def fit(self, X, y): #check_estimator wants var names X, y
mappings, representations = msda.mSDA(X, self.prob_corruption,
self.num_layers,
self.use_nonlinearity)
self.mappings = mappings
msda_train_data = representations[-1]
self.clf.fit(msda_train_data, y)
def predict(self, X):
maps, test_reps = msda.mSDA(X, self.prob_corruption, self.num_layers,
self.use_nonlinearity, self.mappings)
msda_test_data = test_reps[-1]
test_predictions = self.clf.predict(msda_test_data)
return test_predictions
num_layers = 3
subproblem_size = 400
#need to transpose data to be in the right format (#features x #data) for mSDA
#specifically, the deep representation is the output from the last layer
#with low dimensional approximation described in paper
'''
subproblem_mappings, subseq_mappings, representations = msda.mSDA_lowDimApprox(train_data, prob_corruption,
num_layers, subproblem_size)
train_deepRep = representations[:,:,-1]
#use same weights as on training features to transform test data
test_deepRep = msda.mSDA_lowDimApprox(test_data, prob_corruption, num_layers, subproblem_size,
subproblem_mappings, subseq_mappings)[2][:,:,-1]
'''
#without low dimensional approximation
train_mappings, train_reps = msda.mSDA(train_data, prob_corruption, num_layers)
train_deepRep = train_reps[:,:,-1]
#use same weights as on training features to transform test data
test_deepRep = msda.mSDA(test_data, prob_corruption, num_layers, train_mappings)[1][:,:,-1]
#'''
#sklearn requires (#data x #features) so transpose back
train_deepRep = train_deepRep.transpose()
test_deepRep = test_deepRep.transpose()
after_msda = time.time()
print("used msda in %s seconds" % (after_msda - before_msda))
print "Shape of msda train rep: ", train_deepRep.shape
print "Shape of msda test rep: ", test_deepRep.shape
#...and classify with linear SVM
num_layers = 3
subproblem_size = 400
#need to transpose data to be in the right format (#features x #data) for mSDA
#specifically, the deep representation is the output from the last layer
#with low dimensional approximation described in paper
'''
subproblem_mappings, subseq_mappings, representations = msda.mSDA_lowDimApprox(train_data, prob_corruption,
num_layers, subproblem_size)
train_deepRep = representations[:,:,-1]
#use same weights as on training features to transform test data
test_deepRep = msda.mSDA_lowDimApprox(test_data, prob_corruption, num_layers, subproblem_size,
subproblem_mappings, subseq_mappings)[2][:,:,-1]
'''
#without low dimensional approximation
train_mappings, train_reps = msda.mSDA(train_data, prob_corruption, num_layers)
train_deepRep = train_reps[:, :, -1]
#use same weights as on training features to transform test data
test_deepRep = msda.mSDA(test_data, prob_corruption, num_layers,
train_mappings)[1][:, :, -1]
#'''
#sklearn requires (#data x #features) so transpose back
train_deepRep = train_deepRep.transpose()
test_deepRep = test_deepRep.transpose()
after_msda = time.time()
print("used msda in %s seconds" % (after_msda - before_msda))
print "Shape of msda train rep: ", train_deepRep.shape
print "Shape of msda test rep: ", test_deepRep.shape
def predict(self, X): maps, test_reps = msda.mSDA(X, self.prob_corruption, self.num_layers, self.use_nonlinearity, self.mappings) msda_test_data = test_reps[-1] test_predictions = self.clf.predict(msda_test_data) return test_predictions
def fit(self, X, y): #check_estimator wants var names X, y mappings, representations = msda.mSDA(X, self.prob_corruption, self.num_layers, self.use_nonlinearity) self.mappings = mappings msda_train_data = representations[-1] self.clf.fit(msda_train_data, y)