def run(self, session, primers, length, temperature, hp=None):
batch_size = len(primers)
# process in segments to avoid tensorflow eating all the memory
max_segment_length = min(10000, hp.segment_length)
print "conditioning..."
segment_length = min(max_segment_length,
max(len(primer[0]) for primer in primers))
# ensure segment_length is a multiple of chunk_size
segment_length -= segment_length % hp.chunk_size
state = NS(model=self.model.initial_state(batch_size))
for segment in util.segments(primers,
segment_length,
overlap=hp.chunk_size):
x, = util.examples_as_arrays(segment)
feed_dict = {self.tensors.x: x.T}
feed_dict.update(self.model.feed_dict(state.model))
values = NS.FlatCall(
ft.partial(session.run, feed_dict=feed_dict),
self.tensors.cond.Extract("final_state.model final_xchunk"))
state.model = values.final_state.model
sys.stderr.write(".")
sys.stderr.write("\n")
cond_values = values
# make sure length is a multiple of chunk_size
chunky_length = length + hp.chunk_size - length % hp.chunk_size
print "sampling..."
length_left = chunky_length
xhats = []
state = NS(model=cond_values.final_state.model,
initial_xchunk=cond_values.final_xchunk)
while length_left > 0:
segment_length = min(max_segment_length, length_left)
length_left -= segment_length
feed_dict = {
self.tensors.initial_xchunk: state.initial_xchunk,
self.tensors.length: segment_length,
self.tensors.temperature: temperature
}
feed_dict.update(self.model.feed_dict(state.model))
sample_values = NS.FlatCall(
ft.partial(session.run, feed_dict=feed_dict),
self.tensors.sample.Extract(
"final_state.model xhat final_xhatchunk"))
state.model = sample_values.final_state.model
state.initial_xchunk = sample_values.final_xhatchunk
xhats.append(sample_values.xhat)
sys.stderr.write(".")
sys.stderr.write("\n")
xhat = np.concatenate(xhats, axis=0)
# truncate from chunky_length to the desired sample length
xhat = xhat[:length]
return xhat.T
def run(self, session, primers, length, temperature, hp=None):
batch_size = len(primers)
# process in segments to avoid tensorflow eating all the memory
max_segment_length = min(10000, hp.segment_length)
print "conditioning..."
segment_length = min(max_segment_length,
max(len(primer[0]) for primer in primers))
state = NS(model=self.model.initial_state(batch_size))
for segment in util.segments(primers, segment_length,
overlap=LEFTOVER):
x, = util.examples_as_arrays(segment)
feed_dict = {self.tensors.x: x.T}
feed_dict.update(self.model.feed_dict(state.model))
values = tfutil.run(session,
tensors=self.tensors.cond.Extract(
"final_state.model final_xelt"),
feed_dict=feed_dict)
state.model = values.final_state.model
sys.stderr.write(".")
sys.stderr.write("\n")
cond_values = values
print "sampling..."
length_left = length + LEFTOVER
xhats = []
state = NS(model=cond_values.final_state.model,
initial_xelt=cond_values.final_xelt)
while length_left > 0:
segment_length = min(max_segment_length, length_left)
length_left -= segment_length
feed_dict = {
self.tensors.initial_xelt: state.initial_xelt,
self.tensors.length: segment_length,
self.tensors.temperature: temperature
}
feed_dict.update(self.model.feed_dict(state.model))
sample_values = tfutil.run(
session,
tensors=self.tensors.sample.Extract(
"final_state.model xhat final_xhatelt"),
feed_dict=feed_dict),
state.model = sample_values.final_state.model
state.initial_xelt = sample_values.final_xhatelt
xhats.append(sample_values.xhat)
sys.stderr.write(".")
