def fake_signal():
signals = get_signals()
# assume all signals are same length
signals = np.array(signals)
hmm = HMM(5, 3, signals[0].shape[1]) # M, K, D
hmm.init_random(signals)
init = tf.global_variables_initializer()
session = tf.InteractiveSession()
session.run(init)
hmm.set_session(session)
hmm.fit(signals, max_iter=30)
L = hmm.get_cost_multi(signals).sum()
print "LL for fitted params:", L
# test in actual params
_, _, _, pi, A, R, mu, sigma = big_init()
# turn these into their "pre-softmax" forms
pi = np.log(pi)
A = np.log(A)
R = np.log(R)
M, K, D, _ = sigma.shape # need to convert full cov into diag cov
logSigma = np.zeros((M, K, D))
for j in range(M):
for k in range(D):
logSigma[j,k] = np.log(np.diag(sigma[j,k]))
hmm.set(pi, A, R, mu, logSigma)
L = hmm.get_cost_multi(signals).sum()
print "LL for actual params:", L
def fake_signal():
signals = get_signals()
# assume all signals are same length
signals = np.array(signals)
hmm = HMM(5, 3, signals[0].shape[1]) # M, K, D
hmm.init_random(signals)
init = tf.global_variables_initializer()
session = tf.InteractiveSession()
session.run(init)
hmm.set_session(session)
hmm.fit(signals, max_iter=30)
L = hmm.get_cost_multi(signals).sum()
print("LL for fitted params:", L)
# test in actual params
_, _, _, pi, A, R, mu, sigma = big_init()
# turn these into their "pre-softmax" forms
pi = np.log(pi)
A = np.log(A)
R = np.log(R)
M, K, D, _ = sigma.shape # need to convert full cov into diag cov
logSigma = np.zeros((M, K, D))
for j in range(M):
for k in range(D):
logSigma[j,k] = np.log(np.diag(sigma[j,k]))
hmm.set(pi, A, R, mu, logSigma)
L = hmm.get_cost_multi(signals).sum()
print("LL for actual params:", L)
def fake_signal():
signals = get_signals()
hmm = HMM(5, 3)
hmm.fit(signals)
L = hmm.log_likelihood_multi(signals).sum()
print "LL for fitted params:", L
# test in actual params
_, _, _, pi, A, R, mu, sigma = big_init()
hmm.set(pi, A, R, mu, sigma)
L = hmm.log_likelihood_multi(signals).sum()
print "LL for actual params:", L
def fake_signal():
signals = get_signals()
hmm = HMM(5, 3)
hmm.fit(signals)
L = hmm.log_likelihood_multi(signals).sum()
print "LL for fitted params:", L
# test in actual params
_, _, _, pi, A, R, mu, sigma = big_init()
hmm.set(pi, A, R, mu, sigma)
L = hmm.log_likelihood_multi(signals).sum()
print "LL for actual params:", L
def fake_signal():
signals = get_signals()
hmm = HMM(5, 3)
hmm.fit(signals, max_iter=3)
L = hmm.log_likelihood_multi(signals).sum()
print("LL for fitted params:", L)
# test in actual params
_, _, _, pi, A, R, mu, sigma = big_init()
# turn these into their "pre-softmax" forms
pi = np.log(pi)
A = np.log(A)
R = np.log(R)
# decompose sigma using cholesky factorization
sigma = np.linalg.cholesky(sigma)
hmm.set(pi, A, R, mu, sigma)
L = hmm.log_likelihood_multi(signals).sum()
print("LL for actual params:", L)
=======
hmm.fit(signal.reshape(1, T, 1), learning_rate=1e-5, max_iter=20)
>>>>>>> upstream/master
def fake_signal():
signals = get_signals()
hmm = HMM(5, 3)
hmm.fit(signals)
L = hmm.log_likelihood_multi(signals).sum()
<<<<<<< HEAD
print "LL for fitted params:", L
=======
print("LL for fitted params:", L)
>>>>>>> upstream/master
# test in actual params
_, _, _, pi, A, R, mu, sigma = big_init()
hmm.set(pi, A, R, mu, sigma)
L = hmm.log_likelihood_multi(signals).sum()
<<<<<<< HEAD
print "LL for actual params:", L
=======
print("LL for actual params:", L)
>>>>>>> upstream/master
if __name__ == '__main__':
# real_signal()
fake_signal()
