def make_forest(data, n_bootstraps, scoref, min_gain=0.01): """Function to grow a random forest given some training data.""" trees = [] for _ in xrange(n_bootstraps): data_boot = make_boot(data, data.shape[0]) trees.append(dt.buildtree(data_boot, scoref, min_gain)) return Forest(trees)
start = stepSize * i
training = dataset[start:start + stepSize]
trainingLabels = labelInt[start:start + stepSize]
############# Feature Extraction ##############
my_model = PCA(n_components=pca_comps, svd_solver='full')
newSet = my_model.fit_transform(training).tolist()
newTestSet = my_model.transform(test).tolist()
newTrainSet = my_model.transform(training).tolist()
############# Model Building ##############
for k in range(len(newSet)):
newSet[k].append(trainingLabels[k])
passingData = newSet[:]
models.append(dt.buildtree(passingData))
# dt.prune(b,0.1)
############# Classification of Test Records ##############
for j in range(len(newTestSet)):
if j not in test_classify:
test_classify[j] = []
test_classify[j].append(dt.classify(newTestSet[j], models[i]))
############# Accuracy Calculations ##############
d = []
f = []
flat = []
for l in test_classify.values():
tot_count = 0
tot_correct = 0
train_data = parse_file(train_file_path)
test_data = parse_file(test_file_path)
#Calculating the Accuracy at every level
correct = 0
total = 0
TP = 0
TN = 0
FP = 0
FN = 0
depth = 0
for i in range(1, 7):
tree = dtree.buildtree(train_data, 0, i)
for data in test_data:
predicted = list(dtree.decision(tree, data).keys())[0]
actual = data[-1]
total = total + 1
if predicted == 1.0 and actual == 1.0:
correct = correct + 1
TP = TP + 1
if predicted == 0.0 and actual == 0.0:
correct = correct + 1
TN = TN + 1
if predicted == 1.0 and actual == 0.0:
FP = FP + 1
if predicted == 0.0 and actual == 1.0:
FN = FN + 1
tot_correct += correct
def build_BNN(data,
output_condition,
cd=98,
mss=1,
md=10,
relevant_neuron_dictionary={},
with_data=0,
discretization=0,
cluster_means=None):
'''
Starting from the target condition and until the conditions with respect
to the first hidden layer, it extracts a DNF that explains each condition
using conditions of the next shallower layer
param data: instance of DataSet
param output_condition: condition of interest
param cd: class dominance
param mss: minimum dataset size
param md: maximum tree depth
param with_data: Avoid ==0. If == 1, the regular simplification operations are performed, if == 2, post-ppruning is performed
param discretization: method used to determine the thresholds that split the activation range of each neuron
'''
BNN = {}
deep_layer = data.output_layer
target_class = [output_condition]
while deep_layer > 0:
target_split_values = set((l, n, t) for (l, n, t, u) in target_class)
if not target_split_values:
warnings.warn(
'Warning: no split points, returning current dictionary at layer: '
+ str(deep_layer))
print('Target split values', target_split_values)
used_shallow_conditions = set([])
current_data = temp_data(data, deep_layer - 1, target_class)
if discretization == 0:
split_points = dis.all_features_trivial_mid_points(current_data)
elif discretization == 1 or discretization == 3:
split_points = dis.one_time_discretization(
current_data,
discretization,
rnd=relevant_neuron_dictionary,
tsv=list(target_split_values))
elif discretization == 2 or discretization == 4:
split_points = cluster_means[deep_layer - 1]
elif discretization == 6:
colum = [[d[c] for d in current_data]
for c in range(len(current_data[0]) - 1)]
split_points = [[sum(vq.kmeans(v, 2)[0]) / 2] for v in colum]
elif discretization == 5:
if deep_layer == 1:
split_points = [[0.5] for l in range(len(current_data[0]) - 1)]
else:
split_points = [[0] for l in range(len(current_data[0]) - 1)]
print('Split points', [len(l) for l in split_points])
print(split_points)
print('')
for i in target_split_values:
print('')
print('i: ', i)
t = time.time()
i_data = temp_data(data, deep_layer - 1, i)
tree = None
if relevant_neuron_dictionary and discretization == 0:
