def svm(self, h, name=""):
# Function that trains an SVM on the representation
# h given in argument.
print '--- Loading data'
self.jobman.save()
f = theano.function([self.x], [h])
def fix_theano(X, l):
return f(X)[0]
self.svm_data = root(self.jobman.state['dataset'],
prepro_fct_x=[(fix_theano, [])])
res = eval_fold(3, self.svm_data)
cPickle.dump(res, open('results.pkl', 'w'))
self.jobman.state['train_' + name] = res['train']
if self.jobman.state['valid_' + name] <= res['valid']:
self.jobman.state['valid_' + name] = res['valid']
self.jobman.state['test_' + name] = res['test']
self.jobman.save()
print name, res
return True
def __init__(self,jobman):
self.jobman = jobman
self.x = T.matrix()
self.y = T.imatrix()
self.data = root(self.jobman.state['dataset'],prepro_fct_x=[],prepro_fct_y=[])
self.shared_x,self.shared_y = self.data.get_train_labeled()
self.n_i,self.n_h = (self.data.get_sample_size(),self.jobman.state['n_h'])
self.featsplit = int(self.n_h * self.jobman.state['featsplit'])
self.cae = cae(self.n_i,self.n_h)
self.logreg = logistic(self.featsplit,self.jobman.state['nclass'])
def evalreg(self, name=""):
# This function is used only for monitoring purposes
print "--- Loading data"
self.jobman.save()
o = self.logrege.get_hidden(self.cae_emotions.get_hidden(self.x))
f = theano.function([self.x, self.y], hardmax(o, self.y))
def fix_label(Y, l):
return numpy.array(Y - 1, dtype="int32").flatten()
data = root(self.jobman.state["dataset"], prepro_fct_x=[], prepro_fct_y=[(fix_label, [])])
X, Y = data.get_train_labeled(shared_array=False)
print "--- Logreg Valid Accuracy:", f(X, Y)
return True
def evalreg(self,name=""):
# Function to evaluate the classification error
# when using the output of the logistic regression.
# This function is used only for monitoring purposes
print '--- Loading data'
self.jobman.save()
o = self.logrege.get_hidden(self.cae_emotions.get_hidden(self.x))
f = theano.function([self.x,self.y],hardmax(o,self.y))
data = root(self.jobman.state['dataset'],prepro_fct_x=[],prepro_fct_y=[])
X,Y = data.get_train_labeled(shared_array=False)
print '--- Logreg Valid Accuracy:',f(X,Y)
return True
def __init__(self, jobman):
self.jobman = jobman
self.x = T.matrix()
self.y = T.imatrix()
self.data = root(self.jobman.state['dataset'],
prepro_fct_x=[],
prepro_fct_y=[])
self.shared_x, self.shared_y = self.data.get_train_labeled()
self.n_i, self.n_h = (self.data.get_sample_size(),
self.jobman.state['n_h'])
self.featsplit = int(self.n_h * self.jobman.state['featsplit'])
self.cae = cae(self.n_i, self.n_h)
self.logreg = logistic(self.featsplit, self.jobman.state['nclass'])
def __init__(self, jobman):
self.jobman = jobman
self.x = T.matrix()
self.y = T.ivector()
def fix_label(Y, l):
return numpy.array(Y - 1, dtype="int32").flatten()
self.data = root(self.jobman.state["dataset"], prepro_fct_x=[], prepro_fct_y=[(fix_label, [])])
dataset_list = [
"/scratch/rifaisal/data/tfdconv/convfeat_patch14_1024_33_fold0_files_dict.pkl",
"/scratch/rifaisal/data/tfdconv/convfeat_patch14_1024_33_fold1_files_dict.pkl",
"/scratch/rifaisal/data/tfdconv/convfeat_patch14_1024_33_fold2_files_dict.pkl",
"/scratch/rifaisal/data/tfdconv/convfeat_patch14_1024_33_fold3_files_dict.pkl",
"/scratch/rifaisal/data/tfdconv/convfeat_patch14_1024_33_fold4_files_dict.pkl",
]
csplit = self.jobman.state["split"]
assert csplit < 5
self.data.fusion(dataset_list[csplit : csplit + 1])
x, y = shuffle_array(
self.data.get_train_unlabeled(shared_array=False), self.data.get_train_labeled(shared_array=False)
)
self.shared_x, self.shared_y = (theano.shared(value=x), theano.shared(value=y))
self.n_i, self.n_hi, self.n_he = (
self.data.get_sample_size(),
self.jobman.state["n_hi"],
self.jobman.state["n_he"],
)
self.cae_emotions = cae(self.n_i, self.n_he)
self.cae_identity = cae(self.n_i, self.n_hi)
self.logrege = logistic(self.n_he, 7)
if self.jobman.state["identity"]:
self.logregi = logistic(self.n_hi, 900)
def __init__(self, jobman):
self.jobman = jobman
self.x = T.matrix()
self.y = T.ivector()
def fix_label(Y, l):
return numpy.array(Y - 1, dtype='int32').flatten()
self.data = root(self.jobman.state['dataset'],
prepro_fct_x=[],
prepro_fct_y=[(fix_label, [])])
dataset_list = [
'/scratch/rifaisal/data/tfdconv/convfeat_patch14_1024_33_fold0_files_dict.pkl',
'/scratch/rifaisal/data/tfdconv/convfeat_patch14_1024_33_fold1_files_dict.pkl',
'/scratch/rifaisal/data/tfdconv/convfeat_patch14_1024_33_fold2_files_dict.pkl',
'/scratch/rifaisal/data/tfdconv/convfeat_patch14_1024_33_fold3_files_dict.pkl',
'/scratch/rifaisal/data/tfdconv/convfeat_patch14_1024_33_fold4_files_dict.pkl'
]
csplit = self.jobman.state['split']
assert csplit < 5
self.data.fusion(dataset_list[csplit:csplit + 1])
x, y = shuffle_array(self.data.get_train_unlabeled(shared_array=False),
self.data.get_train_labeled(shared_array=False))
self.shared_x, self.shared_y = (theano.shared(value=x),
theano.shared(value=y))
self.n_i, self.n_hi, self.n_he = (self.data.get_sample_size(),
self.jobman.state['n_hi'],
self.jobman.state['n_he'])
self.cae_emotions = cae(self.n_i, self.n_he)
self.cae_identity = cae(self.n_i, self.n_hi)
self.logrege = logistic(self.n_he, 7)
if self.jobman.state['identity']:
self.logregi = logistic(self.n_hi, 900)
def evalreg(self, name=""):
# This function is used only for monitoring purposes
print '--- Loading data'
self.jobman.save()
o = self.logrege.get_hidden(self.cae_emotions.get_hidden(self.x))
f = theano.function([self.x, self.y], hardmax(o, self.y))
def fix_label(Y, l):
return numpy.array(Y - 1, dtype='int32').flatten()
data = root(self.jobman.state['dataset'],
prepro_fct_x=[],
prepro_fct_y=[(fix_label, [])])
X, Y = data.get_train_labeled(shared_array=False)
print '--- Logreg Valid Accuracy:', f(X, Y)
return True
def evalreg(self, name=""):
