SVM problem")
svm_problem = svm.svm_problem(train_labels, train_vectors)
self._log.debug("Training SVM model")
self.svm_model = svmutil.svm_train(svm_problem, svm_params)
self._log.debug("Training SVM model -- Done")
if self.svm_label_map_fp:
self._log.debug("saving file -- labels -- %s", self.svm_label_map_fp)
with open(self.svm_label_map_fp, "wb") as f:
cPickle.dump(self.svm_label_map, f)
if self.svm_model_fp:
self._log.debug("saving file -- model -- %s", self.svm_model_fp)
svmutil.svm_save_model(self.svm_model_fp, self.svm_model)
def train(self, class_examples=None, **kwds):
"""
Train the supervised classifier model.
If a model is already loaded, we will raise an exception in order to
prevent accidental overwrite.
If the same label is provided to both ``class_examples`` and ``kwds``,
the examples given to the reference in ``kwds`` will prevail.
:param class_examples: Dictionary mapping class labels to iterables of
DescriptorElement training examples.
:type class_examples: dict[collections.Hashable,
collections.Iterable[smqtk.representation.DescriptorElement]]
:param kwds: Keyword assignment of labels to iterables of
DescriptorElement training examples.
:type kwds: dict[str,
collections.Iterable[smqtk.representation.DescriptorElement]]
:raises ValueError: There were no class examples provided.
:raises ValueError: Less than 2 classes were given.
:raises RuntimeError: A model already exists in this instance.Following
through with training would overwrite this model. Throwing an
exception for information protection.
"""
class_examples = \
super(LibSvmClassifier, self).train(class_examples, **kwds)
# Offset from 0 for positive class labels to use
# - not using label of 0 because we think libSVM wants positive labels
CLASS_LABEL_OFFSET = 1
# Stuff for debug reporting
etm_ri = None
param_debug = {'-q': ''}
if self._log.getEffectiveLevel() <= logging.DEBUG:
etm_ri = 1.0
param_debug = {}
# Form libSVM problem input values
self._log.debug("Formatting problem input")
train_labels = []
train_vectors = []
train_group_sizes = [] # number of examples per class
self.svm_label_map = {}
# Making SVM label assignment deterministic to alphabetic order
for i, l in enumerate(sorted(class_examples), CLASS_LABEL_OFFSET):
# Map integer SVM label to semantic label
self.svm_label_map[i] = l
self._log.debug('-- class %d (%s)', i, l)
# requires a sequence, so making the iterable ``g`` a tuple
g = class_examples[l]
if not isinstance(g, collections.Sequence):
g = tuple(g)
train_group_sizes.append(float(len(g)))
x = elements_to_matrix(g, report_interval=etm_ri)
x = self._norm_vector(x)
train_labels.extend([i] * x.shape[0])
train_vectors.extend(x.tolist())
del g, x
assert len(train_labels) == len(train_vectors), \
"Count miss-match between parallel labels and descriptor vectors" \
"being sent to libSVM (%d != %d)" \
% (len(train_labels), len(train_vectors))
self._log.debug("Forming train params")
#: :type: dict
params = deepcopy(self.train_params)
params.update(param_debug)
# Calculating class weights for C-SVC SVM
if '-s' not in params or int(params['-s']) == 0:
total_examples = sum(train_group_sizes)
for i, n in enumerate(train_group_sizes, CLASS_LABEL_OFFSET):
# weight is the ratio of between number of other-class examples
# to the number of examples in this class.
other_class_examples = total_examples - n
w = max(1.0, other_class_examples / float(n))
params['-w' + str(i)] = w
self._log.debug("-- class '%s' weight: %s",
self.svm_label_map[i], w)
self._log.debug("Making parameters obj")
svm_params = svmutil.svm_parameter(self._gen_param_string(params))
self._log.debug("Creating SVM problem")
svm_problem = svm.svm_problem(train_labels, train_vectors)
self._log.debug("Training SVM model")
self.svm_model = svmutil.svm_train(svm_problem, svm_params)
self._log.debug("Training SVM model -- Done")
if self.svm_label_map_fp:
self._log.debug("saving file -- labels -- %s",
self.svm_label_map_fp)
with open(self.svm_label_map_fp, 'wb') as f:
cPickle.dump(self.svm_label_map, f, -1)
if self.svm_model_fp:
self._log.debug("saving file -- model -- %s", self.svm_model_fp)
svmutil.svm_save_model(self.svm_model_fp, self.svm_model)