def features_for_problem(problem_instance,
p_hparams,
hparams,
data_filepatterns,
num_datashards,
mode,
batch_size=None,
name="problem_inputs"):
"""Feature map for Problem."""
with tf.name_scope(name):
with tf.device("/cpu:0"): # Input reading on CPU
capacity = (p_hparams.max_expected_batch_size_per_shard *
num_datashards)
batching_scheme = data_reader.hparams_to_batching_scheme(
hparams,
shard_multiplier=num_datashards,
drop_long_sequences=(mode == tf.estimator.ModeKeys.TRAIN
or hparams.eval_drop_long_sequences),
length_multiplier=(p_hparams.batch_size_multiplier))
if batch_size:
# If batch_size is fixed, use a single input bucket
batching_scheme["batch_sizes"] = [batch_size]
batching_scheme["boundaries"] = []
feature_map = data_reader.input_pipeline(problem_instance,
data_filepatterns,
capacity, mode, hparams,
batching_scheme)
# Reverse inputs and targets features if the problem was reversed.
if problem_instance is not None:
problem_instance.maybe_reverse_features(feature_map)
problem_instance.maybe_copy_features(feature_map)
else:
if p_hparams.was_reversed:
inputs = feature_map["inputs"]
targets = feature_map["targets"]
feature_map["inputs"] = targets
feature_map["targets"] = inputs
# Use the inputs as the targets if the problem is a copy problem.
if p_hparams.was_copy:
feature_map["targets"] = feature_map["inputs"]
# Ensure inputs and targets are proper rank.
if problem_instance.has_inputs:
while len(feature_map["inputs"].get_shape()) != 4:
feature_map["inputs"] = tf.expand_dims(feature_map["inputs"],
axis=-1)
while len(feature_map["targets"].get_shape()) != 4:
feature_map["targets"] = tf.expand_dims(feature_map["targets"],
axis=-1)
if problem_instance.has_inputs:
feature_map["input_space_id"] = tf.constant(p_hparams.input_space_id)
feature_map["target_space_id"] = tf.constant(p_hparams.target_space_id)
return feature_map
def features_for_problem(problem_instance,
p_hparams,
hparams,
data_dir,
num_datashards,
mode,
batch_size=None,
dataset_split=None,
shard=None,
name="problem_inputs"):
"""Feature map for Problem."""
with tf.name_scope(name):
with tf.device("/cpu:0"): # Input reading on CPU
capacity = (p_hparams.max_expected_batch_size_per_shard *
num_datashards)
batching_scheme = data_reader.hparams_to_batching_scheme(
hparams,
shard_multiplier=num_datashards,
drop_long_sequences=(mode == tf.estimator.ModeKeys.TRAIN
or hparams.eval_drop_long_sequences),
length_multiplier=(p_hparams.batch_size_multiplier))
if batch_size:
# If batch_size is fixed, use a single input bucket
batching_scheme["batch_sizes"] = [batch_size]
batching_scheme["boundaries"] = []
# Log new batching scheme if updated
tf.logging.info("Updated batching_scheme = %s",
batching_scheme)
feature_map = data_reader.input_pipeline(
problem_instance,
data_dir,
capacity,
mode,
hparams,
batching_scheme,
dataset_split=dataset_split,
shard=shard)
# Ensure inputs and targets are proper rank.
#feature_map["teachers"]=tf.Print(feature_map["teachers"],[feature_map["teachers"]],"#feature_map['teachers']")
if problem_instance.has_inputs:
while len(feature_map["inputs"].get_shape()) != 4:
feature_map["inputs"] = tf.expand_dims(feature_map["inputs"],
axis=-1)
while len(feature_map["targets"].get_shape()) != 4:
feature_map["targets"] = tf.expand_dims(feature_map["targets"],
axis=-1)
while len(feature_map["teachers"].get_shape()) != 4:
feature_map["teachers"] = tf.expand_dims(feature_map["teachers"],
axis=-1)
if problem_instance.has_inputs:
feature_map["input_space_id"] = tf.constant(p_hparams.input_space_id)
feature_map["target_space_id"] = tf.constant(p_hparams.target_space_id)
return feature_map
def features_for_problem(problem_instance,
p_hparams,
hparams,
data_dir,
num_datashards,
mode,
batch_size=None,
dataset_split=None,
shard=None,
name="problem_inputs"):
"""Feature map for Problem."""
with tf.name_scope(name):
with tf.device("/cpu:0"): # Input reading on CPU
capacity = (p_hparams.max_expected_batch_size_per_shard * num_datashards)
batching_scheme = data_reader.hparams_to_batching_scheme(
hparams,
shard_multiplier=num_datashards,
drop_long_sequences=(mode == tf.estimator.ModeKeys.TRAIN or
hparams.eval_drop_long_sequences),
length_multiplier=(p_hparams.batch_size_multiplier))
if batch_size:
# If batch_size is fixed, use a single input bucket
batching_scheme["batch_sizes"] = [batch_size]
batching_scheme["boundaries"] = []
# Log new batching scheme if updated
tf.logging.info("Updated batching_scheme = %s", batching_scheme)
feature_map = data_reader.input_pipeline(
problem_instance,
data_dir,
capacity,
mode,
hparams,
batching_scheme,
dataset_split=dataset_split,
shard=shard)
# Ensure inputs and targets are proper rank.
if problem_instance.has_inputs:
while len(feature_map["inputs"].get_shape()) != 4:
feature_map["inputs"] = tf.expand_dims(feature_map["inputs"], axis=-1)
while len(feature_map["targets"].get_shape()) != 4:
feature_map["targets"] = tf.expand_dims(feature_map["targets"], axis=-1)
if problem_instance.has_inputs:
feature_map["input_space_id"] = tf.constant(p_hparams.input_space_id)
feature_map["target_space_id"] = tf.constant(p_hparams.target_space_id)
return feature_map
def input_fn():
"""Supplies input to our model.
