def test_belief_propagation():
n_states = 4
n_obs = 3
T0 = torch.rand([n_states]).softmax(0)
T = torch.rand([n_states, n_states]).softmax(1)
states = [
CategoricalModel(probs=torch.rand([n_obs]).softmax(0))
for i in range(n_states)
]
model = HiddenMarkovModel(states, T0=T0, T=T)
# compute max path from each node to each node
# will give a matrix, row is the source col is the dest
# res = torch.log(T0 * T)
res = torch.stack([torch.log(T0[i] * T[i, :]) for i in range(len(T0))])
log_T0 = T0.log()
max_b, _ = model._belief_prop_max(log_T0.unsqueeze(0))
assert torch.allclose(res.max(0).values, max_b[0])
sum_b = model._belief_prop_sum(log_T0.unsqueeze(0))
assert torch.allclose(res.logsumexp(0), sum_b[0])
def test_hmm_predict_aima_umbrella_example():
"""
This example was taken from AI a Modern Approach
The example comes from page 572, filtering and prediction section. The
correct values were manually compared propagated one step into future from
the filtering example.
"""
# states = {0: 'No Rain', 1: 'Rain'}
# obs = {0: 'No Umbrella', 1: 'Umbrella'}
T0 = torch.tensor([0.5, 0.5])
T = torch.tensor([[0.7, 0.3], [0.3, 0.7]])
s1_orig = torch.tensor([0.8, 0.2])
s2_orig = torch.tensor([0.1, 0.9])
s1 = CategoricalModel(probs=s1_orig)
s2 = CategoricalModel(probs=s2_orig)
model = HiddenMarkovModel([s1, s2], T0=T0, T=T)
obs_seq = pack_list([torch.tensor([1])])
posterior = model.predict(obs_seq)
normalized = posterior.softmax(1)
correct = torch.tensor([[0.37272727, 0.62727273]])
assert torch.allclose(normalized, correct)
obs_seq = pack_list([torch.tensor([1, 1])])
posterior = model.predict(obs_seq)
normalized = posterior.softmax(1)
correct = torch.tensor([[0.34665718, 0.65334282]])
assert torch.allclose(normalized, correct)
def test_hmm_filter_aima_umbrella_example():
"""
This example was taken from AI a Modern Approach
The example comes from page 572, filtering and prediction section. The
correct values were manually compared to the normalized values from this
example.
"""
# states = {0: 'No Rain', 1: 'Rain'}
# obs = {0: 'No Umbrella', 1: 'Umbrella'}
T0 = torch.tensor([0.5, 0.5])
T = torch.tensor([[0.7, 0.3], [0.3, 0.7]])
s1 = CategoricalModel(probs=torch.tensor([0.8, 0.2]))
s2 = CategoricalModel(probs=torch.tensor([0.1, 0.9]))
model = HiddenMarkovModel([s1, s2], T0=T0, T=T)
obs_seq = pack_list([torch.tensor([1])])
posterior = model.filter(obs_seq)
normalized = posterior.softmax(1)
correct = torch.tensor([[0.18181818, 0.81818182]])
assert torch.allclose(normalized, correct)
obs_seq = pack_list([torch.tensor([1, 1])])
posterior = model.filter(obs_seq)
normalized = posterior.softmax(1)
correct = torch.tensor([[0.11664296, 0.88335704]])
assert torch.allclose(normalized, correct)
def test_hmm_decode_aima_umbrella_example():
"""
This example was taken from AI a Modern Approach
The state sequence comes from figure 15.5(b) on page 577 of the third
edition. The correct values were manually compared to the normalized values
from the figure 15.5(b).
