def non_batched_dataset(train_dev_or_test,
label_vocab,
data_root_dir=DATA_ROOT_DIR):
"""Constructs a dataset of examples.
Args:
train_dev_or_test: one of _DEV_FOLD_VALUES. The source examples to load into
a dataset.
label_vocab: list of string.
data_root_dir: path to tfrecord examples.
Returns:
tf.data.Dataset, where each example is of form
{
SEQUENCE_KEY: one-hot of amino acid characters
SEQUENCE_LENGTH_KEY: length of sequence
SEQUENCE_ID_KEY: unique identifier for protein
LABEL_KEY: rank-1 tensor of integer labels from label_vocab,
}
"""
if train_dev_or_test not in DATA_FOLD_VALUES:
raise ValueError(('Only train, dev, test and * are supported datasets.'
' Received {}.').format(train_dev_or_test))
dataset_files = [
os.path.join(data_root_dir, f)
for f in tf.gfile.ListDirectory(data_root_dir)
if train_dev_or_test in f and ".tfrecord" in f
]
tfrecord_dataset = tf.data.TFRecordDataset(dataset_files)
dataset = tfrecord_dataset.map(lambda record: tf.io.parse_single_example( # pylint: disable=g-long-lambda
record, DATASET_FEATURES))
dataset = dataset.map(_add_sequence_length)
dataset = dataset.filter(_is_sequence_short_enough_for_training)
amino_acid_table = contrib_lookup.index_table_from_tensor(
utils.AMINO_ACID_VOCABULARY,
default_value=len(utils.AMINO_ACID_VOCABULARY))
protein_class_table = contrib_lookup.index_table_from_tensor(
mapping=label_vocab)
dataset = dataset.map(
lambda ex: _map_sequence_to_ints(ex, amino_acid_table))
dataset = dataset.map(
lambda ex: _map_labels_to_ints(ex, protein_class_table))
dataset = dataset.map(_to_one_hot_sequence)
if train_dev_or_test == TRAIN_FOLD:
dataset = dataset.repeat()
dataset = dataset.prefetch(tf.data.experimental.AUTOTUNE)
return dataset
def embeddings(features, configs, embedding_size=None):
assert (embedding_size is not None) and isinstance(embedding_size, int)
embedded = {}
with tf.variable_scope('embeddings'):
for config in configs:
embeddings = tf.get_variable(
name=config.name,
shape=[config.length, embedding_size],
dtype=tf.float32,
trainable=True,
initializer=tf.random_uniform_initializer(minval=-0.5,
maxval=0.5,
dtype=tf.float32),
)
if config.table is not None:
lookup_table = lookup.index_table_from_tensor(
mapping=tf.constant(config.table, dtype=tf.string),
default_value=0)
for field in config.fields:
embedded[field] = tf.nn.embedding_lookup(
embeddings, lookup_table.lookup(features[field]))
else:
for field in config.fields:
embedded[field] = tf.nn.embedding_lookup(
embeddings, features[field])
return embedded
def create_vocab_lookup_tables(vocab):
str_to_int = lookup.index_table_from_tensor(
mapping=vocab,
num_oov_buckets=0,
default_value=OOV_TOKEN_ID,
name='vocab_lookup_str_to_int'
)
int_to_str = lookup.index_to_string_table_from_tensor(
mapping=vocab,
default_value=UNKNOWN_TOKEN,
name='vocab_lookup_int_to_str'
)
word2id = {w: i for i, w in enumerate(vocab)}
vocab_lookup = {
INT_TO_STR: int_to_str,
STR_TO_INT: str_to_int,
RAW_WORD2ID: word2id,
RAW_ID2WORD: vocab
}
graph_utils.add_dict_to_collection(vocab_lookup, VOCAB_LOOKUP_COLL_NAME)
return vocab_lookup
def test_table_roundtrip(self):
export_path = os.path.join(tempfile.mkdtemp(), 'export')
with tf.Graph().as_default():
with tf.Session().as_default() as session:
input_string = tf.placeholder(tf.string)
