def get_best_tree(currtree):
found_better_tree = False
for newtree in dtree.allPruned(currtree):
if dtree.check(newtree, monkval) > dtree.check(currtree, monkval):
found_better_tree = True
currtree = newtree
if found_better_tree:
currtree = get_best_tree(currtree)
return currtree
def pick_best_tree(old_tree, validation_set):
list_of_trees = dec.allPruned(old_tree)
old_accuracy = dec.check(old_tree, validation_set)
validation_accuracy = 0
result = None
for i in range(len(list_of_trees)):
temp = list_of_trees[i]
temp_accuracy = dec.check(temp, validation_set)
if temp_accuracy > validation_accuracy:
validation_accuracy = temp_accuracy
result = temp
if validation_accuracy < old_accuracy:
return old_tree
else:
pick_best_tree(result, validation_set)
return result
def optimum_prune(tree, val_data):
def get_local_opt(tp):
opt = (None, 0)
for tree, perf in tp:
opt = (tree, perf) if perf >= opt[1] else opt
return opt
optimum = (None, 0)
while True:
pruned_trees = allPruned(tree)
performance = [(t, check(t, val_data)) for t in pruned_trees]
local_opt = get_local_opt(performance)
#print('Current optimum: {}, new optimum: {}'.format(optimum[1], local_opt[1]))
if local_opt[1] > optimum[1]:
optimum = local_opt
else:
break
return optimum
def assignment_5():
print("*** ASSIGNMENT 5 ***")
t_monk1 = buildTree(monk1, attributes)
t_monk2 = buildTree(monk2, attributes)
t_monk3 = buildTree(monk3, attributes)
result_text = "{} -- E_train: {}; E_test: {}"
print(
result_text.format('MONK1', 1.0 - check(t_monk1, monk1),
1.0 - check(t_monk1, monk1test)))
print(
result_text.format('MONK2', 1.0 - check(t_monk2, monk2),
1.0 - check(t_monk2, monk2test)))
print(
result_text.format('MONK3', 1.0 - check(t_monk3, monk3),
1.0 - check(t_monk3, monk3test)))
print("\n")
# print(s)
# Onwards
# tree = dtree.buildTree(monkdata.monk1, monkdata.attributes, maxdepth=1) # Two levels
# tree = dtree.buildTree(monkdata.monk1, monkdata.attributes) # All levels
# drawtree.drawTree(tree) # Show tree
# PERFORMANCE CHECK PART
name_sets = ('MONK-1', 'MONK-2', 'MONK-3')
training_sets = (monkdata.monk1, monkdata.monk2, monkdata.monk3)
test_sets = (monkdata.monk1test, monkdata.monk2test, monkdata.monk3test)
trees = list(
dtree.buildTree(training_set, monkdata.attributes)
for training_set in training_sets)
print('# Performance Check')
header = ['Dataset', 'Train', 'Test']
data = []
for tree, name_set, training_set, test_set in zip(trees, name_sets,
training_sets,
test_sets):
data.append([
name_set, 1 - round(dtree.check(tree, training_set), 5),
1 - round(dtree.check(tree, test_set), 5)
])
print(tabulate(data, header))
datasets_test):
data = []
mean_errors = []
stdev = []
for fraction in fractions:
errors = []
for i in range(n):
monktrain, monkval = partition(dataset, fraction)
built_tree = dtree.buildTree(monktrain, m.attributes)
best_tree = get_best_tree(built_tree)
errors.append(1 - dtree.check(best_tree, dataset_test))
mean_error = round(statistics.mean(errors), decimals)
mean_errors.append(round(statistics.mean(errors), decimals))
stdev.append(round(statistics.stdev(errors), decimals))
data.append([fraction, mean_error, statistics.mean(stdev)])
print(tabulate(data, header), '\n')
plt.errorbar(fractions, mean_errors, yerr=stdev, marker='o')
plt.title('{} (n = {})'.format(dataset_name, n))
plt.xlabel('fraction')
plt.ylabel('mean error')
def main():
monk1 = m.monk1
monk2 = m.monk2
monk3 = m.monk3
entropy = dec.entropy(m.monk1)
print("MONK 1: ", entropy)
entropy = dec.entropy(m.monk2)
print("MONK 2: ", entropy)
entropy = dec.entropy(m.monk3)
print("MONK 3: ", entropy)
print(dec.bestAttribute(monk1, m.attributes))
print(("MONK1:"))
for i in range(6):
print("Information gain of a" + str(i + 1) + " is " +
str(dec.averageGain(monk1, m.attributes[i])))
print(("MONK2:"))
for i in range(6):
print("Information gain of a" + str(i + 1) + " is " +
str(dec.averageGain(monk2, m.attributes[i])))
print(("MONK3:"))
for i in range(6):
print("Information gain of a" + str(i + 1) + " is " +
str(dec.averageGain(monk3, m.attributes[i])))
print(("MONK1:"))
tree = dec.buildTree(monk1, m.attributes)
print("Training Error: ", 1 - dec.check(tree, m.monk1))
print("Test Error: ", 1 - dec.check(tree, m.monk1test))
print(("MONK2:"))
