def test_fast_bw():
print("Make op...")
from NativeOp import FastBaumWelchOp
op = FastBaumWelchOp().make_op() # (am_scores, edges, weights, start_end_states, float_idx, state_buffer)
print("Op:", op)
n_batch = 3
seq_len = 5
n_classes = 5
from Fsa import FastBwFsaShared
fsa = FastBwFsaShared()
fsa.add_inf_loop(state_idx=0, num_emission_labels=n_classes)
fast_bw_fsa = fsa.get_fast_bw_fsa(n_batch=n_batch)
edges = fast_bw_fsa.edges.view("float32")
edges_placeholder = T.fmatrix(name="edges")
weights = fast_bw_fsa.weights
weights_placeholder = T.fvector(name="weights")
start_end_states = fast_bw_fsa.start_end_states.view("float32")
start_end_states_placeholder = T.fmatrix(name="start_end_states")
am_scores = numpy.random.normal(size=(seq_len, n_batch, n_classes)).astype("float32") # in -log space
am_scores_placeholder = T.ftensor3(name="am_scores")
float_idx = numpy.ones((seq_len, n_batch), dtype="float32")
float_idx_placeholder = T.fmatrix(name="float_idx")
last_state_idx = numpy.max(fast_bw_fsa.start_end_states[1]) # see get_automata_for_batch
state_buffer = numpy.zeros((2, last_state_idx + 1), dtype="float32")
state_buffer_placeholder = T.fmatrix(name="state_buffer")
print("Construct call...")
fwdbwd, obs_scores = op(
am_scores_placeholder, edges_placeholder, weights_placeholder, start_end_states_placeholder, float_idx_placeholder, state_buffer_placeholder)
f = theano.function(inputs=[am_scores_placeholder, edges_placeholder, weights_placeholder, start_end_states_placeholder, float_idx_placeholder, state_buffer_placeholder], outputs=[fwdbwd, obs_scores])
print("Done.")
print("Eval:")
_, score = f(am_scores, edges, weights, start_end_states, float_idx, state_buffer)
print("score:", score)
def test_FastBaumWelch():
print("Make op...")
op = make_fast_baum_welch_op(compiler_opts=dict(
verbose=True)) # will be cached, used inside :func:`fast_baum_welch`
print("Op:", op)
n_batch = 3
seq_len = 5
n_classes = 10
from Fsa import FastBwFsaShared
fsa = FastBwFsaShared()
fsa.add_inf_loop(state_idx=0, num_emission_labels=n_classes)
fast_bw_fsa = fsa.get_fast_bw_fsa(n_batch=n_batch)
edges = tf.constant(fast_bw_fsa.edges, dtype=tf.int32)
weights = tf.constant(fast_bw_fsa.weights, dtype=tf.float32)
start_end_states = tf.constant(fast_bw_fsa.start_end_states,
dtype=tf.int32)
am_scores = tf.constant(numpy.random.normal(size=(seq_len, n_batch,
n_classes)),
dtype=tf.float32) # in -log space
float_idx = tf.ones((seq_len, n_batch), dtype=tf.float32)
print("Construct call...")
fwdbwd, obs_scores = fast_baum_welch(am_scores=am_scores,
float_idx=float_idx,
edges=edges,
weights=weights,
start_end_states=start_end_states)
print("Done.")
print("Eval:")
_, score = session.run([fwdbwd, obs_scores])
print("score:", score)
def test_fast_bw_uniform():
print("Make op...")
from NativeOp import FastBaumWelchOp
op = FastBaumWelchOp().make_op() # (am_scores, edges, weights, start_end_states, float_idx, state_buffer)
print("Op:", op)
n_batch = 3
seq_len = 7
n_classes = 5
from Fsa import FastBwFsaShared
fsa = FastBwFsaShared()
for i in range(n_classes):
fsa.add_edge(i, i + 1, emission_idx=i) # fwd
fsa.add_edge(i + 1, i + 1, emission_idx=i) # loop
assert n_classes <= seq_len
fast_bw_fsa = fsa.get_fast_bw_fsa(n_batch=n_batch)
print("edges:")
print(fast_bw_fsa.edges)
edges = fast_bw_fsa.edges.view("float32")
edges_placeholder = T.fmatrix(name="edges")
weights = fast_bw_fsa.weights
weights_placeholder = T.fvector(name="weights")
print("start_end_states:")
print(fast_bw_fsa.start_end_states)
start_end_states = fast_bw_fsa.start_end_states.view("float32")
start_end_states_placeholder = T.fmatrix(name="start_end_states")
am_scores = numpy.ones((seq_len, n_batch, n_classes), dtype="float32") * numpy.float32(1.0 / n_classes)
am_scores = -numpy.log(am_scores) # in -log space
am_scores_placeholder = T.ftensor3(name="am_scores")
float_idx = numpy.ones((seq_len, n_batch), dtype="float32")
float_idx_placeholder = T.fmatrix(name="float_idx")
last_state_idx = numpy.max(fast_bw_fsa.start_end_states[1]) # see get_automata_for_batch
state_buffer = numpy.zeros((2, last_state_idx + 1), dtype="float32")
state_buffer_placeholder = T.fmatrix(name="state_buffer")
print("Construct call...")
fwdbwd, obs_scores = op(
am_scores_placeholder, edges_placeholder, weights_placeholder, start_end_states_placeholder, float_idx_placeholder, state_buffer_placeholder)
f = theano.function(inputs=[am_scores_placeholder, edges_placeholder, weights_placeholder, start_end_states_placeholder, float_idx_placeholder, state_buffer_placeholder], outputs=[fwdbwd, obs_scores])
print("Done.")
