def prob_6():
# Repeat experiments for magic telescope and housing using weights (w = 1/dist**2)
arff = Arff('datasets/credit.arff')
arff.shuffle()
test = arff.create_subset_arff(slice(arff.instance_count // 4))
train = arff.create_subset_arff(slice(arff.instance_count // 4, None))
train.normalize()
test.normalize()
krange = np.arange(1, 16, 2)
accs = []
for k in krange:
knn = KNN(k, weighting=True, vdm=True)
predictions = knn.knn(train.get_features(), train.get_labels(),
test.get_features())
acc = predictions == np.ravel(test.get_labels().data)
print("k:", k, "accuracy:", sum(acc) / len(acc))
accs.append(sum(acc) / len(acc))
plt.plot(krange, accs)
plt.title("K Size Versus Accuracy on Credit Approval (Weighted)")
plt.xlabel("K")
plt.ylabel("Accuracy")
plt.show()
def prob5():
arff = Arff(sys.argv[2])
imp_atts = [1, 3, 4, 5, 7, 9, 11, 12, 13]
arff.shuffle()
n = len(arff.get_labels().data)
t = int(n * .55)
v = n - int(n * .20)
train_set = arff.create_subset_arff(row_idx=slice(0, t, 1),
col_idx=imp_atts)
test_set = arff.create_subset_arff(row_idx=slice(t, v, 1),
col_idx=imp_atts)
validation_set = arff.create_subset_arff(row_idx=slice(v, n, 1),
col_idx=imp_atts)
epochs = []
momentums = np.linspace(0, 1.5, 20)
# momentums = [.5, 1]
for momentum in momentums:
print(momentum)
nn = NeuralNetwork(8, [30], 11, LR=.1, momentum=momentum)
all_acc_va, all_mse_va, all_mse_te, all_mse_tr = nn.train_set(
train_set, test_set, validation_set, w=5)
epochs.append(len(all_acc_va))
plt.plot(momentums, epochs)
plt.title("Vowel Momentum vs Epoch Convergence")
plt.xlabel("Momentum")
plt.ylabel("Epochs til Conv.")
plt.show()
def prob_5():
arff = Arff('datasets/cars.arff')
arff.shuffle()
test = arff.create_subset_arff(slice(arff.instance_count//10))
training = arff.create_subset_arff(slice(arff.instance_count//10,None))
tf = test.get_features()
tl = test.get_labels()
splits = k_fold_cv(arff)
arff = arff.create_subset_arff(slice(arff.instance_count//4,None))
d = DecisionTreeLearner()
d.train(arff.get_features(), arff.get_labels())
a = d.tree
arff = Arff('datasets/voting.arff')
arff.shuffle()
arff = arff.create_subset_arff(slice(arff.instance_count//4,None))
d = DecisionTreeLearner()
d.train(arff.get_features(), arff.get_labels())
b = d.tree
return a, b
def prob_1():
arff = Arff('datasets/lenses.arff')
acc = []
for _ in range(10):
arff.shuffle()
testing = arff.create_subset_arff(slice(arff.instance_count//5))
training = arff.create_subset_arff(slice(arff.instance_count//5, None))
d = DecisionTreeLearner()
features = training.get_features()
labels = training.get_labels()
t_feat = testing.get_features()
t_labels = testing.get_labels()
d.train(features, labels)
accuracy = d.get_accuracy(t_feat, t_labels)
print(accuracy)
acc.append(accuracy)
print(sum(acc)/len(acc))
def prob_4(lr):
# read in vowels dataset
arff = Arff('datasets/vowels.arff')
# Leave out the test/train and person features, which are unceccessary.
arff = arff.create_subset_arff(col_idx=slice(2, None), label_count=1)
# Get a 75/25 split
arff.shuffle()
training = arff.create_subset_arff(slice(arff.instance_count // 4))
test = arff.create_subset_arff(slice(arff.instance_count // 4, -1))
t_features = test.get_features()
t_labels = test.get_labels()
# Get a 15% Validation set
validation = training.create_subset_arff(slice(arff.instance_count // 5))
training = training.create_subset_arff(
slice(arff.instance_count // 5, None))
v_features = validation.get_features()
v_labels = validation.get_labels()
domain = 2**np.arange(0, 8)
training_mse = []
validation_mse = []
test_mse = []
for nodes in domain:
mse = 0
vmse = 0
tmse = 0
for _ in range(100):
learner = MultilayerPerceptronLearner(
[training.features_count, nodes, 11], momentum=0)
# learner.zero_weights()
learner.lr = lr
learner.max_epoch = 500
learner.train(training.get_features(), training.get_labels(),
validation.get_features(), validation.get_labels())
tmse += learner.get_mse(test.get_features(), test.get_labels())
mse += learner.get_mse(training.get_features(),
training.get_labels())
vmse += learner.get_mse(validation.get_features(),
validation.get_labels())
training_mse.append(mse / 100)
validation_mse.append(vmse / 100)
test_mse.append(tmse / 100)
plt.semilogx(domain, test_mse, basex=2, label="Test Set MSE")
plt.semilogx(domain, training_mse, basex=2, label="Training Set MSE")
plt.semilogx(domain, validation_mse, basex=2, label="Validation Set MSE")
plt.title("MSE vs Number of Hidden Nodes")
plt.xlabel("Number of Hidden Nodes")
plt.ylabel("Mean Squared Error")
plt.legend()
plt.show()
def prob_6_b():
# read in vowels dataset
arff = Arff('datasets/vowels.arff')