pruned_split_points = [
_sp(j, i, split_points, relevant_neuron_dictionary)
for j in range(len(split_points))
]
print(pruned_split_points)
tree = dt.buildtree(i_data,
pruned_split_points,
class_dominance=cd,
min_set_size=mss,
max_depth=md,
root=True)
else:
tree = dt.buildtree(i_data,
split_points,
class_dominance=cd,
min_set_size=mss,
max_depth=md,
root=True)
if not tree:
cero_class = sum(1 for x in i_data if x[-1] == 0)
one_class = sum(1 for x in i_data if x[-1] == 1)
if cero_class > one_class:
BNN[(i[0], i[1], i[2], True)] = False
BNN[(i[0], i[1], i[2], False)] = True
else:
BNN[(i[0], i[1], i[2], False)] = True
BNN[(i[0], i[1], i[2], True)] = False
break
print('Tree is formed')
print('Time: ', time.time() - t)
dnfs = dt.get_dnfs(deep_layer - 1, tree)
if (i[0], i[1], i[2], False) in target_class:
print('False case')
pruned = None
if isinstance(dnfs[0], list):
# print('Fidelity pre-pruning:', ef.accuracy_of_dnf(data, (i[0], i[1], i[2], False), dnfs[0], True, False, False, True))
# print('Precision pre-pruning:', ef.precision_of_dnf(data, (i[0], i[1], i[2], False), dnfs[0], True, False, False, True))
# print('Recall pre-pruning:', ef.recall_of_dnf(data, (i[0], i[1], i[2], False), dnfs[0], True, False, False, True))
data.update_dictionary([(l, n, t) for conj in dnfs[0]
for (l, n, t, u) in conj])
if with_data == 0:
pruned = s.boolean_simplify_basic(dnfs[0])
elif with_data >= 1:
pruned = s.boolean_simplify_complex(dnfs[0])
if with_data == 2:
pruned = p.post_prune(pruned,
(i[0], i[1], i[2], False),
data.example_cond_dict,
data.dict_indexes,
data=None)
used_shallow_conditions.update(
set(c for conj in pruned for c in conj))
else:
pruned = dnfs[0]
# print('Fidelity post-pruning:', ef.accuracy_of_dnf(data, (i[0], i[1], i[2], False), pruned, True, False, False, True))
# print('Precision post-pruning:', ef.precision_of_dnf(data, (i[0], i[1], i[2], False), pruned, True, False, False, True))
# print('Recall post-pruning:', ef.recall_of_dnf(data, (i[0], i[1], i[2], False), pruned, True, False, False, True))
BNN[(i[0], i[1], i[2], False)] = pruned
print((i[0], i[1], i[2], False), pruned)
if (i[0], i[1], i[2], True) in target_class:
print('True case')
pruned = None
if isinstance(dnfs[1], list):
# print('Fidelity pre-pruning:', ef.accuracy_of_dnf(data, (i[0], i[1], i[2], True), dnfs[1], True, False, False, True))
# print('Precision pre-pruning:', ef.precision_of_dnf(data, (i[0], i[1], i[2], True), dnfs[1], True, False, False, True))
# print('Recall pre-pruning:', ef.recall_of_dnf(data, (i[0], i[1], i[2], True), dnfs[1], True, False, False, True))
data.update_dictionary([(l, n, t) for conj in dnfs[1]
for (l, n, t, u) in conj])
if with_data == 0:
pruned = s.boolean_simplify_basic(dnfs[1])
elif with_data >= 1:
pruned = s.boolean_simplify_complex(dnfs[1])
if with_data == 2:
pruned = p.post_prune(pruned, (i[0], i[1], i[2], True),
data.example_cond_dict,
data.dict_indexes,
data=None)
used_shallow_conditions.update(
set(c for conj in pruned for c in conj))
else:
pruned = dnfs[1]
# print('Fidelity post-pruning:', ef.accuracy_of_dnf(data, (i[0], i[1], i[2], True), pruned, True, False, False, True))
# print('Precision post-pruning:', ef.precision_of_dnf(data, (i[0], i[1], i[2], True), pruned, True, False, False, True))
# print('Recall post-pruning:', ef.recall_of_dnf(data, (i[0], i[1], i[2], True), pruned, True, False, False, True))
BNN[(i[0], i[1], i[2], True)] = pruned
print((i[0], i[1], i[2], True), pruned)
deep_layer -= 1
target_class = list(used_shallow_conditions)
return BNN
trainingLabels = [labelInt[i] for i in training_idx]
testLabels = [labelInt[i] for i in test_idx]
############# Feature Extraction ##############
my_model = PCA(n_components=pca_comps, svd_solver='full')
newSet = my_model.fit_transform(training).tolist()
newTestSet = my_model.transform(test).tolist()
newTrainSet = my_model.transform(training).tolist()
############# Model Building ##############
for i in range(len(newSet)):
newSet[i].append(trainingLabels[i])