# Map string through a table, in this case based on a constant tensor.
table = lookup.index_table_from_tensor(
tf.constant(['cat', 'dog', 'giraffe']))
output = table.lookup(input_string)
inputs = {'input': input_string}
outputs = {'output': output}
saved_transform_io.write_saved_transform_from_session(
session, inputs, outputs, export_path)
with tf.Graph().as_default():
with tf.Session().as_default() as session:
# Using a computed input gives confidence that the graphs are fused.
input_string = tf.constant('dog')
inputs = {'input': input_string}
outputs = saved_transform_io.apply_saved_transform(
export_path, inputs)
session.run(tf.tables_initializer())
result = session.run(outputs['output'])
self.assertEqual(1, result)
def __init__(self,
speaker_list,
file_pattern,
T=2**12,
batch_size=1,
num_epoch=None,
buffer_size=4000):
''' `T`: sequence length '''
with tf.device('/cpu'):
with tf.name_scope('ByteInputPipeline'):
self.speaker_list = tf.constant(speaker_list)
self.table = index_table_from_tensor(mapping=self.speaker_list)
self.T = T
filenames = tf.gfile.Glob(file_pattern)
if filenames:
print('Data Loader: {} files found\n'.format(
len(filenames)))
else:
raise ValueError('No files found: {}'.format(file_pattern))
dataset = (tf.data.TFRecordDataset(filenames).map(
self._parse_function).shuffle(buffer_size).batch(
batch_size).repeat(num_epoch))
self.iterator = dataset.make_initializable_iterator()
self.x, self.y = self.iterator.get_next()
def testCreatePhasesWithUnwrappedTable(self):
# Create a graph with a table that is not wrapped in `apply_function`.
string_placeholder = tf.placeholder(tf.string, shape=(None, ))
table = lookup.index_table_from_tensor(['a', 'b'])
table.lookup(string_placeholder)
with self.assertRaisesRegexp(ValueError, 'Found table initializers'):
impl_helper.create_phases()
def make_item2id(self, seq_list):
count_dict = collections.Counter(
reduce(lambda seq1, seq2: seq1 + seq2, seq_list))
item_mapper = list(
map(lambda i: i[0] if i[1] >= self.min_count else '0',
count_dict.items()))
item2id = index_table_from_tensor(mapping=item_mapper, default_value=0)
self.vocab_size = item2id.size()
return item2id
def add_vocab_lookups(self):
with open(self.config.filename_words) as f:
words = [word.strip() for idx, word in enumerate(f)]
with open(self.config.filename_tags) as f:
labels = [label.strip() for idx, label in enumerate(f)]
with open(self.config.filename_chars) as f:
chars = [char.strip() for idx, char in enumerate(f)]
self.label_list = tf.constant(labels)
self.word_table = lookup.index_table_from_tensor(
mapping=words, default_value=words.index(UNK))
self.char_table = lookup.index_table_from_tensor(mapping=chars,
default_value=-1)
self.label_table = lookup.index_table_from_tensor(
mapping=self.label_list, num_oov_buckets=1)
def fasta_indexer():
"""Get a function for converting tokenized protein strings to indices."""
mapping = tf.constant(FULL_RESIDUE_VOCAB)
table = contrib_lookup.index_table_from_tensor(mapping)
def mapper(residues):
return tf.ragged.map_flat_values(table.lookup, residues)
return mapper
def __init__(self,
split_name,
preprocess_fn,
num_epochs,
shuffle,
random_seed=None,
filter_filename=None,
drop_remainder=True,
sup=False):
"""Initialize the dataset object.
Args:
split_name: A string split name, to load from the dataset.
preprocess_fn: Preprocess a single example. The example is already
parsed into a dictionary.
num_epochs: An int, defaults to `None`. Number of epochs to cycle
through the dataset before stopping. If set to `None` this will read
samples indefinitely.
shuffle: A boolean, defaults to `False`. Whether output data are
shuffled.
random_seed: Optional int. Random seed for shuffle operation.
filter_filename: Optional filename to use for filtering.
drop_remainder: If true, then the last incomplete batch is dropped.
