def square():
graph = [
[0, 1, 2],
[1, ~0, 3],
[2, ~1, 3],
[3, ~1, ~0],
]
NUM_QUESTION = len(set(g[0] for g in graph))
assert NUM_QUESTION == 4
NUM_ANSWER = len(set(~g[i] for g in graph for i in [1, 2] if g[i] < 0))
assert NUM_ANSWER == 2
data = graph2data(graph, 0, NUM_QUESTION, NUM_ANSWER)
NUM_DATA = len(data)
mat_phi = np.array([d[0] for d in data])
mat_t = np.array([d[1] for d in data])
weights = [logreg.learn((mat_phi, mat_t[:, i])) for i in range(NUM_ANSWER)]
for d in mat_phi:
probs = logreg.get_multi_prob(d, weights)
print d
print ' '.join('%0.2f' % p for p in probs)
print weights
def sampling():
"sampling from all data"
SIZE = 10
NUM_QUESTION = SIZE
NUM_ANSWER = SIZE + 1
graph = generate_tree_graph(SIZE)
data = list(graph2all_data(graph))
from random import shuffle
shuffle(data)
data = data[:11]
mat_phi = np.array([d[0] for d in data])
print mat_phi
mat_t = np.array([d[1] for d in data])
print mat_t
weights = [
logreg.learn((mat_phi, mat_t[:, i])) for i in range(mat_t.shape[1])
]
for d in mat_phi:
probs = logreg.get_multi_prob(d, weights)
print d
print ' '.join('%0.2f' % p for p in probs)
print_mat(weights)
print correct_ratio(NUM_QUESTION, NUM_ANSWER, weights)
return weights
def change_scale(size=10, add_offset_column=True):
SIZE = 10
NUM_QUESTION = SIZE
NUM_ANSWER = SIZE + 1
graph = [[i, ~i, i + 1] for i in range(SIZE)]
graph[-1][-1] = ~SIZE
data = list(graph2all_data(graph, add_offset_column=add_offset_column))
mat_phi = np.array([d[0] for d in data])
mat_phi[:, 1] *= 10
mat_phi[:, 3] *= 100
print mat_phi
mat_t = np.array([d[1] for d in data])
print mat_t
weights = [
logreg.learn((mat_phi, mat_t[:, i]), 1000)
for i in range(mat_t.shape[1])
]
for d in mat_phi:
probs = logreg.get_multi_prob(d, weights)
print d
print ' '.join('%0.2f' % p for p in probs)
print_mat(weights)
print correct_ratio(NUM_QUESTION,
NUM_ANSWER,
weights,
add_offset_column=add_offset_column)
def square():
graph = [
[0, 1, 2],
[1, ~0, 3],
[2, ~1, 3],
[3, ~1, ~0],
]
NUM_QUESTION = len(set(g[0] for g in graph))
assert NUM_QUESTION == 4
NUM_ANSWER = len(set(~g[i] for g in graph for i in [1, 2] if g[i] < 0))
assert NUM_ANSWER == 2
data = graph2data(graph, 0, NUM_QUESTION, NUM_ANSWER)
NUM_DATA = len(data)
mat_phi = np.array([d[0] for d in data])
mat_t = np.array([d[1] for d in data])
weights = [logreg.learn((mat_phi, mat_t[:, i])) for i in range(NUM_ANSWER)]
for d in mat_phi:
probs = logreg.get_multi_prob(d, weights)
print d
print ' '.join('%0.2f' % p for p in probs)
print weights
def one_way_tree2():
SIZE = 10
NUM_QUESTION = SIZE - 1
NUM_ANSWER = SIZE
NUM_DATA = SIZE
mat_phi = np.zeros((NUM_DATA, NUM_QUESTION))
mat_t = np.zeros((NUM_DATA, NUM_ANSWER))
for i in range(NUM_DATA):
mat_t[i][i] = 1.0
for j in range(i):
mat_phi[i, j] = -1.0
if i < NUM_QUESTION:
mat_phi[i, i] = +1.0
weights = [logreg.learn((mat_phi, mat_t[:, i])) for i in range(NUM_ANSWER)]
for d in mat_phi:
probs = logreg.get_multi_prob(d, weights)
print d
print ' '.join('%0.2f' % p for p in probs)
from itertools import product
qs = product([-1, 1], repeat=NUM_QUESTION)
num_ok = 0
for q in qs:
if not 1.0 in q:
a = NUM_ANSWER - 1
#else:
# a = q.index(1.0)
probs = logreg.get_multi_prob(np.array(q), weights)
if np.array(probs).argmax() == a:
num_ok += 1
print num_ok / (2.0**NUM_QUESTION)
def change_scale(size=10, add_offset_column=True):
SIZE = 10
NUM_QUESTION = SIZE
NUM_ANSWER = SIZE + 1
graph = [
[i, ~i, i + 1]
for i in range(SIZE)
]
graph[-1][-1] = ~SIZE
data = list(graph2all_data(graph, add_offset_column=add_offset_column))
mat_phi = np.array([d[0] for d in data])
mat_phi[:,1] *= 10
mat_phi[:,3] *= 100
print mat_phi
