def SrcOnly(target, model='ML'):
# three domains
domains = ['MALE', 'FEMALE', 'MIXED']
if target not in domains:
print("Wrong target !")
return
print("Model:SrcOnly -- Domain:{}".format(target))
src = []
for i in domains:
if i != target:
src.append(i)
_, dev, _ = pp.get_splited_data("{}.csv".format(target))
train = []
for s in src:
t, _, _ = pp.get_splited_data("{}.csv".format(s))
train += t
random.shuffle(train)
if model == 'NN':
NN(train, dev, 100)
elif model == 'ML':
BayesianRidge(train, dev)
def FSMM(target):
# three domains
domains = ['MALE', 'FEMALE', 'MIXED']
if target not in domains:
print("Wrong target !")
return
print("Model:FSMM -- Domain:{}".format(target))
src = []
for i in domains:
if i != target:
src.append(i)
tra, dev, _ = pp.get_splited_data("{}.csv".format(target))
# 100 subsamples from target domain
sub_tra = random.sample(tra, 100)
train = []
for s in src:
t, _, _ = pp.get_splited_data("{}.csv".format(s))
train += t
random.shuffle(tra)
random.shuffle(train)
# train the model on source samples
pred_mol = predict_model(train)
# train the model to learn mapping
map_mol = mapping_model(tra, train)
dev = np.array(dev)
X_dev, Y_dev = dev[:, :-1], dev[:, -1]
# mapping target samples to source samples
mapping_res = map_mol.predict(X_dev)
# prediction process
pred = pred_mol.predict(mapping_res)
# evaluation
zero, two, five, ten = 0, 0, 0, 0
for i in range(len(Y_dev)):
if abs(Y_dev[i] - pred[i]) <= 0.5:
zero += 1
elif abs(Y_dev[i] - pred[i]) <= 2:
two += 1
elif abs(Y_dev[i] - pred[i]) <= 5:
five += 1
elif abs(Y_dev[i] - pred[i]) <= 10:
ten += 1
print("Using FSMM:")
print(
"[Mark difference : Sample number] pairs between prediction and true mark:"
)
print("Zero:{} Two:{} Five:{} Ten:{}".format(zero, two, five, ten))
def Pred(target, model='ML'):
# three domains
domains = ['MALE', 'FEMALE', 'MIXED']
if target not in domains:
print("Wrong target !")
return
print("Model:Pred -- Domain:{}".format(target))
src = []
for i in domains:
if i != target:
src.append(i)
tra, dev, _ = pp.get_splited_data("{}.csv".format(target))
target = tra + dev
# trained on source domain
train = []
for s in src:
t, _, _ = pp.get_splited_data("{}.csv".format(s))
train += t
random.shuffle(train)
if model == 'NN':
pred = NN(train, target, 100, False)
# add predicted feature
for i in range(len(target)):
target[i] = [round(pred[i][0], 2)] + target[i]
# split updated target domain samples (last 100 samples are dev samples)
new_tra, new_dev = target[:-100], target[-100:]
random.shuffle(new_tra)
# trained on new target domain
NN(new_tra, new_dev, 100)
elif model == 'ML':
pred = BayesianRidge(train, target, False)
print(pred)
# add predicted feature
for i in range(len(target)):
target[i] = [round(pred[i], 2)] + target[i]
# split updated target domain samples (last 100 samples are dev samples)
new_tra, new_dev = target[:-100], target[-100:]
random.shuffle(new_tra)
# trained on new target domain
BayesianRidge(new_tra, new_dev)
def Weighted(target, model='ML'):
# three domains
domains = ['MALE', 'FEMALE', 'MIXED']
if target not in domains:
print("Wrong target !")
return
print("Model:Weighted -- Domain:{}".format(target))
src = []
for i in domains:
if i != target:
src.append(i)
tra, dev, _ = pp.get_splited_data("{}.csv".format(target))
train = []
for s in src:
t, _, _ = pp.get_splited_data("{}.csv".format(s))
train += t
# weight proportion constant
# Actually, w = 1 / (len(train) / 100)
# where len(train) is the number of source samples and 100 is the number of target sub-samples
w_constant = 100 / len(train)
# weighted process
for i in range(len(train)):
# Don't change the label value
for j in range(len(train[0]) - 1):
train[i][j] = round(w_constant * train[i][j], 2)
# add target samples without weighting
train += random.sample(tra, 100)
random.shuffle(train)
if model == 'NN':
NN(train, dev, 100)
elif model == 'ML':
BayesianRidge(train, dev)
def FEDA(target, model='ML'):
# three domains
domains = ['MALE', 'FEMALE', 'MIXED']
if target not in domains:
print("Wrong target !")