"""
states = {0: 'No Rain', 1: 'Rain'}
# obs = {0: 'No Umbrella', 1: 'Umbrella'}
T0 = torch.tensor([0.5, 0.5])
T = torch.tensor([[0.7, 0.3], [0.3, 0.7]])
s1 = CategoricalModel(probs=torch.tensor([0.8, 0.2]))
s2 = CategoricalModel(probs=torch.tensor([0.1, 0.9]))
model = HiddenMarkovModel([s1, s2], T0=T0, T=T)
obs_seq = pack_list([torch.tensor([1, 1, 0, 1, 1])])
states_seq, path_ll = model.decode(obs_seq)
ss_unpacked, _ = pad_packed_sequence(states_seq, batch_first=True)
path_unpacked, _ = pad_packed_sequence(path_ll, batch_first=True)
most_likely_states = [states[s.item()] for s in ss_unpacked[0]]
assert most_likely_states == ['Rain', 'Rain', 'No Rain', 'Rain', 'Rain']
normalized = path_unpacked[0].softmax(1)
correct = torch.tensor([[0.18181818, 0.81818182], [0.08695652, 0.91304348],
[0.77419355, 0.22580645], [0.34146341, 0.65853659],
[0.10332103, 0.89667897]])
assert torch.allclose(normalized, correct)
def test_hmm_parameters():
T0 = torch.tensor([0.5, 0.5])
T = torch.tensor([[0.7, 0.3], [0.3, 0.7]])
s1_orig = torch.tensor([0.8, 0.2])
s2_orig = torch.tensor([0.1, 0.9])
s1 = CategoricalModel(probs=s1_orig)
s2 = CategoricalModel(probs=s2_orig)
model = HiddenMarkovModel([s1, s2], T0=T0, T=T)
actual = [T0.log(), T.log(), s1_orig, s2_orig]
for i, p in enumerate(model.parameters()):
if i > 1:
p = torch.stack(p)
assert torch.isclose(actual[i], p).all()
def test_hmm_decode():
states = {0: 'Healthy', 1: 'Fever'}
# obs = {0: 'normal', 1: 'cold', 2: 'dizzy'}
T0 = torch.tensor([0.6, 0.4])
T = torch.tensor([[0.7, 0.3], [0.4, 0.6]])
s1 = CategoricalModel(probs=torch.tensor([0.5, 0.4, 0.1]))
s2 = CategoricalModel(probs=torch.tensor([0.1, 0.3, 0.6]))
model = HiddenMarkovModel([s1, s2], T0=T0, T=T)
obs_seq = pack_list([torch.tensor([1, 0, 1, 2, 2])])
states_seq, _ = model.decode(obs_seq)
ss_unpacked, _ = pad_packed_sequence(states_seq, batch_first=True)
most_likely_states = [states[s.item()] for s in ss_unpacked[0]]
assert most_likely_states == [
'Healthy', 'Healthy', 'Healthy', 'Fever', 'Fever'
]
def test_hmm_fit_viterbi_diagnormal():
T0 = torch.tensor([0.75, 0.25])
T = torch.tensor([[0.85, 0.15], [0.12, 0.88]])
s1_means = torch.tensor([0.0, 0.0, 0.0])
s1_precs = torch.tensor([1.0, 1.0, 1.0])
s2_means = torch.tensor([10.0, 10.0, 10.0])
s2_precs = torch.tensor([1.0, 1.0, 1.0])
s1 = DiagNormalModel(means=s1_means, precs=s1_precs)
s2 = DiagNormalModel(means=s2_means, precs=s2_precs)
model = HiddenMarkovModel([s1, s2], T0=T0, T=T)
obs_seq, states = model.sample(50, 100)
T0 = torch.tensor([0.75, 0.25])
T = torch.tensor([[0.85, 0.15], [0.12, 0.88]])
s1_means = torch.tensor([3.0, 3.0, 3.0])
s1_precs = torch.tensor([1.0, 1.0, 1.0])
s2_means = torch.tensor([6.0, 6.0, 6.0])
s2_precs = torch.tensor([1.0, 1.0, 1.0])
s1 = DiagNormalModel(means=s1_means, precs=s1_precs)
s2 = DiagNormalModel(means=s2_means, precs=s2_precs)
model = HiddenMarkovModel([s1, s2], T0=T0, T=T)
converge = model.fit(obs_seq, max_steps=500, epsilon=1e-2, restarts=3)
# Not enough samples (only 1) to test
# assert np.allclose(trans0.data.numpy(), True_pi)
print("Pi Matrix: ")
print(model.T0)
print("Transition Matrix: ")
print(model.T)
# assert np.allclose(transition.exp().data.numpy(), True_T, atol=0.1)
print()
print("Emission: ")
for s in model.states:
p = list(s.parameters())
print("Means", p[0])
print("Cov", p[1].abs())
# assert np.allclose(emission.exp().data.numpy(), True_E, atol=0.1)
print()
print("Reached Convergence: ")
print(converge)
assert converge
