def tune_learner(train_X): train_len = len(train_X) new_train_index = np.random.choice(range(train_len), train_len * 0.7) new_tune_index = list(set(range(train_len)) - set(new_train_index)) new_train_X = train_X[new_train_index] new_train_Y = train_Y[new_train_index] new_tune_X = train_X[new_tune_index] new_tune_Y = train_Y[new_tune_index] clf = learner(new_train_X, new_train_Y, new_tune_X, new_tune_Y, goal) tuner = DE_Tune_ML(clf, clf.get_param(), target_class, goal) return tuner.Tune()
def test_unknown_data():
n = 100
l = 1
d = int(np.ceil(n**(0.15)))
rho = 0.01
N = int(np.log(n)/rho**2)
Y = generate_Y(n, l, d, rho, N)
Gp, Hp = learner(n, l, d, rho, Y, np.zeros((l, n,n)), np.zeros((N,n)))
print 'Learned the network'
Yp = decoder(Gp, Hp)
print 'learner: training error %.4f' % (l1_loss(Y, Yp))
def test_mnist():
train_set, valid_set, test_set = generate_mnist(fname)
Y, target = train_set
Yt, targett = test_set
#pdb.set_trace()
#Y = np.concatenate((Y, 0*Y), axis=1)
#Yt = np.concatenate((Yt, 0*Yt), axis=1)
N, n = Y.shape
l = 4
d = int(np.ceil(n**(0.19)))
rho = estimate_rho(l, d, Y)
#N = int(np.log(n)/rho**2)
if N >= Y.shape[0]:
print 'samples N: %.d, Y: %.d' %(N, Y.shape[0])
else:
N = int(np.log(n)/rho**2)
Y = Y[0:N,:]
target = target[0:N]
Yt = Yt[0:N,:]
targett = targett[0:N]
print n, l, d, rho, N
Gp, Hp = learner(n, l, d, rho, Y)
print 'Learned the network'
Yp = decoder(Gp, Hp)
print 'NN training error: %.4f' % l1_loss(Y, Yp)
if 1:
clf = svm.LinearSVC(loss='l2', penalty='l1', dual=False)
clf.fit(Hp, target)
targetp = clf.predict(Hp)
print 'training error: %.4f' % (zero_one_loss(target, targetp))
Htp = encoder(d, Gp, Yt)
targettp = clf.predict(Htp)
print 'test error: %.4f' % (zero_one_loss(targett, targettp))
def test_leaner():
training_data, test_data = create_data()
G, H, Y, n, l, d, rho, N = training_data
Gt, Ht, Yt, nt, lt, dt, rhot, Nt = test_data
print n, l, d, rho, N
l = 1
d = 2
rho = estimate_rho(l, d, Y)
Gp, Hp = learner(n, l, d, rho, Y, G, H)
print 'Learned the network'
Yp = decoder(Gp, Hp)
print 'learner: training error %.4f' % (l1_loss(Y, Yp))
Htp = encoder(d, Gp, Yt)
Ytp = decoder(Gp, Htp)
print 'learner: test error %.4f' % (l1_loss(Yt, Ytp))
'''
def cross_val(pd_data, learner, target_class, goal, isWhat="", fold=5,
repeats=2):
"""
do 5-fold cross_validation
"""
def tune_learner(train_X):
train_len = len(train_X)
new_train_index = np.random.choice(range(train_len), train_len * 0.7)
new_tune_index = list(set(range(train_len)) - set(new_train_index))
new_train_X = train_X[new_train_index]
new_train_Y = train_Y[new_train_index]
new_tune_X = train_X[new_tune_index]
new_tune_Y = train_Y[new_tune_index]
clf = learner(new_train_X, new_train_Y, new_tune_X, new_tune_Y, goal)
tuner = DE_Tune_ML(clf, clf.get_param(), target_class, goal)
return tuner.Tune()
def tune_SMOTE(train_pd):
train_len = len(train_pd)
new_train_index = random.sample(train_pd.index, int(train_len * 0.7))
new_train = train_pd.ix[new_train_index]
if "_TunedSmote" in isWhat:
new_train_X = new_train.ix[:, new_train.columns[:-1]].values
new_train_Y = new_train.ix[:, new_train.columns[-1]].values
new_tune = train_pd.drop(new_train_index)
new_tune_X = new_tune.ix[:, new_tune.columns[:-1]].values
new_tune_Y = new_tune.ix[:, new_tune.columns[-1]].values
# clf = learner(new_train_X, new_train_Y, new_tune_X, new_tune_Y)
A_smote = smote(new_train)
num_range = [[int(A_smote.get_majority_num() * 0.5),
int(A_smote.get_majority_num() * 1.5)]] * (
A_smote.label_num - 1)
params_to_tune = {"k": [2, 20], "up_to_num": num_range}
# pdb.set_trace()
tuner = DE_Tune_SMOTE(learner, smote, params_to_tune, new_train,
new_tune, target_class, goal)
params = tuner.Tune()
return params, new_train
F = {}
total_evaluation = 0
for i in xrange(repeats): # repeat 5 times here
kf = StratifiedKFold(pd_data.ix[:, pd_data.columns[-1]].values, fold, shuffle=True)
for train_index, test_index in kf:
train_pd = pd_data.ix[train_index]
test_pd = pd_data.ix[test_index]
if "Smote" in isWhat:
k = 5
up_to_num = []
if "_TunedSmote" in isWhat:
params, train_pd = tune_SMOTE(train_pd)
# use new training data not original, because some are used as tuning
k = params["k"]
up_to_num = params["up_to_num"]
train_pd = smote(train_pd, k, up_to_num).run()
train_X = train_pd.ix[:, train_pd.columns[:-1]].values
train_Y = train_pd.ix[:, train_pd.columns[-1]].values
test_X = test_pd.ix[:, test_pd.columns[:-1]].values
test_Y = test_pd.ix[:, test_pd.columns[-1]].values
params, evaluation = tune_learner(train_X) if "_TunedLearner" in isWhat else ({},0)
F = learner(train_X, train_Y, test_X, test_Y, goal).learn(F, **params)
total_evaluation +=evaluation
# pdb.set_trace()
avg_evaluation = total_evaluation / (repeats * fold)
return avg_evaluation, F