tree = dec.buildTree(monk2, m.attributes)
print("Training Error: ", 1 - dec.check(tree, m.monk2))
print("Test Error: ", 1 - dec.check(tree, m.monk2test))
print(("MONK3:"))
tree = dec.buildTree(monk3, m.attributes)
print("Training Error: ", 1 - dec.check(tree, m.monk3))
print("Test Error: ", 1 - dec.check(tree, m.monk3test))
def partition(data, fraction):
ldata = list(data)
random.shuffle(ldata)
breakPoint = int(len(ldata) * fraction)
return ldata[:breakPoint], ldata[breakPoint:]
values = [0.3, 0.4, 0.5, 0.6, 0.7, 0.8]
best_test_error = 0
model_score = []
best_fraction = 0
for fraction in values:
total_test_error = 0
for j in range(1000):
monk3train, monk3val = partition(m.monk3, fraction)
tree = dec.buildTree(monk3train, m.attributes)
result = pick_best_tree(tree, monk3val)
total_test_error += dec.check(result, m.monk3test)
avg_test_error = total_test_error / 1000
model_score.append(avg_test_error)
if avg_test_error > best_test_error:
best_test_error = avg_test_error
best_fraction = fraction
plt.scatter(values, model_score)
plt.xlabel("Split fraction")
plt.ylabel("Test accuracy")
plt.savefig("Monk3.png")
plt.show()
best_test_error = 0
model_score = []
best_fraction = 0
for fraction in values:
total_test_error = 0
for j in range(1000):
monk1train, monk1val = partition(m.monk1, fraction)
tree = dec.buildTree(monk1train, m.attributes)
result = pick_best_tree(tree, monk1val)
total_test_error += dec.check(result, m.monk1test)
avg_test_error = total_test_error / 1000
model_score.append(avg_test_error)
if avg_test_error > best_test_error:
best_test_error = avg_test_error
best_fraction = fraction
plt.scatter(values, model_score)
plt.xlabel("Split fraction")
plt.ylabel("Test accuracy")
plt.savefig("Monk1.png")
plt.show()
print(best_fraction)
raw.set_title('Mean and raw values')
raw.set_xlabel('fraction')
raw.set_ylabel('error')
stat = plt.subplot(2, 1, 2)
stat.set_title('Standard deviation')
stat.set_xlabel('fraction')
stat.set_ylabel('standard deviation')
for j in range(3, 9):
print j
total = 0
for k in range(size):
monk1train, monk1val = partition(m.monk3, j / 10.0)
t = d.buildTree(monk1train, m.attributes)
checkT = d.check(t, monk1val)
while True:
hold = checkT
tprune = d.allPruned(t)
for i in tprune:
temp = (d.check(i, monk1val))
if (temp > checkT):
checkT = temp
t = i
#print checkT
if (checkT == hold):
break
from python import dtree as d
from python import monkdata as m
print("Error rate MONK 1")
t= d.buildTree(m.monk1, m.attributes)
print(1-d.check(t,m.monk1test))
print("Error rate MONK 2")
t= d.buildTree(m.monk2, m.attributes)
print(1-d.check(t,m.monk2test))
print("Error rate MONK 3")
t= d.buildTree(m.monk3, m.attributes)
print(1-d.check(t,m.monk3test))
def assignment_7():
print("*** ASSIGNMENT 7 ***")
samples = 100
fractions = (0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.925, 0.95,
0.99)
datasets = {
'monk1': {
'training': monk1,
'test': monk1test
},
'monk3': {
'training': monk3,
'test': monk3test
}
}
results = {}
for dataset_name, dataset_data in datasets.items():
dataset = dataset_data['training']
dataset_test = dataset_data['test']
results[dataset_name] = {}
for fraction in fractions:
errors = []
for _ in range(samples):
train, validation = partition(dataset, fraction)
tree = buildTree(train, attributes)
opt_tree, _ = optimum_prune(tree, validation)
errors.append(1.0 - check(opt_tree, dataset_test))
results[dataset_name][fraction] = {
'mean': np.mean(errors),
'median': np.median(errors),
'std': np.std(errors),
'max': max(errors),
'min': min(errors)
}
pp = PrettyPrinter(indent=4)
pp.pprint(results)
y_monk1 = [(stats['mean'], stats['std'])
for fraction, stats in results['monk1'].items()]
y_monk3 = [(stats['mean'], stats['std'])
for fraction, stats in results['monk3'].items()]
plt.figure()
plt.errorbar(fractions, [e[0] for e in y_monk1],
yerr=[e[1] for e in y_monk1],
fmt='or',
capsize=5,
label='MONK-1')
plt.errorbar(fractions, [e[0] for e in y_monk3],
yerr=[e[1] for e in y_monk3],
fmt='ob',
capsize=5,
label='MONK-3')
plt.xlabel('Pruning fraction size (relative size of training set)')
plt.ylabel('Classification error')
plt.legend()
plt.title(
'Error vs. fraction size (mean of {} samples, errorbars represent one standard deviation)'
.format(samples))
plt.show()