def state_placeholders(self):
"""Get the Tensorflow placeholders for the model's states.
Returns:
A Namespace tree containing the placeholders.
"""
return NS.Copy(self._state_placeholders)
def __call__(self, x, state, context=None):
# construct the usual graph without unrolling
state = NS.Copy(state)
state.cells = Wayback.transition(state.time,
state.cells,
self.cells,
below=x,
above=context,
hp=self.hp,
symbolic=True)
state.time += 1
state.time %= self.period
return state
def make_transition_graph(state,
transition,
x=None,
context=None,
temperature=1.0,
hp=None):
"""Make the graph that processes a single sequence element.
Args:
state: `_make_sequence_graph` loop state.
transition: Model transition function mapping (xelt, model_state,
context) to (output, new_model_state).
x: Sequence of integer (categorical) inputs. Axes [time, batch].
context: Optional Tensor denoting context, shaped [batch, ?].
temperature: Softmax temperature to use for sampling.
hp: Model hyperparameters.
Returns:
Updated loop state.
"""
state = NS.Copy(state)
xelt = tfutil.shaped_one_hot(
state.xhats.read(state.i) if x is None else x[state.i, :],
[None, hp.data_dim])
embedding = tfutil.layers([xelt], sizes=hp.io_sizes, use_bn=hp.use_bn)
h, state.model = transition(embedding, state.model, context=context)
# predict the next elt
with tf.variable_scope("xhat") as scope:
embedding = tfutil.layers([h], sizes=hp.io_sizes, use_bn=hp.use_bn)
exhat = tfutil.project(embedding, output_dim=hp.data_dim)
xhat = tfutil.sample(exhat, temperature)
state.xhats = state.xhats.write(state.i + LEFTOVER, xhat)
if x is not None:
target = tfutil.shaped_one_hot(x[state.i + 1], [None, hp.data_dim])
state.losses = state.losses.write(
state.i, tf.nn.softmax_cross_entropy_with_logits(exhat, target))
state.errors = state.errors.write(
state.i,
tf.not_equal(tf.nn.top_k(exhat)[1],
tf.nn.top_k(target)[1]))
state.exhats = state.exhats.write(state.i, exhat)
state.i += 1
return state
def __call__(self, x, state, context=None):
state = NS.Copy(state)
for i, _ in enumerate(self.cells):
cell_inputs = []
if i == 0:
cell_inputs.append(x)
if context is not None and i == len(self.cells) - 1:
cell_inputs.append(context)
if self.hp.vskip:
# feed in state of all other layers
cell_inputs.extend(self.cells[j].get_output(state.cells[j])
for j in range(len(self.cells)) if j != i)
else:
# feed in state of layer below
if i > 0:
cell_inputs.append(self.cells[i - 1].get_output(
state.cells[i - 1]))
state.cells[i] = self.cells[i].transition(cell_inputs,
state.cells[i],
scope="cell%i" % i)
return state
def _make_sequence_graph_with_unroll(self,
model_state=None,
x=None,
initial_xelt=None,
context=None,
length=None,
temperature=1.0,
hp=None,
back_prop=False):
"""Create a sequence graph by unrolling upper layers.
This method is similar to `_make_sequence_graph`, except that `length` must be provided. The
resulting graph behaves in the same way as that constructed by `_make_sequence_graph`, except
that the upper layers are outside of the while loop and so the gradient can actually be
truncated between runs of lower layers.
If `x` is given, the graph processes the sequence `x` one element at a time. At step `i`, the
model receives the `i`th element as input, and its output is used to predict the `i + 1`th
element.
The last element is not processed, as there would be no further element available to compare
against and compute loss. To ensure all data is processed during TBPTT, segments `x` fed into
successive computations of the graph should overlap by 1.
If `x` is not given, `initial_xelt` must be given as the first input to the model. Further
elements are constructed from the model's predictions.
Args:
model_state: initial state of the model.
x: Sequence of integer (categorical) inputs. Not needed if sampling.