# Leave out the test/train and person features, which are unceccessary.
arff = arff.create_subset_arff(col_idx=slice(2, None), label_count=1)
# Get a 75/25 split
arff.shuffle()
training = arff.create_subset_arff(slice(arff.instance_count // 4))
test = arff.create_subset_arff(slice(arff.instance_count // 4, -1))
t_features = test.get_features()
t_labels = test.get_labels()
# Get a 15% Validation set
validation = training.create_subset_arff(slice(arff.instance_count // 5))
training = training.create_subset_arff(
slice(arff.instance_count // 5, None))
v_features = validation.get_features()
v_labels = validation.get_labels()
features = training.get_features()
labels = training.get_labels()
taccuracy = []
accuracy = []
domain = 2**np.arange(1, 7)
for i in domain:
tacc = 0
acc = 0
for _ in range(3):
learner = MultilayerPerceptronLearner([training.features_count] +
[i] * (32 // i) + [11],
momentum=.85)
# learner.zero_weights()
learner.lr = .1
learner.max_epoch = 500
learner.train(training.get_features(), training.get_labels(),
validation.get_features(), validation.get_labels())
tacc += learner.get_accuracy(t_features, t_labels)
acc += learner.get_accuracy(features, labels)
accuracy.append(acc / 3)
taccuracy.append(tacc / 3)
plt.semilogx(domain, accuracy, basex=2, label="Training Set Accuracy")
plt.semilogx(domain, taccuracy, basex=2, label="Test Set Accuracy")
plt.title("Node Distribution vs Accuracy")
plt.xlabel("Number of Nodes per Hidden Layer")
plt.ylabel("Accuracy")
plt.legend()
plt.show()
def prob0haccomplete():
arff = Arff('datasets/labor.arff', label_count=1)
# Trim the id column
arff = arff.create_subset_arff(col_idx=slice(1, None))
arff = arff.get_features()
hac = HAC(simple=False)
hac.train(arff, verbose=True, printk=[5])
def prob0():
arff = Arff('datasets/labor.arff', label_count=1)
# Trim the id column
arff = arff.create_subset_arff(col_idx=slice(1, None))
arff = arff.get_features()
km = KMeans(5)
km.train(arff, verbose=True, centers=arff.data[:5])
def prob3():
""" """
arff = Arff(sys.argv[2])
imp_atts = [1, 3, 4, 5, 7, 9, 11, 12, 13]
arff.shuffle()
n = len(arff.get_labels().data)
t = int(n * .55)
v = n - int(n * .20)
train_set = arff.create_subset_arff(row_idx=slice(0, t, 1),
col_idx=imp_atts)
test_set = arff.create_subset_arff(row_idx=slice(t, v, 1),
col_idx=imp_atts)
validation_set = arff.create_subset_arff(row_idx=slice(v, n, 1),
col_idx=imp_atts)
best_mse_te = []
best_mse_tr = []
best_mse_va = []
epochs = []
LRS = [.01, .1, .5, .8, 1.5]
for LR in LRS:
# print(LR)
nn = NeuralNetwork(8, [16], 11, LR=LR, momentum=0)
all_acc_va, all_mse_va, all_mse_te, all_mse_tr = nn.train_set(
train_set, test_set, validation_set, w=5)
best_mse_te.append(min(all_mse_te))
best_mse_tr.append(min(all_mse_tr))
best_mse_va.append(min(all_mse_va))
epochs.append(len(all_mse_va))
plt.plot(LRS, best_mse_te, label="MSE Te")
plt.plot(LRS, best_mse_tr, label="MSE Tr")
plt.plot(LRS, best_mse_va, label="MSE V.A")
plt.title("Vowel MSE vs Learning Rate")
plt.xlabel("Learning Rate")
plt.ylabel("MSE")
plt.legend()
plt.show()
plt.plot(LRS, epochs)
plt.title("Vowel Epochs vs Learning Rate")
plt.xlabel("Learning Rate")
plt.ylabel("Epochs")
plt.legend()
plt.show()
def prob_5(lr, hidden):
# read in vowels dataset
arff = Arff('datasets/vowels.arff')