print("Eval:")
fwdbwd, score = f(am_scores, edges, weights, start_end_states, float_idx, state_buffer)
print("score:")
print(repr(score))
assert_equal(score.shape, (seq_len, n_batch))
bw = numpy.exp(-fwdbwd)
print("Baum-Welch soft alignment:")
print(repr(bw))
assert_equal(bw.shape, (seq_len, n_batch, n_classes))
from numpy import array, float32
if seq_len == n_classes:
print("Extra check identity...")
for i in range(n_batch):
assert_almost_equal(numpy.identity(n_classes), bw[:, i])
if seq_len == 7 and n_classes == 5:
print("Extra check ref_align (7,5)...")
assert_allclose(score, 8.55801582, rtol=1e-5) # should be the same everywhere
ref_align = \
array([[[1., 0., 0., 0., 0.]],
[[0.33333316, 0.66666663, 0., 0., 0.]],
[[0.06666669, 0.53333354, 0.40000018, 0., 0.]],
[[0., 0.20000014, 0.60000014, 0.19999999, 0.]],
[[0., 0., 0.39999962, 0.53333312, 0.06666663]],
[[0., 0., 0., 0.66666633, 0.33333316]],
[[0., 0., 0., 0., 0.99999982]]], dtype=float32)
assert_equal(ref_align.shape, (seq_len, 1, n_classes))
ref_align = numpy.tile(ref_align, (1, n_batch, 1))
assert_equal(ref_align.shape, bw.shape)
# print("Reference alignment:")
# print(repr(ref_align))
print("mean square diff:", numpy.mean(numpy.square(ref_align - bw)))
print("max square diff:", numpy.max(numpy.square(ref_align - bw)))
assert_allclose(ref_align, bw, rtol=1e-5)
print("Done.")
def test_fast_bw_uniform():
print("Make op...")
op = make_fast_baum_welch_op(compiler_opts=dict(
verbose=True)) # will be cached, used inside :func:`fast_baum_welch`
# args: (am_scores, edges, weights, start_end_states, float_idx, state_buffer)
print("Op:", op)
n_batch = 3
seq_len = 7
n_classes = 5
from Fsa import FastBwFsaShared
fsa = FastBwFsaShared()
for i in range(n_classes):
fsa.add_edge(i, i + 1, emission_idx=i) # fwd
fsa.add_edge(i + 1, i + 1, emission_idx=i) # loop
assert n_classes <= seq_len
fast_bw_fsa = fsa.get_fast_bw_fsa(n_batch=n_batch)
edges = tf.constant(fast_bw_fsa.edges, dtype=tf.int32)
weights = tf.constant(fast_bw_fsa.weights, dtype=tf.float32)
start_end_states = tf.constant(fast_bw_fsa.start_end_states,
dtype=tf.int32)
am_scores = numpy.ones((seq_len, n_batch, n_classes),
dtype="float32") * numpy.float32(1.0 / n_classes)
am_scores = -numpy.log(am_scores) # in -log space
am_scores = tf.constant(am_scores, dtype=tf.float32)
float_idx = tf.ones((seq_len, n_batch), dtype=tf.float32)
# from TFUtil import sequence_mask_time_major
# float_idx = tf.cast(sequence_mask_time_major(tf.convert_to_tensor(list(range(seq_len - n_batch + 1, seq_len + 1)))), dtype=tf.float32)
print("Construct call...")
fwdbwd, obs_scores = fast_baum_welch(am_scores=am_scores,
float_idx=float_idx,
edges=edges,
weights=weights,
start_end_states=start_end_states)
print("Done.")
print("Eval:")
fwdbwd, score = session.run([fwdbwd, obs_scores])
print("score:")
print(repr(score))
assert_equal(score.shape, (seq_len, n_batch))
bw = numpy.exp(-fwdbwd)
print("Baum-Welch soft alignment:")
print(repr(bw))
assert_equal(bw.shape, (seq_len, n_batch, n_classes))
from numpy import array, float32
if seq_len == n_classes:
print("Extra check identity...")
for i in range(n_batch):
assert_almost_equal(numpy.identity(n_classes), bw[:, i])
if seq_len == 7 and n_classes == 5:
print("Extra check ref_align (7,5)...")
assert_allclose(score, 8.55801582,
rtol=1e-5) # should be the same everywhere
ref_align = \
array([[[1., 0., 0., 0., 0.]],
[[0.33333316, 0.66666663, 0., 0., 0.]],
[[0.06666669, 0.53333354, 0.40000018, 0., 0.]],
[[0., 0.20000014, 0.60000014, 0.19999999, 0.]],
[[0., 0., 0.39999962, 0.53333312, 0.06666663]],
[[0., 0., 0., 0.66666633, 0.33333316]],
[[0., 0., 0., 0., 0.99999982]]], dtype=float32)
assert_equal(ref_align.shape, (seq_len, 1, n_classes))
ref_align = numpy.tile(ref_align, (1, n_batch, 1))
assert_equal(ref_align.shape, bw.shape)
# print("Reference alignment:")
# print(repr(ref_align))
print("mean square diff:", numpy.mean(numpy.square(ref_align - bw)))
print("max square diff:", numpy.max(numpy.square(ref_align - bw)))
assert_allclose(ref_align, bw, rtol=1e-5)
print("Done.")