def test_backward_pass(T=1000, K=5, D=2):
from pyhsmm.internals.hmm_messages_interface import messages_backwards_log
# Make parameters
As = npr.rand(K, K)
As /= As.sum(axis=-1, keepdims=True)
ll = npr.randn(T, K)
# Use pyhsmm to compute
true_betas = np.zeros((T, K))
messages_backwards_log(As, ll, true_betas)
# Use ssm to compute
test_betas = np.zeros((T, K))
backward_pass(As[None, :, :], ll, test_betas)
assert np.allclose(true_betas, test_betas)
def test_hmm_mp_perf(T=10000, K=100, D=20):
# Make parameters
pi0 = np.ones(K) / K
Ps = npr.rand(T-1, K, K)
Ps /= Ps.sum(axis=2, keepdims=True)
ll = npr.randn(T, K)
out1 = np.zeros((T, K))
out2 = np.zeros((T, K))
# Run the PyHSMM message passing code
from pyhsmm.internals.hmm_messages_interface import messages_forwards_log, messages_backwards_log
tic = time()
messages_forwards_log(Ps, ll, pi0, out1)
pyhsmm_dt = time() - tic
print("PyHSMM Fwd: ", pyhsmm_dt, "sec")
# Run the SSM message passing code
from ssm.messages import forward_pass, backward_pass
forward_pass(pi0, Ps, ll, out2) # Call once to compile, then time it
tic = time()
forward_pass(pi0, Ps, ll, out2)
smm_dt = time() - tic
print("SMM Fwd: ", smm_dt, "sec")
assert np.allclose(out1, out2)
# Backward pass
tic = time()
messages_backwards_log(Ps, ll, out1)
pyhsmm_dt = time() - tic
print("PyHSMM Bwd: ", pyhsmm_dt, "sec")
backward_pass(Ps, ll, out2) # Call once to compile, then time it
tic = time()
backward_pass(Ps, ll, out2)
smm_dt = time() - tic
print("SMM (Numba) Bwd: ", smm_dt, "sec")
assert np.allclose(out1, out2)
def hmm_estep(natparam):
C = lambda x: np.require(x, np.double, 'C')
init_params, pair_params, node_params = list(map(C, natparam))
# compute messages
alphal = messages_forwards_log(np.exp(pair_params), node_params,
np.exp(init_params),
np.zeros_like(node_params))
betal = messages_backwards_log(np.exp(pair_params), node_params,
np.zeros_like(node_params))
# compute expected statistics from messages
expected_states, expected_transcounts, log_normalizer = \
expected_statistics_log(
pair_params, node_params, alphal, betal,
np.zeros_like(node_params), np.zeros_like(pair_params))
expected_stats = expected_states[0], expected_transcounts, expected_states
return log_normalizer, expected_stats
def _messages_backwards_log(trans_matrix,log_likelihoods):
from pyhsmm.internals.hmm_messages_interface import messages_backwards_log
return messages_backwards_log(
trans_matrix,log_likelihoods,
np.empty_like(log_likelihoods))