# Leave out the test/train and person features, which are unceccessary.
arff = arff.create_subset_arff(col_idx=slice(2, None), label_count=1)
# Get a 75/25 split
arff.shuffle()
training = arff.create_subset_arff(slice(arff.instance_count // 4))
test = arff.create_subset_arff(slice(arff.instance_count // 4, -1))
t_features = test.get_features()
t_labels = test.get_labels()
# Get a 15% Validation set
validation = training.create_subset_arff(slice(arff.instance_count // 5))
training = training.create_subset_arff(
slice(arff.instance_count // 5, None))
v_features = validation.get_features()
v_labels = validation.get_labels()
epochs = []
accuracy = []
domain = np.linspace(0, 1, 20)
for momentum in domain:
e = 0
acc = 0
for _ in range(10):
learner = MultilayerPerceptronLearner(
[training.features_count, hidden, 11], momentum=momentum)
# learner.zero_weights()
learner.lr = lr
learner.max_epoch = 500
learner.train(training.get_features(), training.get_labels(),
validation.get_features(), validation.get_labels())
acc += learner.get_accuracy(t_features, t_labels)
e += learner.epochs
epochs.append(e / 10)
accuracy.append(acc / 10)
print(accuracy)
plt.plot(domain, epochs)
plt.title("Number of Training Epochs vs Momentum")
plt.xlabel("Momentum Constant")
plt.ylabel("Number of Training Epochs")
plt.show()
def setup():
arff = Arff('datasets/labor.arff', label_count=1)
# Trim the id column
arff = arff.create_subset_arff(col_idx=slice(1, None))
arff = arff.get_features()
hac = HAC()
hac.nominal_indicies = np.where(np.array(arff.attr_types) == 'nominal')[0]
print('33,44', hac.get_distance(arff.data[33], arff.data[44]))
print('25,34', hac.get_distance(arff.data[25], arff.data[34]))
def prob_5():
cont_mask = [1, 2, 7, 10, 13, 14, 16]
cate_mask = [0, 3, 4, 5, 6, 8, 9, 11, 12, 15]
arff = Arff("credit_approval_data.arff")
arff.shuffle()
arff.normalize()
n = len(arff.get_labels().data)
t = int(n * .7)
train_data = arff.create_subset_arff(row_idx=slice(0, t, 1))
test_data = arff.create_subset_arff(row_idx=slice(t, n, 1))
test_data = np.hstack((test_data.get_features().data, test_data.get_labels().data))
train_data = np.hstack((train_data.get_features().data, train_data.get_labels().data))
#b,30.83,0,u,g,w,v,1.25,t,t,01,f,g,00202,0,+
dist_matrix = np.ones((16, 16))
np.fill_diagonal(dist_matrix, 0)
KNNC = KNNClassifier(8, train_data, test_data)
print(KNNC.get_accuracy_mixed(cate_mask, cont_mask, dist_matrix))
def prob_2():
# Get arff from iris database
iris = Arff('datasets/iris.arff')
# Get a random 75/25 split
iris.shuffle()
validation = iris.create_subset_arff(slice(iris.instance_count // 4))
v_features = validation.get_features()
v_labels = validation.get_labels()
training = iris.create_subset_arff(slice(iris.instance_count // 4, None))
features = training.get_features()
labels = training.get_labels()
# Make the learner
# shape = [4 8 3]
learner = MultilayerPerceptronLearner([4, 8, 3], momentum=0)
training_mse = []
validation_mse = []
classification_accuracy = []
for epoch in range(learner.max_epoch):
learner.train_one_epoch(features, labels)
training_mse.append(learner.get_mse(features, labels))
validation_mse.append(learner.get_mse(v_features, v_labels))
classification_accuracy.append(
learner.get_accuracy(v_features, v_labels))
if learner.check_for_convergence(v_features, v_labels):
break
fig, (ax1, ax2) = plt.subplots(nrows=2, sharex=True)
ax1.plot(training_mse, label="Training Data MSE")
ax1.plot(validation_mse, label="Validation Data MSE")
ax1.legend()
ax2.plot(classification_accuracy, label="Classification Accuracy")
plt.legend()
plt.show()
def prob2():
arff = Arff(sys.argv[1])
arff.shuffle()
n = len(arff.get_labels().data)