"""
# This is an instance-variable instead of a class-variable because it
# depends on FLAGS, which is not parsed yet at class-parse-time.
dataset_dir = FLAGS.sup_dataset_dir if sup else FLAGS.unsup_dataset_dir
files = os.path.join(os.path.expanduser(dataset_dir), '%s@%i')
filenames = {
'train': generate_sharded_filenames(files % ('train', 1024))[:-40],
'val': generate_sharded_filenames(files % ('train', 1024))[-40:],
'trainval': generate_sharded_filenames(files % ('train', 1024)),
'test': generate_sharded_filenames(files % ('validation', 128))
}
super(DatasetWalmartFashion,
self).__init__(filenames=filenames[split_name],
reader=tf.data.TFRecordDataset,
num_epochs=num_epochs,
shuffle=shuffle,
random_seed=random_seed,
filter_fn=self.get_filter()
if filter_filename is not None else None,
drop_remainder=drop_remainder)
self.split_name = split_name
self.preprocess_fn = preprocess_fn
self.filename_list = None
if filter_filename is not None:
with tf.gfile.Open(filter_filename, 'r') as f:
filename_list = json.load(f)
filename_list = tf.constant(filename_list['values'])
filename_list = index_table_from_tensor(mapping=filename_list,
num_oov_buckets=0,
default_value=-1)
self.filename_list = filename_list
def _map_to_int(x):
"""Maps string tensor into indexes using vocab.
Args:
x : a Tensor/SparseTensor of string.
Returns:
a Tensor/SparseTensor of indexes (int) of the same shape as x.
"""
table = lookup.index_table_from_tensor(vocab, default_value=len(vocab))
return table.lookup(x)
def get_vocab_lookup(vocab, name=None, reuse=None):
with tf.variable_scope(name, 'vocab_lookup', reuse=reuse):
vocab_lookup = lookup.index_table_from_tensor(
mapping=vocab,
num_oov_buckets=0,
default_value=OOV_TOKEN_ID,
name=name
)
return vocab_lookup
def convert_label(label):
"""Parses a string tensor into the label tensor
Args:
label_string_tensor: Tensor of dtype string. Result of parsing the
CSV column specified by LABEL_COLUMN
Returns:
A Tensor of the same shape as label_string_tensor, should return
an int64 Tensor representing the label index for classification tasks
"""
table = lookup.index_table_from_tensor(['<=50K', '>50K'])
return table.lookup(label)
def get_word_embeddings(word_tensor, embedding_words, embedding_vectors):
"""Convert a string tensor of words into a float32 tensor of word embeddings."""
word_lookup = lookup.index_table_from_tensor(embedding_words, default_value=0)
word_ids = word_lookup.lookup(word_tensor)
word_embeddings = tf.concat((
tf.get_variable(name="unknown_word_embedding", initializer=embedding_vectors[:1], trainable=True),
tf.get_variable(name="known_word_embeddings", initializer=embedding_vectors[1:], trainable=False)
), axis=0)
return tf.nn.embedding_lookup(word_embeddings, word_ids)
def convert_label(label):
"""Parses a string tensor into the label tensor
Args:
label_string_tensor: Tensor of dtype string. Result of parsing the
CSV column specified by LABEL_COLUMN
Returns:
A Tensor of the same shape as label_string_tensor, should return
an int64 Tensor representing the label index for classification tasks
"""
table = lookup.index_table_from_tensor(['<=50K', '>50K'])
return table.lookup(label)
def _map_to_int(x):
"""Maps string tensor into indexes using vocab.
Args:
x : a Tensor/SparseTensor of string.
Returns:
a Tensor/SparseTensor of indexes (int) of the same shape as x.