states_seq, _ = model.decode(obs_seq)
# state_summary = np.array([model.prob_state_1[i].cpu().numpy() for i in
# range(len(model.prob_state_1))])
# pred = (1 - state_summary[-2]) > 0.5
# pred = torch.cat(states_seq, 0).data.numpy()
# true = np.concatenate(states, 0)
pred = states_seq
true = states
accuracy = torch.mean(torch.abs(pred.data - true.data).float())
print("Accuracy: ", accuracy)
assert accuracy >= 0.9 or accuracy <= 0.1
def test_hmm_smooth():
T0 = torch.tensor([0.5, 0.5])
T = torch.tensor([[0.7, 0.3], [0.3, 0.7]])
s1_orig = torch.tensor([0.8, 0.2])
s2_orig = torch.tensor([0.1, 0.9])
s1 = CategoricalModel(probs=s1_orig)
s2 = CategoricalModel(probs=s2_orig)
model = HiddenMarkovModel([s1, s2], T0=T0, T=T)
obs_seq = pack_list([torch.tensor([1, 1])])
# New approach
posterior_ll = model.smooth(obs_seq)
post_unpacked, post_lengths = pad_packed_sequence(posterior_ll,
batch_first=True)
posterior_prob = post_unpacked[0].softmax(1)
first_correct = torch.tensor([[0.11664296, 0.88335704]])
assert torch.allclose(posterior_prob[0], first_correct)
def test_hmm_sample():
# create categorical models for the states
T0 = torch.tensor([1.0, 0.0])
T = torch.tensor([[1.0, 0.0], [0.0, 1.0]])
s1 = CategoricalModel(probs=torch.tensor([1.0, 0.0]))
s2 = CategoricalModel(probs=torch.tensor([0.0, 1.0]))
model = HiddenMarkovModel([s1, s2], T0=T0, T=T)
obs_seq, states = model.sample(3, 10)
assert isinstance(obs_seq, PackedSequence)
assert obs_seq.data.shape[0] == 30
assert isinstance(states, PackedSequence)
assert states.data.shape[0] == 30
assert 1 not in obs_seq.data
assert 1 not in states.data
T0 = torch.tensor([0.5, 0.5])
T = torch.tensor([[0.5, 0.5], [0.5, 0.5]])
s1 = CategoricalModel(probs=torch.tensor([1.0, 0.0]))
s2 = CategoricalModel(probs=torch.tensor([0.0, 1.0]))
model = HiddenMarkovModel([s1, s2], T0=T0, T=T)
obs_seq, states = model.sample(4, 20)
assert isinstance(obs_seq, PackedSequence)
assert obs_seq.data.shape[0] == 80
assert isinstance(states, PackedSequence)
assert states.data.shape[0] == 80
assert 1 in obs_seq.data and 0 in obs_seq.data
assert 1 in states.data and 0 in states.data
T0 = torch.tensor([0.9, 0.1])
T = torch.tensor([[0.9, 0.1], [0.5, 0.5]])
s1 = CategoricalModel(probs=torch.tensor([1.0, 0.0]))
s2 = CategoricalModel(probs=torch.tensor([0.0, 1.0]))
model = HiddenMarkovModel([s1, s2], T0=T0, T=T)
obs_seq, states = model.sample(1, 20)
assert isinstance(obs_seq, PackedSequence)
assert obs_seq.data.shape[0] == 20
assert isinstance(states, PackedSequence)
assert states.data.shape[0] == 20
assert (states.data == 0).sum() > (states.data == 1).sum()
def test_init():
good_T0 = torch.tensor([1.0, 0.0])
good_T = torch.tensor([[1.0, 0.0], [0.0, 1.0]])
good_E1 = CategoricalModel(probs=torch.tensor([1.0, 0.0]))
good_E2 = CategoricalModel(probs=torch.tensor([0.0, 1.0]))
bad1_T0 = torch.tensor([1.0, 1.0])
bad1_T = torch.tensor([[1.0, 1.0], [1.0, 1.0]])
bad_E1 = torch.tensor([1.0, 0.0])
bad2_T0 = torch.tensor([1.0])
bad2_T = torch.tensor([[1.0], [1.0]])
HiddenMarkovModel([good_E1, good_E2], T0=good_T0, T=good_T)
with pytest.raises(ValueError):
HiddenMarkovModel([bad_E1, good_E2], T0=bad1_T0, T=good_T)
with pytest.raises(ValueError):
HiddenMarkovModel([good_E1, good_E2], T0=bad1_T0, T=good_T)
with pytest.raises(ValueError):
HiddenMarkovModel([good_E1, good_E2], T0=good_T0, T=bad1_T)
with pytest.raises(ValueError):
HiddenMarkovModel([good_E1, good_E2], T0=bad2_T0, T=good_T)
with pytest.raises(ValueError):
