def train_TG(
classes, train_fname, model_fname, iterations=2, L1=0, heldout_fname=None, crossValidation=None, verbose=False
):
"""
Train using binary maximum entropy model (i.e. logistic regression)
using stocastic gradient decent method. If heldout_fname is given
then we will report the accuracy on heldout data after each iteration.
If cross-validation is set to a number
(e.g. 5 for five-fold cross-validation)
then we will perform cross-validation and will report accuracy for each fold
as well as the average. You cannot specify both cross-validation and holdout
evaluation at the same time. If you do so then an error will be reported.
"""
TrainSeqFileReader = SEQUENTIAL_FILE_READER(train_fname)
trainVects = TrainSeqFileReader.read()
TrainSeqFileReader.close()
if heldout_fname:
HeldoutSeqFileReader = SEQUENTIAL_FILE_READER(heldout_fname)
heldoutVects = HeldoutSeqFileReader.read()
HeldoutSeqFileReader.close()
Learner = TruncatedGradient(classes)
Learner.total_iterations = iterations
Learner.c = L1
Learner.verbose = verbose
if crossValidation:
Learner.folds = crossValidation
if heldout_fname:
Learner.heldoutVects = heldoutVects["vects"]
no_features = classes * len(trainVects["featIDs"])
Learner.train(trainVects["vects"], no_features)
Learner.writeModel(no_features, model_fname)
pass
def load_model(self, model_fname, train_fname):
"""
Load the matrix from the model file.
"""
self.Cinv = genfromtxt("%s.matrix" % model_fname)
train_file = SEQUENTIAL_FILE_READER(train_fname)
train_vects = train_file.read()
train_file.close()
(self.D, self.t) = get_train_data(train_vects)
para_file = open(model_fname)
self.beta = float(para_file.readline().split()[1])
kernel_type = para_file.readline().strip().split("\t")[1]
if kernel_type == "GAUSSIAN_QUADRATIC_KERNEL":
self.kernel = GAUSSIAN_QUADRATIC_KERNEL()
for i in range(0, 4):
(para, val) = para_file.readline().strip().split()
if para == "theta_0":
self.kernel.theta_0 = float(val)
elif para == "theta_1":
self.kernel.theta_1 = float(val)
elif para == "theta_2":
self.kernel.theta_2 = float(val)
elif para == "theta_3":
self.kernel.theta_3 = float(val)
para_file.close()
pass
def predict_GPR(test_fname,
train_fname,
model_fname,
output_fname=None,
accuracy=False):
"""
Predict the outputs for the test instances.
If the output is not specified, then
write to the standard output.
"""
test_file = SEQUENTIAL_FILE_READER(test_fname)
test_vects = test_file.read()
test_file.close()
learner = GPR()
learner.load_model(model_fname, train_fname)
count = 0
error = 0
if output_fname:
output_file = SEQUENTIAL_FILE_WRITER(output_fname)
else:
output_file = SEQUENTIAL_FILE_WRITER(None, "STDOUT")
for v in test_vects["vects"]:
(mean, variance) = learner.predict(v)
output_file.writeLine("%f\t%f\n" % (mean, variance))
if accuracy:
error += (v.label - mean)**2
count += 1
error = sqrt(error) / float(count)
if accuracy:
output_file.writeLine("RMSE = %f\n" % error)
output_file.close()
pass
def train_GPR(train_fname,
model_fname,
verbose=True,
beta=1,
theta_0=None,
theta_1=None,
theta_2=None,
theta_3=None):
"""
This is the utility function used to train a regression
model using Gaussian Process.
"""
train_file = SEQUENTIAL_FILE_READER(train_fname)
train_vects = train_file.read()
train_file.close()
learner = GPR()
learner.verbose = verbose
learner.beta = beta
kernel = GAUSSIAN_QUADRATIC_KERNEL()
if theta_0:
kernel.theta_0 = theta_0
if theta_1:
kernel.theta_1 = theta_1
if theta_2:
kernel.theta_2 = theta_2
if theta_3:
kernel.theta_3 = theta_3
learner.set_kernel(kernel)
learner.train(train_vects)
learner.save_model(model_fname)
pass
def load_model(self, model_fname, train_fname):
"""
Load the matrix from the model file.