sys.stderr.write("\n")
xhat = np.concatenate(xhats, axis=0)
return xhat.T
def run(self, session, examples, max_step_count=None, hooks=None, hp=None):
tensors = self.tensors.Extract(
"loss error summaries global_step training_op learning_rate final_state.model"
)
state = NS(global_step=tf.train.global_step(session,
self.tensors.global_step),
model=self.model.initial_state(hp.batch_size))
while True:
for batch in util.batches(examples, hp.batch_size):
for segment in util.segments(batch,
self.segment_length,
overlap=LEFTOVER):
if max_step_count is not None and state.global_step >= max_step_count:
return
hooks.Get("step.before", util.noop)(state)
x, = util.examples_as_arrays(segment)
feed_dict = {self.tensors.x: x.T}
feed_dict.update(self.model.feed_dict(state.model))
values = tfutil.run(session, tensors, feed_dict=feed_dict)
state.model = values.final_state.model
state.global_step = values.global_step
hooks.Get("step.after", util.noop)(state, values)
print("step #%d loss %f error %f learning rate %e" %
(values.global_step, values.loss, values.error,
values.learning_rate))
if np.isnan(values.loss):
raise ValueError("loss has become NaN")
def run(self,
session,
examples,
max_step_count=None,
hp=None,
aggregates=None):
aggregates = NS(aggregates or {})
for key in "loss error".split():
if key not in aggregates:
aggregates[key] = util.MeanAggregate()
tensors = self.tensors.Extract(*[key for key in aggregates.Keys()])
tensors.Update(self.tensors.Extract("final_state.model"))
state = NS(step=0, model=self.model.initial_state(hp.batch_size))
try:
for batch in util.batches(examples, hp.batch_size):
for segment in util.segments(batch,
hp.segment_length,
overlap=hp.chunk_size):
if max_step_count is not None and state.step >= max_step_count:
raise StopIteration()
x, = util.examples_as_arrays(segment)
feed_dict = {self.tensors.x: x.T}
feed_dict.update(self.model.feed_dict(state.model))
values = NS.FlatCall(
ft.partial(session.run, feed_dict=feed_dict), tensors)
for key in aggregates.Keys():
aggregates[key].add(values[key])
sys.stderr.write(".")
state.model = values.final_state.model
state.step += 1
except StopIteration:
pass
sys.stderr.write("\n")
values = NS(
(key, aggregate.value) for key, aggregate in aggregates.Items())
values.summaries = [
tf.Summary.Value(tag="%s_valid" % key, simple_value=values[key])
for key in "loss error".split()
]
print "### evaluation loss %6.5f error %6.5f" % (values.loss,
values.error)
if np.isnan(values.loss):
raise ValueError("loss has become NaN")
return values
def run(self, session, examples, max_step_count=None, hooks=None, hp=None):
state = NS(global_step=tf.train.global_step(session,
self.tensors.global_step),
model=self.model.initial_state(hp.batch_size))
while True:
for batch in util.batches(examples, hp.batch_size):
for segment in util.segments(
batch,
# the last chunk is not processed, so grab
# one more to ensure we backpropagate
# through at least one full model cycle.
# TODO(cotim): rename segment_length to
# backprop_length?
hp.segment_length + hp.chunk_size,
overlap=hp.chunk_size):
if max_step_count is not None and state.global_step >= max_step_count:
return
hooks.Get("step.before", util.noop)(state)
x, = util.examples_as_arrays(segment)
feed_dict = {self.tensors.x: x.T}
feed_dict.update(self.model.feed_dict(state.model))
values = NS.FlatCall(
ft.partial(session.run, feed_dict=feed_dict),
self.tensors.Extract(
"loss error summaries global_step training_op learning_rate final_state.model"
))
state.model = values.final_state.model
state.global_step = values.global_step
hooks.Get("step.after", util.noop)(state, values)
print("step #%d loss %f error %f learning rate %e" %
(values.global_step, values.loss, values.error,
values.learning_rate))
if np.isnan(values.loss):
raise ValueError("loss has become NaN")
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