Axes [time, batch].
initial_xelt: When sampling, x is not given; initial_xelt specifies
the input x[0] to the first timestep.
context: a `Tensor` denoting context, e.g. for conditioning.
Axes [batch, features].
length: Optional length of sequence. Inferred from `x` if possible.
temperature: Softmax temperature to use for sampling.
hp: Model hyperparameters.
back_prop: Whether the graph will be backpropagated through.
Raises:
ValueError: if `length` is not an int.
Returns:
Namespace containing relevant symbolic variables.
"""
if length is None or not isinstance(length, int):
raise ValueError(
"For partial unrolling, length must be known at graph construction time."
)
if model_state is None:
model_state = self.state_placeholders()
state = NS(model=model_state,
inner_initial_xelt=initial_xelt,
xhats=[],
losses=[],
errors=[])
# i suspect ugly gradient biases may occur if gradients are truncated
# somewhere halfway through the cycle. ensure we start at a cycle boundary.
state.model.time = tfutil.assertion(state.model.time,
tf.equal(state.model.time, 0),
[state.model.time],
name="outer_alignment_assertion")
# ensure we end at a cycle boundary too.
assert (length - LEFTOVER) % self.period == 0
inner_period = int(np.prod(hp.periods[:self.outer_indices[0] + 1]))
# hp.boundaries specifies truncation boundaries relative to the end of the sequence and in terms
# of each layer's own steps; translate this to be relative to the beginning of the sequence and
# in terms of sequence elements. note that due to the dynamic unrolling of the inner graph, the
# inner layers necessarily get truncated at the topmost inner layer's boundary.
boundaries = [
length - 1 - hp.boundaries[i] * int(np.prod(hp.periods[:i + 1]))
for i in range(len(hp.periods))
]
assert all(0 <= boundary and boundary < length - LEFTOVER
for boundary in boundaries)
assert boundaries == list(reversed(sorted(boundaries)))
print "length %s periods %s boundaries %s %s inner period %s" % (
length, hp.periods, hp.boundaries, boundaries, inner_period)
outer_step_count = length // inner_period
for outer_time in range(outer_step_count):
if outer_time > 0:
tf.get_variable_scope().reuse_variables()
# update outer layers (wrap in seq scope to be consistent with the fully
# symbolic version of this graph)
with tf.variable_scope("seq"):
# truncate gradient (only effective on outer layers)
for i in range(len(self.cells)):
if outer_time * inner_period <= boundaries[i]:
state.model.cells[i] = list(
map(tf.stop_gradient, state.model.cells[i]))
state.model.cells = Wayback.transition(
outer_time * inner_period,
state.model.cells,
self.cells,
below=None,
above=context,
subset=self.outer_indices,
hp=hp,
symbolic=False)
# run inner layers on subsequence
if x is None:
inner_x = None
else:
start = inner_period * outer_time
stop = inner_period * (outer_time + 1) + LEFTOVER
inner_x = x[start:stop, :]
# grab a copy of the outer states. they will not be updated in the inner
# loop, so we can put back the copy after the inner loop completes.
# this avoids the gradient truncation due to calling `while_loop` with
# `back_prop=False`.
outer_cell_states = NS.Copy(state.model.cells[self.outer_slice])
def _inner_transition(input_, state, context=None):
assert not context
state.cells = Wayback.transition(state.time,
state.cells,
self.cells,
below=input_,
above=None,
subset=self.inner_indices,
hp=hp,
symbolic=True)
state.time += 1
state.time %= self.period
h = self.get_output(state)
return h, state
inner_back_prop = back_prop and outer_time * inner_period >= boundaries[
self.inner_indices[-1]]
inner_ts = _make_sequence_graph(
transition=_inner_transition,
model_state=state.model,
x=inner_x,
initial_xelt=state.inner_initial_xelt,
temperature=temperature,
hp=hp,
back_prop=inner_back_prop)
state.model = inner_ts.final_state.model
state.inner_initial_xelt = inner_ts.final_xelt if x is not None else inner_ts.final_xhatelt
state.final_xhatelt = inner_ts.final_xhatelt
if x is not None:
state.final_x = inner_x
state.final_xelt = inner_ts.final_xelt
# track only losses and errors after the boundary to avoid bypassing the truncation boundary.
if inner_back_prop:
state.losses.append(inner_ts.loss)
state.errors.append(inner_ts.error)
state.xhats.append(inner_ts.xhat)
# restore static outer states
state.model.cells[self.outer_slice] = outer_cell_states
# double check alignment to be safe
state.model.time = tfutil.assertion(
state.model.time,
tf.equal(state.model.time % inner_period, 0),
[state.model.time, tf.shape(inner_x)],
name="inner_alignment_assertion")
ts = NS()
ts.xhat = tf.concat(0, state.xhats)
ts.final_xhatelt = state.final_xhatelt
ts.final_state = state
if x is not None:
ts.final_x = state.final_x
ts.final_xelt = state.final_xelt
# inner means are all on the same sample size, so taking their mean is valid
ts.loss = tf.reduce_mean(state.losses)
ts.error = tf.reduce_mean(state.errors)
return ts
def testCopy(self):
before = NS(v=2, w=NS(x=1, y=NS(z=0)))
after = NS.Copy(before)
self.assertEqual(before, after)
self.assertTrue(
all(a is b for a, b in zip(NS.Flatten(after), NS.Flatten(before))))
def make_transition_graph(state,
transition,
x=None,
context=None,
temperature=1.0,
hp=None):
"""Make the graph that processes a single sequence element.
Args:
state: `_make_sequence_graph` loop state.
transition: Model transition function mapping (xchunk, model_state,
context) to (output, new_model_state).
x: Sequence of integer (categorical) inputs. Axes [time, batch].
context: Optional Tensor denoting context, shaped [batch, ?].
temperature: Softmax temperature to use for sampling.
hp: Model hyperparameters.
Returns:
Updated loop state.
"""
state = NS.Copy(state)
xchunk = _get_flat_chunk(state.xhats if x is None else x,
state.i * hp.chunk_size,
hp.chunk_size,
depth=hp.data_dim)
embedding = tfutil.layers([xchunk], sizes=hp.io_sizes, use_bn=hp.use_bn)
h, state.model = transition(embedding, state.model, context=context)
# predict the next chunk
exhats = []
with tf.variable_scope("xhat") as scope:
for j in range(hp.chunk_size):
if j > 0:
scope.reuse_variables()
xchunk = _get_flat_chunk(state.xhats if x is None else x,
state.i * hp.chunk_size + j,
hp.chunk_size,
depth=hp.data_dim)
embedding = tfutil.layers([h, xchunk],
sizes=hp.io_sizes,
use_bn=hp.use_bn)
exhat = tfutil.project(embedding, output_dim=hp.data_dim)
exhats.append(exhat)
state.xhats = state.xhats.write((state.i + 1) * hp.chunk_size + j,
tfutil.sample(exhat, temperature))
if x is not None:
targets = tf.unpack(_get_1hot_chunk(x, (state.i + 1) * hp.chunk_size,
hp.chunk_size,
depth=hp.data_dim),
num=hp.chunk_size,
axis=1)
state.losses = _put_chunk(state.losses, state.i * hp.chunk_size, [
tf.nn.softmax_cross_entropy_with_logits(exhat, target)
for exhat, target in util.equizip(exhats, targets)
])
state.errors = _put_chunk(state.errors, state.i * hp.chunk_size, [
tf.not_equal(tf.nn.top_k(exhat)[1],
tf.nn.top_k(target)[1])
for exhat, target in util.equizip(exhats, targets)
])
state.exhats = _put_chunk(state.exhats, state.i * hp.chunk_size,
exhats)
state.i += 1
return state