"""
self.Cinv = genfromtxt("%s.matrix" % model_fname)
train_file = SEQUENTIAL_FILE_READER(train_fname)
train_vects = train_file.read()
train_file.close()
(self.D, self.t) = get_train_data(train_vects)
para_file = open(model_fname)
self.beta = float(para_file.readline().split()[1])
kernel_type = para_file.readline().strip().split("\t")[1]
if kernel_type == "GAUSSIAN_QUADRATIC_KERNEL":
self.kernel = GAUSSIAN_QUADRATIC_KERNEL()
for i in range(0,4):
(para,val) = para_file.readline().strip().split()
if para == "theta_0":
self.kernel.theta_0 = float(val)
elif para == "theta_1":
self.kernel.theta_1 = float(val)
elif para == "theta_2":
self.kernel.theta_2 = float(val)
elif para == "theta_3":
self.kernel.theta_3 = float(val)
para_file.close()
pass
def predict_GPR(test_fname,
train_fname,
model_fname,
output_fname=None,
accuracy=False):
"""
Predict the outputs for the test instances.
If the output is not specified, then
write to the standard output.
"""
test_file = SEQUENTIAL_FILE_READER(test_fname)
test_vects = test_file.read()
test_file.close()
learner = GPR()
learner.load_model(model_fname, train_fname)
count = 0
error = 0
if output_fname:
output_file = SEQUENTIAL_FILE_WRITER(output_fname)
else:
output_file = SEQUENTIAL_FILE_WRITER(None, "STDOUT")
for v in test_vects["vects"]:
(mean, variance) = learner.predict(v)
output_file.writeLine("%f\t%f\n" % (mean, variance))
if accuracy:
error += (v.label - mean) ** 2
count += 1
error = sqrt(error) / float(count)
if accuracy:
output_file.writeLine("RMSE = %f\n" % error)
output_file.close()
pass
def train_GPR(train_fname, model_fname,
verbose=True,
beta=1,
theta_0=None, theta_1=None,
theta_2=None, theta_3=None):
"""
This is the utility function used to train a regression
model using Gaussian Process.
"""
train_file = SEQUENTIAL_FILE_READER(train_fname)
train_vects = train_file.read()
train_file.close()
learner = GPR()
learner.verbose = verbose
learner.beta = beta
kernel = GAUSSIAN_QUADRATIC_KERNEL()
if theta_0:
kernel.theta_0 = theta_0
if theta_1:
kernel.theta_1 = theta_1
if theta_2:
kernel.theta_2 = theta_2
if theta_3:
kernel.theta_3 = theta_3
learner.set_kernel(kernel)
learner.train(train_vects)
learner.save_model(model_fname)
pass
def train_SGD(classes,
train_fname,
model_fname,
iterations=2,
L2=0,
heldout_fname=None,
crossValidation=None,
verbose=False):
"""
Train using binary maximum entropy model (i.e. logistic regression)
using stocastic gradient decent method. If heldout_fname is given
then we will report the accuracy on heldout data after each iteration.
If cross-validation is set to a number
(e.g. 5 for five-fold cross-validation)
then we will perform cross-validation and will report accuracy for each fold
as well as the average. You cannot specify both cross-validation and holdout
evaluation at the same time. If you do so then an error will be reported.
"""
TrainSeqFileReader = SEQUENTIAL_FILE_READER(train_fname)
trainVects = TrainSeqFileReader.read()
TrainSeqFileReader.close()
heldoutVects = None
if heldout_fname:
HeldoutSeqFileReader = SEQUENTIAL_FILE_READER(heldout_fname)
heldoutVects = HeldoutSeqFileReader.read()
HeldoutSeqFileReader.close()
Learner = SGD(classes)
Learner.total_iterations = iterations
Learner.c = L2
Learner.verbose = verbose
if crossValidation:
Learner.folds = crossValidation
if heldout_fname:
Learner.heldoutVects = heldoutVects["vects"]
no_features = classes * len(trainVects["featIDs"])
Learner.train(trainVects["vects"], no_features)
print "Writing the model... %s" % model_fname
Learner.writeModel(no_features, model_fname)
pass