"""
table = lookup.index_table_from_tensor(
vocab,
default_value=len(vocab))
return table.lookup(x)
def string2index(feature_strings, feature):
"""
Convert a `Tensor` of type `tf.string` to a corresponding Tensor of ids (`tf.int32`)
:param feature_strings: string `Tensor`
:param feature: feature extractor with string to index vocabulary
:return: feature id Tensor
"""
with variable_scope('lookup'):
feats = list(feature.ordered_feats())
lookup = index_table_from_tensor(mapping=tf.constant(feats),
default_value=feature.unk_index())
return lookup.lookup(feature_strings)
def __init__(self, speaker_list, filenames, num_epoch=1):
with tf.device('/cpu'):
with tf.name_scope('ByteInputPipeline'):
self.speaker_list = tf.constant(speaker_list)
self.table = index_table_from_tensor(mapping=self.speaker_list)
print('{} files found'.format(len(filenames)))
dataset = (tf.data.TFRecordDataset(filenames).map(
self._parse_function).batch(1).repeat(num_epoch))
self.iterator = dataset.make_initializable_iterator()
self.x, self.y, self.f, self.w, self.t = self.iterator.get_next(
)
def build_tensorize_text_fn(embeddings):
"""Builds a function to turn text into word/char ids."""
tbl = contrib_lookup.index_table_from_tensor(
mapping=embeddings.get_vocab(), num_oov_buckets=1)
def fn(string_tensor):
"""Builds the output tensor dictionary."""
out = {}
if FLAGS.lowercase:
string_tensor = ops.lowercase_op(string_tensor)
out["wids"] = tf.to_int32(tbl.lookup(string_tensor))
out["cids"] = char_utils.batch_word_to_char_ids(string_tensor, 50)
out["len"] = tf.shape(string_tensor)[-1]
return out
return fn
def add_id_lookups(self):
table = lookup.index_table_from_tensor(mapping=tf.constant(['']),
default_value=1)
sentences_shape = tf.shape(self.padded_sentences, out_type=tf.int64)
removed_char_sentences = remove_unknown_chars(self.padded_sentences,
self.char_table)
split_words = tf.string_split(tf.reshape(removed_char_sentences, [-1]),
delimiter="")
dense_split_words = tf.sparse_tensor_to_dense(split_words,
default_value='')
max_word_len = tf.gather_nd(split_words.dense_shape, tf.constant([1]))
chars_shape = tf.concat([sentences_shape, [max_word_len]], 0)
chars = tf.reshape(dense_split_words, chars_shape)
self.word_lengths = tf.reduce_sum(table.lookup(chars), 2)
lowercase_sentences = lowercase(self.padded_sentences)
sanitised_sentences = tf.regex_replace(lowercase_sentences, '^[0-9]+$',
NUM)
self.sequence_lengths = tf.reduce_sum(
table.lookup(sanitised_sentences), 1)
self.word_ids = self.word_table.lookup(sanitised_sentences)
self.char_ids = self.char_table.lookup(chars)
word_mask = tf.sequence_mask(self.sequence_lengths)
char_mask = tf.sequence_mask(self.word_lengths)
self.word_ids = tf.where(word_mask, self.word_ids,
tf.zeros_like(self.word_ids))
self.char_ids = tf.where(char_mask, self.char_ids,
tf.zeros_like(self.char_ids))
label_lengths = tf.reduce_sum(table.lookup(self.label_codes), 1)
labels_mask = tf.sequence_mask(label_lengths)
self.labels = self.label_table.lookup(self.label_codes)
self.labels = tf.where(labels_mask, self.labels,
tf.zeros_like(self.labels))
def provide_dataset(self):
"""Provides dataset (audio, labels) of nsynth."""
length = 64000
channels = 1
pitch_counts = self.get_pitch_counts()
pitches = sorted(pitch_counts.keys())
label_index_table = contrib_lookup.index_table_from_tensor(
sorted(pitches), dtype=tf.int64)
def _parse_nsynth(record):
"""Parsing function for NSynth dataset."""