t = int(n * .55)
v = n - int(n * .20)
train_set = arff.create_subset_arff(row_idx=slice(0, t, 1))
test_set = arff.create_subset_arff(row_idx=slice(t, v, 1))
validation_set = arff.create_subset_arff(row_idx=slice(v, n, 1))
nn = NeuralNetwork(4, [9], 3, LR=.1)
all_acc_va, all_mse_va, all_mse_te, all_mse_tr = nn.train_set(
train_set, test_set, validation_set)
d = [x for x in range(len(all_acc_va))]
plt.plot(d, all_mse_te, label="test MSE")
plt.plot(d, all_mse_va, label="Val. MSE")
plt.plot(d, all_acc_va, label="Val. Accuracy")
plt.title("Iris Dataset")
plt.xlabel("Epochs")
plt.ylabel("%")
plt.legend()
plt.show()
def prob4():
arff = Arff(sys.argv[2])
imp_atts = [1, 3, 4, 5, 7, 9, 11, 12, 13]
arff.shuffle()
n = len(arff.get_labels().data)
t = int(n * .55)
v = n - int(n * .20)
train_set = arff.create_subset_arff(row_idx=slice(0, t, 1),
col_idx=imp_atts)
test_set = arff.create_subset_arff(row_idx=slice(t, v, 1),
col_idx=imp_atts)
validation_set = arff.create_subset_arff(row_idx=slice(v, n, 1),
col_idx=imp_atts)
best_mse_te = []
best_mse_tr = []
best_mse_va = []
hidden_nodes = [1, 3, 6, 10, 13, 15, 16, 18, 20, 22, 25, 30, 40]
for nodes in hidden_nodes:
# print(nodes)
nn = NeuralNetwork(8, [nodes], 11, LR=.1, momentum=0)
all_acc_va, all_mse_va, all_mse_te, all_mse_tr = nn.train_set(
train_set, test_set, validation_set, w=5)
best_mse_te.append(min(all_mse_te))
best_mse_tr.append(min(all_mse_tr))
best_mse_va.append(min(all_mse_va))
plt.plot(hidden_nodes, best_mse_te, label="MSE Te")
plt.plot(hidden_nodes, best_mse_tr, label="MSE Tr")
plt.plot(hidden_nodes, best_mse_va, label="MSE V.A")
plt.title("Vowel MSE vs Hidden Nodes")
plt.xlabel("Hidden Nodes")
plt.ylabel("MSE")
plt.legend()
plt.show()
def prob_3():
# read in vowels dataset
arff = Arff('datasets/vowels.arff')
# Leave out the test/train and person features, which are unceccessary.
arff = arff.create_subset_arff(col_idx=slice(2, None), label_count=1)
# Get a 75/25 split
arff.shuffle()
training = arff.create_subset_arff(slice(arff.instance_count // 4))
test = arff.create_subset_arff(slice(arff.instance_count // 4, -1))
t_features = test.get_features()
t_labels = test.get_labels()
# Get a 15% Validation set
validation = training.create_subset_arff(slice(arff.instance_count // 5))
training = training.create_subset_arff(
slice(arff.instance_count // 5, None))
v_features = validation.get_features()
v_labels = validation.get_labels()
training_mse = []
validation_mse = []
test_mse = []
epochs = []
domain = np.logspace(-3, 0)
for lr in domain:
mse = 0
vmse = 0
tmse = 0
e = 0
for _ in range(3):
learner = MultilayerPerceptronLearner(
[training.features_count, 2 * training.features_count, 11],
momentum=0)
learner.zero_weights()
learner.lr = lr
learner.max_epoch = 500
learner.train(training.get_features(), training.get_labels(),
validation.get_features(), validation.get_labels())
e += learner.epochs
tmse += learner.get_mse(test.get_features(), test.get_labels())
mse += learner.get_mse(training.get_features(),
training.get_labels())
vmse += learner.get_mse(validation.get_features(),
validation.get_labels())
epochs.append(e / 3)
training_mse.append(mse / 3)
validation_mse.append(vmse / 3)
test_mse.append(tmse / 3)
plt.semilogx(domain, test_mse, label="Test Set MSE")
plt.semilogx(domain, training_mse, label="Training Set MSE")
plt.semilogx(domain, validation_mse, label="Validation Set MSE")
plt.title("MSE vs Learning Rate")
plt.xlabel("Learning Rate")
plt.ylabel("Mean Squared Error")
plt.legend()
plt.show()
plt.semilogx(domain, epochs)
plt.title("Number of Training Epochs vs Learning Rate")
plt.xlabel("Learning Rate")
plt.ylabel("Number of Training Epochs")
plt.show()