def update_weight(self, xseq, yseq, **kwargs):
"""
xseq (list): list of np.arrays, each of size (n_feat)
yseq (list): list of class labels (int)
"""
L = len(yseq)
phi_start = np.zeros(self.n_classes)
phi_start[yseq[0]] = 1
initial_scores = np.multiply(self.W_start, phi_start)
phi_stop = np.zeros(self.n_classes)
phi_stop[yseq[-1]] = 1
final_scores = np.multiply(self.W_stop, phi_stop)
phi_transition = np.zeros((L - 1, self.n_classes, self.n_classes))
s_t = np.zeros((L - 1, self.n_classes, self.n_classes))
for i in range(L - 1):
phi_transition[i][yseq[i + 1], yseq[i]] = 1
s_t[i] = np.dot(self.W_bigrams, phi_transition[i])
s_u = np.dot(np.array(xseq), self.W_unigrams.T)
yseq_hat = viterbi(initial_scores, s_t, final_scores, s_u)
if yseq_hat != yseq:
for t in range(L):
if yseq_hat[t] != yseq[t]:
self.W_unigrams[yseq_hat[t]] -= xseq[t]
self.W_unigrams[yseq[t]] += xseq[t]
if t == 0:
self.W_start[yseq_hat[t]] -= 1
self.W_start[yseq[t]] += 1
else:
self.W_bigrams[yseq_hat[t],
yseq_hat[t - 1]] -= phi_transition[
t - 1][yseq_hat[t], yseq_hat[t - 1]]
self.W_bigrams[yseq[t],
yseq[t -
1]] += phi_transition[t -
1][yseq[t],
yseq[t -
1]]
if t == L - 1:
self.W_stop[yseq[t]] += 1
self.W_stop[yseq_hat[t]] -= 1
return 1
return 0
def main():
weather = ["Sunny", "Windy", "Rainy"]
activities = ["Surf", "Beach", "Videogame", "Study"]
str2state = {w: i for i, w in enumerate(weather)}
str2emission = {e: i for i, e in enumerate(activities)}
# emissions probabilities from the handout: emission_probabilities[i, j]
# is the probability of the emission i given the state j
emission_probabilities = np.array([[0.4, 0.5, 0.1], [0.4, 0.1, 0.1],
[0.1, 0.2, 0.3], [0.1, 0.2, 0.5]])
# transition probabilities from the handout: transition_probabilities[i, j]
# is the probability of transitioning to state i from state j
transition_probabilities = np.array([[0.6, 0.3, 0.2], [0.3, 0.5, 0.3],
[0.1, 0.2, 0.5]])
observations = [
"Videogame", "Study", "Study", "Surf", "Beach", "Videogame", "Beach"
]
initial_weather = "Rainy"
final_weather = "Sunny"
x = [str2emission[observation] for observation in observations]
emission_scores = np.log(emission_probabilities[x])
transition_scores = np.log(
np.array([transition_probabilities for observation in observations]))
initial_state = str2state[initial_weather]
final_state = str2state[final_weather]
initial_scores = np.log(transition_probabilities[:, initial_state])
final_scores = np.log(transition_probabilities[final_state])
viterbi_path = viterbi(initial_scores, transition_scores, final_scores,
emission_scores)
from IPython.core.debugger import Pdb
Pdb().set_trace()
viterbi_weather = [weather[i] for i in viterbi_path]
posteriors, _, _ = forward_backward(initial_scores, transition_scores,
final_scores, emission_scores)
posterior_path = posteriors.argmax(1)
posterior_weather = [weather[i] for i in posterior_path]
print("Sequence given by Viterbi: %s" % " -> ".join(viterbi_weather))
print("Sequence given by posterior: %s" % " -> ".join(posterior_weather))
def evaluate(self, X, y):
"""Evaluate model on data."""
correct = 0
total = 0
for xseq, yseq in zip(X, y):
if self.feature_function is not None:
xseq = [self.feature_function(x) for x in xseq]
emission_scores = np.zeros((len(xseq), self.n_classes))
transition_scores = np.zeros(
(len(xseq) - 1, self.n_classes, self.n_classes))
initial_scores = np.zeros(self.n_classes)
final_scores = np.zeros(self.n_classes)
initial_scores[:] = self.W_start
for t in range(len(xseq)):
emission_scores[t] = self.W_unigrams.dot(xseq[t])
if t > 0:
transition_scores[t - 1] = self.W_bigrams
final_scores[:] = self.W_stop
yseq_hat = viterbi(initial_scores, transition_scores, final_scores,
emission_scores)
correct += sum([yseq[t] == yseq_hat[t] for t in range(len(yseq))])
total += len(yseq)
return correct / total