features = {
'pitch': tf.FixedLenFeature([1], dtype=tf.int64),
'audio': tf.FixedLenFeature([length], dtype=tf.float32),
'qualities': tf.FixedLenFeature([10], dtype=tf.int64),
'instrument_source': tf.FixedLenFeature([1], dtype=tf.int64),
'instrument_family': tf.FixedLenFeature([1], dtype=tf.int64),
}
example = tf.parse_single_example(record, features)
wave, label = example['audio'], example['pitch']
wave = spectral_ops.crop_or_pad(wave[tf.newaxis, :, tf.newaxis],
length,
channels)[0]
one_hot_label = tf.one_hot(
label_index_table.lookup(label), depth=len(pitches))[0]
return wave, one_hot_label, label, example['instrument_source']
dataset = self._get_dataset_from_path()
dataset = dataset.map(_parse_nsynth, num_parallel_calls=4)
# Filter just specified instrument sources
def _is_wanted_source(s):
return tf.reduce_any(list(map(lambda q: tf.equal(s, q)[0], self._instrument_sources)))
dataset = dataset.filter(lambda w, l, p, s: _is_wanted_source(s))
# Filter just specified pitches
dataset = dataset.filter(lambda w, l, p, s: tf.greater_equal(p, self._min_pitch)[0])
dataset = dataset.filter(lambda w, l, p, s: tf.less_equal(p, self._max_pitch)[0])
dataset = dataset.map(lambda w, l, p, s: (w, l))
return dataset
def string_to_int_mapper(keys_to_map,
mapping,
num_oov_buckets=1,
suffix="_id"):
"""Creates a mapping function to convert strings to ints in a tf.data.Dataset.
For `dataset` outputs of type `str`, uses the list of strings in the given
input `mapping` to look up the strings using tf.contrib.lookup and convert
them to same-shape tensors of size tf.int32.
Example:
vocab = ['the', 'fox', 'jumped']
dataset = dataset.map(string_to_int_mapper(['words'], mapping=vocab))
dataset['words_id'] # <-- 'the' is mapped to 0, 'fox' to 1, etc...
Args:
keys_to_map: List of strings that are keys for tf.string Tensors to lookup.
mapping: List of strings (or string tensors) to do the lookup. If the
mapping is already a lookup table, then we directly use it.
num_oov_buckets: Number of OOV buckets to use (default = 1).
suffix: String to append to the given keys to indicate the mapped Tensors.
Returns:
_mapper: A mapping function that can be used with the tf.data.Dataset API.
"""
if isinstance(mapping, LookupInterface):
table = mapping
else:
table = contrib_lookup.index_table_from_tensor(
mapping=mapping, num_oov_buckets=num_oov_buckets)
def _mapper(dataset):
for k in keys_to_map:
dataset[k + suffix] = tf.to_int32(table.lookup(dataset[k]))
return dataset
return _mapper
def make_iterator_from_text_dataset(text_dataset,
batch_size,
unit_dict,
shuffle=False,
bucket_width=-1,
num_cores=4):
from tensorflow.contrib.lookup import index_table_from_tensor
table = index_table_from_tensor(mapping=list(unit_dict.values()))
dataset = tf.data.TextLineDataset(text_dataset)
dataset = dataset.map(
lambda str: tf.string_split([str], delimiter='').values)
dataset = dataset.map(lambda chars: (chars, tf.size(chars)))
dataset = dataset.map(lambda chars, size: (table.lookup(chars), size))
if shuffle is True:
dataset = dataset.shuffle(buffer_size=1000000,
reshuffle_each_iteration=True)
def batching_fun(x):
labels_shape = (
tf.TensorShape([None]),
tf.TensorShape([]),
)
return x.padded_batch(batch_size=batch_size,
padded_shapes=(labels_shape))
if bucket_width == -1:
dataset = batching_fun(dataset)
else:
def key_func(labels, labels_len):
# labels_len = tf.shape(labels)[0]
bucket_id = labels_len // bucket_width
return tf.cast(bucket_id, dtype=tf.int64)
def reduce_func(unused_key, windowed_dataset):
return batching_fun(windowed_dataset)
dataset = tf.data.Dataset.apply(
dataset,
tf.data.experimental.group_by_window(key_func=key_func,
reduce_func=reduce_func,
window_size=batch_size))
dataset = dataset.prefetch(128)
iterator = dataset.make_initializable_iterator()
labels, labels_len = iterator.get_next()