HiddenMarkovModel([good_E1, good_E2], T0=good_T0, T=bad2_T)
def test_hmm_fit_viterbi():
T0 = torch.tensor([0.75, 0.25])
T = torch.tensor([[0.85, 0.15], [0.12, 0.88]])
s1_orig = torch.tensor([0.99, 0.01])
s2_orig = torch.tensor([0.05, 0.95])
s1 = CategoricalModel(probs=s1_orig)
s2 = CategoricalModel(probs=s2_orig)
model = HiddenMarkovModel([s1, s2], T0=T0, T=T)
obs_seq, states = model.sample(50, 100)
T0 = torch.tensor([0.5, 0.5])
T = torch.tensor([[0.6, 0.4], [0.5, 0.5]])
s1_orig = torch.tensor([0.6, 0.4])
s2_orig = torch.tensor([0.5, 0.5])
s1 = CategoricalModel(probs=s1_orig)
s2 = CategoricalModel(probs=s2_orig)
model = HiddenMarkovModel([s1, s2], T0=T0, T=T)
converge = model.fit(obs_seq, max_steps=500, epsilon=1e-2, restarts=3)
# Not enough samples (only 1) to test
# assert np.allclose(trans0.data.numpy(), True_pi)
print("Pi Matrix: ")
print(model.T0)
print("Transition Matrix: ")
print(model.T)
# assert np.allclose(transition.exp().data.numpy(), True_T, atol=0.1)
print()
print("Emission Matrix: ")
for s in model.states:
print([p.softmax(0) for p in s.parameters()])
# assert np.allclose(emission.exp().data.numpy(), True_E, atol=0.1)
print()
print("Reached Convergence: ")
print(converge)
assert converge
states_seq, _ = model.decode(obs_seq)
# state_summary = np.array([model.prob_state_1[i].cpu().numpy() for i in
# range(len(model.prob_state_1))])
# pred = (1 - state_summary[-2]) > 0.5
# pred = torch.cat(states_seq, 0).data.numpy()
# true = np.concatenate(states, 0)
pred = states_seq
true = states
accuracy = torch.mean(torch.abs(pred.data - true.data).float())
print("Accuracy: ", accuracy)
assert accuracy >= 0.9 or accuracy <= 0.1
def test_log_prob_aima_umbrella_example():
"""
This example was taken from AI a Modern Approach
The example comes from page 572, filtering and prediction section. The
correct values were manually computed by summing the filtered posterior.
"""
T0 = torch.tensor([0.5, 0.5])
T = torch.tensor([[0.7, 0.3], [0.3, 0.7]])
s1 = CategoricalModel(probs=torch.tensor([0.8, 0.2]))
s2 = CategoricalModel(probs=torch.tensor([0.1, 0.9]))
model = HiddenMarkovModel([s1, s2], T0=T0, T=T)
obs_seq = pack_list([torch.tensor([1])])
ll_score = model.log_prob(obs_seq).item()
assert abs(ll_score - -0.5978) < 0.001
obs_seq = pack_list([torch.tensor([1, 1])])
ll_score = model.log_prob(obs_seq).item()
assert abs(ll_score - -1.045545) < 0.001
obs_seq = pack_list([torch.tensor([1]), torch.tensor([1])])
ll_score = model.log_prob(obs_seq)
assert abs(ll_score - (2 * -0.5978)) < 0.001
X = pack_list(X)
return super().fit(X, max_steps, epsilon, randomize_first, restarts,
rand_fun, **kwargs)
if __name__ == "__main__":
T0 = torch.tensor([0.75, 0.25])
T = torch.tensor([[0.85, 0.15], [0.12, 0.88]])
s1_orig = torch.tensor([1.0, 0.0])
s2_orig = torch.tensor([0.0, 1.0])
s1 = CategoricalModel(probs=s1_orig, prior=torch.zeros_like(s1_orig))
s2 = CategoricalModel(probs=s2_orig, prior=torch.zeros_like(s2_orig))
model = HiddenMarkovModel([s1, s2],
T0=T0,
T=T,
T0_prior=torch.tensor([0., 0.]),
T_prior=torch.tensor([[0., 0.], [0., 0.]]))
obs_seq, states = model.sample(10, 10)
print("First 5 Obersvations of seq 0: ", obs_seq[0][:5])
print("First 5 Hidden States of seq 0: ", states[0][:5])
print()
T0 = torch.tensor([0.49, 0.51])
T = torch.tensor([[0.6, 0.4], [0.48, 0.52]])
s1_orig = torch.tensor([0.9, 0.1])
s2_orig = torch.tensor([0.2, 0.8])
s1 = CategoricalModel(probs=s1_orig)
s2 = CategoricalModel(probs=s2_orig)
model = HiddenMarkovModel([s1, s2], T0=T0, T=T)