# Function to evaluate the classification error
# when using the output of the logistic regression.
# This function is used only for monitoring purposes
print '--- Loading data'
self.jobman.save()
o = self.logrege.get_hidden(self.cae_emotions.get_hidden(self.x))
f = theano.function([self.x, self.y], hardmax(o, self.y))
data = root(self.jobman.state['dataset'],
prepro_fct_x=[],
prepro_fct_y=[])
X, Y = data.get_train_labeled(shared_array=False)
print '--- Logreg Valid Accuracy:', f(X, Y)
return True
def svm(self, h, name=""):
# Function that trains an SVM on the representation
# h given in argument.
print "--- Loading data"
self.jobman.save()
f = theano.function([self.x], [h])
def fix_theano(X, l):
return f(X)[0]
self.svm_data = root(self.jobman.state["dataset"], prepro_fct_x=[(fix_theano, [])])
res = eval_fold(3, self.svm_data)
cPickle.dump(res, open("results.pkl", "w"))
self.jobman.state["train_" + name] = res["train"]
if self.jobman.state["valid_" + name] <= res["valid"]:
self.jobman.state["valid_" + name] = res["valid"]
self.jobman.state["test_" + name] = res["test"]
self.jobman.save()
print name, res
return True
def svm(self,h,name=""):
# Function that trains an SVM on the representation
# h given in argument.
print '--- Loading data'
self.jobman.save()
f = theano.function([self.x],[h])
def fix_theano(X,l):
return f(X)[0]
self.svm_data = root(self.jobman.state['dataset'],prepro_fct_x=[(fix_theano,[])])
res = eval_fold(3,self.svm_data)
cPickle.dump(res,open('results.pkl','w'))
self.jobman.state['train_'+name] = res['train']
if self.jobman.state['valid_'+name] <= res['valid']:
self.jobman.state['valid_'+name] = res['valid']
self.jobman.state['test_'+name] = res['test']
self.jobman.save()
print name,res
return True
valid_set_x, valid_set_y = dataset.get_all_valid(shared_array=False)
test_set_x, test_set_y = dataset.get_all_test(shared_array=False)
train_set_y = train_set_y.flatten()
valid_set_y = valid_set_y.flatten()
test_set_y = test_set_y.flatten()
bestC, res = OptimizeLinearSVM(([train_set_x, train_set_y],\
[valid_set_x, valid_set_y],\
[test_set_x, test_set_y]),\
MAXSTEPS=nsteps)
best = {
'C': bestC,
'train': res[bestC]['train_acc'],
'valid': res[bestC]['valid_acc'],
'test': res[bestC]['test_acc'],
'Optimize': res
}
return best
if __name__ == '__main__':
data = root(
'/scratch/rifaisal/data/paradata/msrparaphrasefullconvl2_files_dict.pkl',
prepro_fct_x=[(scale, [])])
print 'Starting evaluation...'
eval_fold(nsteps=30, dataset=data)
print dic
'''
print "Loading data..."
train_set_x,train_set_y = dataset.get_all_train_labeled(shared_array=False)
valid_set_x,valid_set_y = dataset.get_all_valid(shared_array=False)
test_set_x,test_set_y = dataset.get_all_test(shared_array=False)
train_set_y = train_set_y.flatten()
valid_set_y = valid_set_y.flatten()
test_set_y = test_set_y.flatten()
bestC, res = OptimizeLinearSVM(([train_set_x, train_set_y],\
[valid_set_x, valid_set_y],\
[test_set_x, test_set_y]),\
MAXSTEPS=nsteps)
best = {'C':bestC,'train':res[bestC]['train_acc'],
'valid':res[bestC]['valid_acc'],
'test':res[bestC]['test_acc'],
'Optimize':res}
return best
if __name__ == '__main__':
data = root('/scratch/rifaisal/data/paradata/msrparaphrasefullconvl2_files_dict.pkl',prepro_fct_x=[(scale,[])])
print 'Starting evaluation...'
eval_fold(nsteps=30, dataset = data)
print dic