passingData = newSet[:]
b = dt.buildtree(passingData)
dt.prune(b, 0.1)
############# Classification of Train Records ##############
count = 0
for i in range(len(newTrainSet)):
a = dt.classify(newTrainSet[i], b)
for key in a.keys():
if (key == trainingLabels[i]):
count = count + 1
############# Accuracy Calculations for Training DataSet ##############
accuracy = (count / len(newTrainSet)) * 100
final_train_acc += accuracy
print('Train accuracy:', accuracy)
# remove index column
features_train = features_train[:,1:]
labels_train = np.genfromtxt('../census-dataset/census-train-labels.csv', delimiter=' ', skip_header=1)
# remove index column
labels_train = labels_train[:,1:][:,0]
# split to obtain train and test set
x_train, x_test, y_train, y_test = train_test_split(features_train, labels_train, test_size=0.33)
# concatenate features and labels
data_train = np.column_stack((x_train, y_train))
data_test = np.column_stack((x_test, y_test))
# build decision tree using entropy
decision_tree = dt.buildtree(data_train, dt.entropy, 0.01)
min_gain_error = {}
# test minimal gain values for pruning
for min_gain_value in np.arange(0,1, 0.01):
dt_temp = copy.copy(decision_tree)
dt.prune(dt_temp, min_gain_value)
# classify test data
y_hat = map(lambda obs : dt.classify(obs, dt_temp), x_test)
y_hat = map(dt.convertToLabel, y_hat)
y_hat = np.array(y_hat)
error = (y_hat != y_test).sum() / float(y_test.shape[0])
min_gain_error[min_gain_value] = error
# prune tree with optimal min_gain value
label, pixels = dataset[test_idx[i]]
record = (pixels.flatten()).tolist()
testing_labels.append(label)
rows_test_total.append(record)
############# Feature Extraction ##############
FinalTrain = []
my_model = PCA(n_components=pca_comps, svd_solver='full')
newSet = my_model.fit_transform(rows_total).tolist()
newtestSet = my_model.transform(rows_test_total).tolist()
############# Model Building ##############
for i in range(len(rows_total)):
newSet[i].append(training_labels[i])
b = dt.buildtree(newSet)
dt.prune(b, 0.1)
############# Classification of Test Records ##############
number = 0
accuracy = 0
for i in range(testSize):
a = dt.classify(newtestSet[i], b)
for key in a.keys():
if (key == testing_labels[i]):
number = number + 1
############# Accuracy Calculations ##############
accuracy = (number / testSize) * 100
final_test_acc += accuracy
def buildtree(depth, test_data, train_data, current_index):
tot_count = 0
tot_correct = 0
#Calculating the Accuracy at every level
correct = 0
total = 0
TP = 0
TN = 0
FP = 0
FN = 0
#print ("Depth Entered is :" ,depth)
predicted_list = []
predicted_list_1 = []
for i in range(depth):
tree = dtree.buildtree(train_data, 0, i)
for data in train_data:
predicted = list(dtree.decision(tree, data).keys())[-1]
predicted_list.append(predicted)
for data in test_data:
predicted = list(dtree.decision(tree, data).keys())[0]
predicted_list_1.append(predicted)
one_count_testdata = predicted_list_1.count(1)
zero_count_testdata = predicted_list_1.count(0)
actual = data[-1]
total = total + 1
if predicted == 1.0 and actual == 1.0:
correct = correct + 1
TP = TP + 1
if predicted == 0.0 and actual == 0.0:
correct = correct + 1
TN = TN + 1
if predicted == 1.0 and actual == 0.0:
FP = FP + 1
if predicted == 0.0 and actual == 1.0:
FN = FN + 1
tot_correct += correct
tot_count += total
Accuracy = round(100 * correct / total, 2)
Depth_list.append(depth)
Accuracy_list.append(Accuracy)
depth = depth + 1
#print (Accuracy_list)
#print (Depth_list)
#printing the confusion matrix
print("Accuracy::", str(Accuracy) + '%')
print("False Negatives ", str(FN))
print("False positives ", str(FP))
print("True Negatives ", str(TN))
print("True Positives ", str(TP))
print("Confusion Matrix for bagging")
print("------")
print("| ", TP, "|", FN, "|")
print("------")
print("| ", FP, "|", TN, "|")
print("------")
def setUp(self):
my_data = np.genfromtxt('decision_tree_example.txt', dtype=None)
self.rows = my_data.tolist()
self.tree = decision_tree.buildtree(self.rows)
