def q2_5():
gold_words = load_gold('test')
test_size = len(gold_words)
test_chars_size = 0
word_acc = 0.0
char_acc = 0.0
for i in range(1, test_size + 1):
data_fname = "../data/test_img%d.txt" % (i)
data = load_data(data_fname)
f_weights = load_feature_weights()
t_weights = load_transition_weights()
node_potentials = crf.compute_node_potentials(data, f_weights)
clique_potentials = crf.compute_clique_potentials(
node_potentials, t_weights)
forward, backward = crf.compute_messages(clique_potentials)
beliefs = crf.compute_beliefs(clique_potentials, forward, backward)
pos_probs, _ = crf.compute_marginals(clique_potentials, beliefs)
word = crf.classify(pos_probs)
gold = ''.join(map(lambda c: char_map[c], gold_words[i - 1]))
word_acc += 1.0 if word == gold else 0.0
chars_correct = sum(
[1 if word[j] == gold[j] else 0 for j in range(len(word))])
char_acc += chars_correct
test_chars_size += len(word)
if i < 6: # only print first 5 test words
print word, gold
word_acc /= test_size
char_acc /= test_chars_size
print "Word accuracy:", word_acc
print "Character accuracy:", char_acc
def q2_5():
gold_words = load_gold('test')
test_size = len(gold_words)
test_chars_size = 0
word_acc = 0.0
char_acc = 0.0
for i in range(1,test_size+1):
data_fname = "../data/test_img%d.txt" % (i)
data = load_data(data_fname)
f_weights = load_feature_weights()
t_weights = load_transition_weights()
node_potentials = crf.compute_node_potentials(data, f_weights)
clique_potentials = crf.compute_clique_potentials(node_potentials, t_weights)
forward, backward = crf.compute_messages(clique_potentials)
beliefs = crf.compute_beliefs(clique_potentials, forward, backward)
pos_probs, _ = crf.compute_marginals(clique_potentials, beliefs)
word = crf.classify(pos_probs)
gold = ''.join(map(lambda c: char_map[c], gold_words[i-1]))
word_acc += 1.0 if word==gold else 0.0
chars_correct = sum([1 if word[j]==gold[j] else 0 for j in range(len(word))])
char_acc += chars_correct
test_chars_size += len(word)
if i < 6 : # only print first 5 test words
print word, gold
word_acc /= test_size
char_acc /= test_chars_size
print "Word accuracy:", word_acc
print "Character accuracy:", char_acc
def q2_3():
data_fname = "../data/test_img1.txt"
data = load_data(data_fname)
f_weights = load_feature_weights()
t_weights = load_transition_weights()
node_potentials = crf.compute_node_potentials(data, f_weights)
clique_potentials = crf.compute_clique_potentials(node_potentials, t_weights)
forward, backward = crf.compute_messages(clique_potentials)
beliefs = crf.compute_beliefs(clique_potentials, forward, backward)
print "beliefs:\n", beliefs[:,:2,:2] # only want e, t
def q2_2():
data_fname = "../data/test_img1.txt"
data = load_data(data_fname)
f_weights = load_feature_weights()
t_weights = load_transition_weights()
node_potentials = crf.compute_node_potentials(data, f_weights)
clique_potentials = crf.compute_clique_potentials(node_potentials, t_weights)
forward, backward = crf.compute_messages(clique_potentials)
print "forward:\n", forward
print "backward:\n", backward
def q2_3():
data_fname = "../data/test_img1.txt"
data = load_data(data_fname)
f_weights = load_feature_weights()
t_weights = load_transition_weights()
node_potentials = crf.compute_node_potentials(data, f_weights)
clique_potentials = crf.compute_clique_potentials(node_potentials,
t_weights)
forward, backward = crf.compute_messages(clique_potentials)
beliefs = crf.compute_beliefs(clique_potentials, forward, backward)
print "beliefs:\n", beliefs[:, :2, :2] # only want e, t
def q2_2():
data_fname = "../data/test_img1.txt"
data = load_data(data_fname)
f_weights = load_feature_weights()
t_weights = load_transition_weights()
node_potentials = crf.compute_node_potentials(data, f_weights)
clique_potentials = crf.compute_clique_potentials(node_potentials,
t_weights)
forward, backward = crf.compute_messages(clique_potentials)
print "forward:\n", forward
print "backward:\n", backward
def q2_4():
data_fname = "../data/test_img1.txt"
data = load_data(data_fname)
f_weights = load_feature_weights()
t_weights = load_transition_weights()
node_potentials = crf.compute_node_potentials(data, f_weights)
clique_potentials = crf.compute_clique_potentials(node_potentials, t_weights)
forward, backward = crf.compute_messages(clique_potentials)
beliefs = crf.compute_beliefs(clique_potentials, forward, backward)
pos_probs, trans_probs = crf.compute_marginals(clique_potentials, beliefs)
print "marginals:\n", pos_probs
# uncomment to print LaTeX tabular version
#print_marginals_table(pos_probs)
indices = (0,1,8) # only want e, t, r