mat_t = np.array([d[1] for d in data])
print mat_t
weights = [logreg.learn((mat_phi, mat_t[:, i]), 1000) for i in range(mat_t.shape[1])]
for d in mat_phi:
probs = logreg.get_multi_prob(d, weights)
print d
print ' '.join('%0.2f' % p for p in probs)
print_mat(weights)
print correct_ratio(NUM_QUESTION, NUM_ANSWER, weights, add_offset_column=add_offset_column)
def sampling():
"sampling from all data"
SIZE = 10
NUM_QUESTION = SIZE
NUM_ANSWER = SIZE + 1
graph = generate_tree_graph(SIZE)
data = list(graph2all_data(graph))
from random import shuffle
shuffle(data)
data = data[:11]
mat_phi = np.array([d[0] for d in data])
print mat_phi
mat_t = np.array([d[1] for d in data])
print mat_t
weights = [logreg.learn((mat_phi, mat_t[:, i])) for i in range(mat_t.shape[1])]
for d in mat_phi:
probs = logreg.get_multi_prob(d, weights)
print d
print ' '.join('%0.2f' % p for p in probs)
print_mat(weights)
print correct_ratio(NUM_QUESTION, NUM_ANSWER, weights)
return weights
def one_way_tree2():
SIZE = 10
NUM_QUESTION = SIZE - 1
NUM_ANSWER = SIZE
NUM_DATA = SIZE
mat_phi = np.zeros((NUM_DATA, NUM_QUESTION))
mat_t = np.zeros((NUM_DATA, NUM_ANSWER))
for i in range(NUM_DATA):
mat_t[i][i] = 1.0
for j in range(i):
mat_phi[i, j] = -1.0
if i < NUM_QUESTION:
mat_phi[i, i] = +1.0
weights = [logreg.learn((mat_phi, mat_t[:, i])) for i in range(NUM_ANSWER)]
for d in mat_phi:
probs = logreg.get_multi_prob(d, weights)
print d
print ' '.join('%0.2f' % p for p in probs)
from itertools import product
qs = product([-1, 1], repeat=NUM_QUESTION)
num_ok = 0
for q in qs:
if not 1.0 in q:
a = NUM_ANSWER - 1
#else:
# a = q.index(1.0)
probs = logreg.get_multi_prob(np.array(q), weights)
if np.array(probs).argmax() == a:
num_ok += 1
print num_ok / (2.0 ** NUM_QUESTION)
def with_all_data(graph=graph):
"with all data"
data = list(graph2all_data(graph))
mat_phi = np.array([d[0] for d in data])
print mat_phi
mat_t = np.array([d[1] for d in data])
print mat_t
weights = [logreg.learn((mat_phi, mat_t[:, i])) for i in range(mat_t.shape[1])]
for d in mat_phi:
probs = logreg.get_multi_prob(d, weights)
print d
print ' '.join('%0.2f' % p for p in probs)
print_mat(weights)
return weights
def with_all_data(graph=graph):
"with all data"
data = list(graph2all_data(graph))
mat_phi = np.array([d[0] for d in data])
print mat_phi
mat_t = np.array([d[1] for d in data])
print mat_t
weights = [
logreg.learn((mat_phi, mat_t[:, i])) for i in range(mat_t.shape[1])
]
for d in mat_phi:
probs = logreg.get_multi_prob(d, weights)
print d
print ' '.join('%0.2f' % p for p in probs)
print_mat(weights)
return weights
def one_way_tree():
SIZE = 4
NUM_QUESTION = SIZE - 1
NUM_ANSWER = SIZE
NUM_DATA = SIZE
mat_phi = np.zeros((NUM_DATA, NUM_QUESTION))
mat_t = np.zeros((NUM_DATA, NUM_ANSWER))
for i in range(NUM_DATA):
mat_t[i][i] = 1.0
for j in range(i):
mat_phi[i, j] = -1.0
if i < NUM_QUESTION:
mat_phi[i, i] = +1.0
weights = [logreg.learn((mat_phi, mat_t[:, i])) for i in range(NUM_ANSWER)]
for d in mat_phi:
probs = logreg.get_multi_prob(d, weights)
print d
print ' '.join('%0.2f' % p for p in probs)
print_mat(np.array(weights))
def one_way_tree():
SIZE = 4
NUM_QUESTION = SIZE - 1
NUM_ANSWER = SIZE
NUM_DATA = SIZE
mat_phi = np.zeros((NUM_DATA, NUM_QUESTION))
mat_t = np.zeros((NUM_DATA, NUM_ANSWER))
for i in range(NUM_DATA):
mat_t[i][i] = 1.0
for j in range(i):
mat_phi[i, j] = -1.0
if i < NUM_QUESTION:
mat_phi[i, i] = +1.0
weights = [logreg.learn((mat_phi, mat_t[:, i])) for i in range(NUM_ANSWER)]
for d in mat_phi:
probs = logreg.get_multi_prob(d, weights)
print d
print ' '.join('%0.2f' % p for p in probs)
print_mat(np.array(weights))