This function supplies input to our model, where this input is a
function of the mode. For example, we supply different data if
we're performing training versus evaluation.
Returns:
A tuple consisting of 1) a dictionary of tensors whose keys are
the feature names, and 2) a tensor of target labels if the mode
is not INFER (and None, otherwise).
Raises:
ValueError: if one of the parameters has an unsupported value.
"""
problem_count, batches = len(hparams.problems), []
with tf.name_scope("input_reader"):
for n in xrange(problem_count):
if fixed_problem is not None and n != fixed_problem:
continue
problem_instance = hparams.problem_instances[n]
p_hparams = hparams.problems[n]
with tf.name_scope("problem_%d" % n):
with tf.device("/cpu:0"): # Input reading on CPU
capacity = (
p_hparams.max_expected_batch_size_per_shard *
num_datashards)
feature_map = data_reader.input_pipeline(
problem_instance, data_file_patterns
and data_file_patterns[n], capacity, mode, hparams,
data_reader.hparams_to_batching_scheme(
hparams,
shard_multiplier=num_datashards,
drop_long_sequences=(
mode == tf.estimator.ModeKeys.TRAIN
or hparams.eval_drop_long_sequences),
length_multiplier=(
p_hparams.batch_size_multiplier)))
# Reverse inputs and targets features if the problem was reversed.
if problem_instance is not None:
problem_instance.maybe_reverse_features(feature_map)
problem_instance.maybe_copy_features(feature_map)
else:
if p_hparams.was_reversed:
inputs = feature_map["inputs"]
targets = feature_map["targets"]
feature_map["inputs"] = targets
feature_map["targets"] = inputs
# Use the inputs as the targets if the problem is a copy problem.
if p_hparams.was_copy:
feature_map["targets"] = feature_map["inputs"]
# Ensure inputs and targets are proper rank.
while len(feature_map["inputs"].get_shape()) != 4:
feature_map["inputs"] = tf.expand_dims(
feature_map["inputs"], axis=-1)
while len(feature_map["targets"].get_shape()) != 4:
feature_map["targets"] = tf.expand_dims(
feature_map["targets"], axis=-1)
batches.append(
(feature_map["inputs"], feature_map["targets"],
tf.constant(n), tf.constant(p_hparams.input_space_id),
tf.constant(p_hparams.target_space_id)))
# We choose which problem to process.
loss_moving_avgs = [] # Need loss moving averages for that.
for n in xrange(problem_count):
with tf.variable_scope("losses_avg"):
loss_moving_avgs.append(
tf.get_variable("problem_%d/total_loss" % n,
initializer=100.0,
trainable=False))
if fixed_problem is None:
if (hparams.problem_choice == "uniform"
or mode != tf.estimator.ModeKeys.TRAIN):
problem_choice = tf.random_uniform([],
maxval=problem_count,
dtype=tf.int32)
elif hparams.problem_choice == "adaptive":
loss_moving_avgs = tf.stack(loss_moving_avgs)
problem_choice = tf.multinomial(
tf.reshape(loss_moving_avgs, [1, -1]), 1)
problem_choice = tf.to_int32(tf.squeeze(problem_choice))
elif hparams.problem_choice == "distributed":
assert worker_replicas >= problem_count
assert worker_replicas % problem_count == 0
problem_choice = tf.to_int32(worker_id % problem_count)
else:
raise ValueError(
"Value of hparams.problem_choice is %s and must be "
"one of [uniform, adaptive, distributed]" %
hparams.problem_choice)
# Inputs and targets conditional on problem_choice.
rand_inputs, rand_target, choice, inp_id, tgt_id = cond_on_index(
lambda n: batches[n], problem_choice, 0, problem_count - 1)
else:
problem_choice = tf.constant(fixed_problem)
# Take the only constructed batch, which is the fixed_problem.
rand_inputs, rand_target, choice, inp_id, tgt_id = batches[0]
# Set shapes so the ranks are clear.
rand_inputs.set_shape([None, None, None, None])
rand_target.set_shape([None, None, None, None])
choice.set_shape([])
inp_id.set_shape([])
tgt_id.set_shape([])
# Forced shape obfuscation is necessary for inference.
if mode == tf.estimator.ModeKeys.PREDICT:
rand_inputs._shape = tf.TensorShape([None, None, None, None]) # pylint: disable=protected-access
rand_target._shape = tf.TensorShape([None, None, None, None]) # pylint: disable=protected-access
# Final feature map.
rand_feature_map = {
"inputs": rand_inputs,
"problem_choice": choice,
"input_space_id": inp_id,
"target_space_id": tgt_id
}
if mode == tf.estimator.ModeKeys.PREDICT:
rand_feature_map["infer_targets"] = rand_target
rand_target = None
# This is because of a bug in the Estimator that short-circuits prediction
# if it doesn't see a QueueRunner. DummyQueueRunner implements the
# minimal expected interface but does nothing.
tf.add_to_collection(tf.GraphKeys.QUEUE_RUNNERS,
DummyQueueRunner())
return rand_feature_map, rand_target