print "pairwise:\n", trans_probs[...,indices][:,indices]
def q2_4():
data_fname = "../data/test_img1.txt"
data = load_data(data_fname)
f_weights = load_feature_weights()
t_weights = load_transition_weights()
node_potentials = crf.compute_node_potentials(data, f_weights)
clique_potentials = crf.compute_clique_potentials(node_potentials,
t_weights)
forward, backward = crf.compute_messages(clique_potentials)
beliefs = crf.compute_beliefs(clique_potentials, forward, backward)
pos_probs, trans_probs = crf.compute_marginals(clique_potentials, beliefs)
print "marginals:\n", pos_probs
# uncomment to print LaTeX tabular version
#print_marginals_table(pos_probs)
indices = (0, 1, 8) # only want e, t, r
print "pairwise:\n", trans_probs[..., indices][:, indices]
def test_model(f_weights, t_weights):
test_chars_size = 0
word_acc = 0.0
char_acc = 0.0
for i in range(test_size):
node_potentials = crf.compute_node_potentials(test_data[i], f_weights)
clique_potentials = crf.compute_clique_potentials(node_potentials, t_weights)
forward, backward = crf.compute_messages(clique_potentials)
beliefs = crf.compute_beliefs(clique_potentials, forward, backward)
pos_probs, _ = crf.compute_marginals(clique_potentials, beliefs)
word = crf.classify(pos_probs)
gold = ''.join(map(lambda c: char_map[c], gold_test_words[i]))
word_acc += 1.0 if word == gold else 0.0
chars_correct = sum([1 if word[j] == gold[j] else 0 for j in range(len(word))])
char_acc += chars_correct
test_chars_size += len(word)
word_acc /= test_size
char_acc /= test_chars_size
return 1-char_acc
def get_sum_prod_marginals(num_examples, f_weights, t_weights):
all_node_potentials = [
crf.compute_node_potentials(example, f_weights)
for example in training_data[:num_examples]
]
all_clique_potentials = [
crf.compute_clique_potentials(node_potentials, t_weights)
for node_potentials in all_node_potentials
]
all_fb = [
crf.compute_messages(clique_potentials)
for clique_potentials in all_clique_potentials
]
all_beliefs = [
crf.compute_beliefs(all_clique_potentials[i], all_fb[i][0],
all_fb[i][1]) for i in range(num_examples)
]
return np.array([
crf.compute_marginals(all_clique_potentials[i], all_beliefs[i])
for i in range(num_examples)
])
def test_model(f_weights, t_weights):
test_chars_size = 0
word_acc = 0.0
char_acc = 0.0
for i in range(test_size):
node_potentials = crf.compute_node_potentials(test_data[i], f_weights)
clique_potentials = crf.compute_clique_potentials(
node_potentials, t_weights)
forward, backward = crf.compute_messages(clique_potentials)
beliefs = crf.compute_beliefs(clique_potentials, forward, backward)
pos_probs, _ = crf.compute_marginals(clique_potentials, beliefs)
word = crf.classify(pos_probs)
gold = ''.join(map(lambda c: char_map[c], gold_test_words[i]))
word_acc += 1.0 if word == gold else 0.0
chars_correct = sum(
[1 if word[j] == gold[j] else 0 for j in range(len(word))])
char_acc += chars_correct
test_chars_size += len(word)
word_acc /= test_size
char_acc /= test_chars_size
return 1 - char_acc
def get_sum_prod_marginals(num_examples, f_weights, t_weights):
all_node_potentials = [crf.compute_node_potentials(example, f_weights) for example in training_data[:num_examples]]
all_clique_potentials = [crf.compute_clique_potentials(node_potentials, t_weights) for node_potentials in all_node_potentials]
all_fb = [crf.compute_messages(clique_potentials) for clique_potentials in all_clique_potentials]
all_beliefs = [crf.compute_beliefs(all_clique_potentials[i], all_fb[i][0], all_fb[i][1]) for i in range(num_examples)]
return np.array([crf.compute_marginals(all_clique_potentials[i], all_beliefs[i]) for i in range(num_examples)])
'''
CS688 HW02: Chain CRF
Question 3: Maximum Log Likelihood Learning Derivation
@author: Emma Strubell
'''
from data_loader import *
import chain_crf as crf
# compute log likelihood over first 50 training examples
gold_words = load_gold('train')
train_size = 50
f_weights = load_feature_weights()
t_weights = load_transition_weights()
total = 0.0
for i in range(train_size):
data_fname = "../data/train_img%d.txt" % (i+1)
example = load_data(data_fname)
node_potentials = crf.compute_node_potentials(example, f_weights)
clique_potentials = crf.compute_clique_potentials(node_potentials, t_weights)
forward, backward = crf.compute_messages(clique_potentials)
beliefs = crf.compute_beliefs(clique_potentials, forward, backward)
pos_probs, _ = crf.compute_marginals(clique_potentials, beliefs)
total += np.sum(np.log([pos_probs[i,j] for i,j in enumerate(gold_words[i])]))
total /= train_size
print "Log likelihood:", total