return
print("Model:FEDA -- Domain:{}".format(target))
src = []
for i in domains:
if i != target:
src.append(i)
tra, dev, _ = pp.get_splited_data("{}.csv".format(target))
# 100 subsamples from target domain
sub_tra = random.sample(tra, 100)
# feature augmentation
sub_tra = feature_aug(sub_tra, target)
dev = feature_aug(dev, target)
train = []
for s in src:
t, _, _ = pp.get_splited_data("{}.csv".format(s))
t = feature_aug(t, s)
train += t
train += sub_tra
random.shuffle(train)
if model == 'NN':
NN(train, dev, 100)
elif model == 'ML':
BayesianRidge(train, dev)
def TgtOnly(target, model='ML'):
# three domains
domains = ['MALE', 'FEMALE', 'MIXED']
if target not in domains:
print("Wrong target !")
return
print("Model:TgtOnly -- Domain:{}".format(target))
train, dev, _ = pp.get_splited_data("{}.csv".format(target))
random.shuffle(train)
if model == 'NN':
NN(train, dev, 100)
elif model == 'ML':
BayesianRidge(train, dev)
def LinInt(target, model='ML'):
# three domains
domains = ['MALE', 'FEMALE', 'MIXED']
if target not in domains:
print("Wrong target !")
return
print("Model:LinInt -- Domain:{}".format(target))
src = []
for i in domains:
if i != target:
src.append(i)
tra, dev, _ = pp.get_splited_data("{}.csv".format(target))
train = []
for s in src:
t, _, _ = pp.get_splited_data("{}.csv".format(s))
train += t
random.shuffle(train)
if model == 'NN':
# SrcOnly pred
pred1 = NN(train, dev, 100, False)
# TgtOnly pred
pred2 = NN(tra, dev, 100, False)
two_pred = np.hstack((pred1, pred2))
#print(pred1, pred2)
#print(two_pred, two_pred.shape)
# final pred = w1*pred1 + w2*pred2 + bias
# train a new model to get W and b
model = tf.keras.Sequential()
model.add(tf.keras.layers.Dense(1))
model.compile(optimizer='adam', loss='mse', metrics=['accuracy'])
dev_label = np.array(dev)[:, -1]
# print(dev_label,dev_label.shape)
for _ in range(500):
model.fit(two_pred, dev_label, verbose=0)
W, b = model.layers[0].get_weights()
# use W and b to interpolate SrcOnly and TgtOnly predictions
new_pred1 = NN(train, dev, 100, False)
new_pred2 = NN(tra, dev, 100, False)
new_two_pred = np.hstack((new_pred1, new_pred2))
# print(W, b)
final_pred = np.dot(new_two_pred, W) + b
zero, two, five, ten = 0, 0, 0, 0
for i in range(len(final_pred)):
if abs(dev_label[i] - final_pred[i]) <= 0.5:
zero += 1
elif abs(dev_label[i] - final_pred[i]) <= 2:
two += 1
elif abs(dev_label[i] - final_pred[i]) <= 5:
five += 1
elif abs(dev_label[i] - final_pred[i]) <= 10:
ten += 1
print("Using Nerual Network:")
print(
"[Mark difference : Sample number] pairs between prediction and true mark:"
)
print("Zero:{} Two:{} Five:{} Ten:{}".format(zero, two, five, ten))
if model == 'ML':
# SrcOnly pred
pred1 = BayesianRidge(train, dev, False)
# TgtOnly pred
pred2 = BayesianRidge(tra, dev, False)
pred1 = pred1.reshape(100, 1)
pred2 = pred2.reshape(100, 1)
two_pred = np.hstack((pred1, pred2))
dev_label = np.array(dev)[:, -1]
clf = linear_model.BayesianRidge()
# clf to learn to combine the pred of SrcOny and TgtOnly
clf.fit(two_pred, dev_label)
new_pred1 = BayesianRidge(train, dev, False)
new_pred2 = BayesianRidge(tra, dev, False)
new_pred1 = new_pred1.reshape(100, 1)
new_pred2 = new_pred2.reshape(100, 1)
new_two_pred = np.hstack((new_pred1, new_pred2))
final_pred = clf.predict(new_two_pred)
zero, two, five, ten = 0, 0, 0, 0
for i in range(len(final_pred)):
if abs(dev_label[i] - final_pred[i]) <= 0.5:
zero += 1
elif abs(dev_label[i] - final_pred[i]) <= 2:
two += 1
elif abs(dev_label[i] - final_pred[i]) <= 5:
five += 1
elif abs(dev_label[i] - final_pred[i]) <= 10:
ten += 1
print("Using Machine Learning Method -- BayesianRidge:")
print(
"[Mark difference : Sample number] pairs between prediction and true mark:"
)
print("Zero:{} Two:{} Five:{} Ten:{}".format(zero, two, five, ten))