return BatchedData(iterator_initializer=iterator.initializer,
inputs_filenames=None,
labels_filenames=None,
inputs=None,
payload=None,
inputs_length=None,
labels=labels,
labels_length=labels_len)
def decode_target(target_string):
table = lookup.index_table_from_tensor(tf.constant(commons.TARGET_LABELS))
return table.lookup(target_string)
def main():
input = [["emersoN", "lAke", "aNd", "palmer"],
["i", "haVe", "a", "343yaCht123", "m%an", "2543"]]
sentences_padded, _ = pad_sequences(input, '')
sentences = tf.constant(sentences_padded)
lowercase_sentences = lowercase(sentences)
table = lookup.index_table_from_tensor(mapping=tf.constant(['']),
default_value=1)
sequence_lengths = tf.reduce_sum(table.lookup(sentences), 1)
word_table = lookup.index_table_from_file(vocabulary_file="data/words.txt",
num_oov_buckets=1)
char_table = lookup.index_table_from_file(vocabulary_file="data/chars.txt",
default_value=-1)
sentences_shape = tf.shape(sentences, out_type=tf.int64)
# We need to remove chars not in vocab
removed_char_sentences = remove_unknown_chars(sentences, char_table)
split_words = tf.string_split(tf.reshape(removed_char_sentences, [-1]),
delimiter="")
dense_split_words = tf.sparse_tensor_to_dense(split_words,
default_value='')
max_word_len = tf.gather_nd(split_words.dense_shape, [1])
chars_shape = tf.concat([sentences_shape, [max_word_len]], 0)
chars = tf.reshape(dense_split_words, chars_shape)
word_lengths = tf.reduce_sum(table.lookup(chars), 2)
word_ids = word_table.lookup(sentences)
char_ids = char_table.lookup(chars)
word_mask = tf.sequence_mask(sequence_lengths)
word_ids = tf.where(word_mask, word_ids, tf.zeros_like(word_ids))
char_mask = tf.sequence_mask(word_lengths)
char_ids = tf.where(char_mask, char_ids, tf.zeros_like(char_ids))
config = Config()
# build model
model = NERModel(config)
model.build()
dev = CoNLLDataset(config.filename_dev, max_iter=config.max_iter)
train = CoNLLDataset(config.filename_train, max_iter=config.max_iter)
batch_size = model.config.batch_size
# iterate over dataset
for i, (words, labels) in enumerate(minibatches(train, batch_size)):
print "Start"
fd, _ = model.get_feed_dict(words, labels, model.config.lr,
model.config.dropout)
_, train_loss = model.sess.run([model.train_op, model.loss],
feed_dict=fd)
print "train loss", train_loss
metrics = model.run_evaluate(dev)
msg = " - ".join(
["{} {:04.2f}".format(k, v) for k, v in metrics.items()])
print msg
def construct_input(sequence_feature_map, categorical_values,
categorical_seq_feature, feature_value, mode, normalize,
momentum, min_value, max_value, input_keep_prob):
"""Returns a function to build the model.
Args:
sequence_feature_map: A dictionary of (Sparse)Tensors of dense shape
[batch_size, max_sequence_length, None] keyed by the feature name.
categorical_values: Potential values of the categorical_seq_feature.
categorical_seq_feature: Name of feature of observation code.
feature_value: Name of feature of observation value.
mode: The execution mode, as defined in tf.estimator.ModeKeys.
normalize: Whether to normalize each lab test.
momentum: For the batch normalization mean and variance will be updated as
momentum*old_value + (1-momentum) * new_value.
min_value: Observation values smaller than this will be capped to min_value.
max_value: Observation values larger than this will be capped to max_value.
input_keep_prob: Keep probability for input observation values.
Returns:
- diff_delta_time: Tensor of shape [batch_size, max_seq_length, 1]
with the
- obs_values: A dense representation of the observation_values with
obs_values[b, t, :] has at most one non-zero value at the
position of the corresponding lab test from obs_code_ids with
the value of the lab result. A padded Tensor of shape
[batch_size, max_sequence_length, vocab_size] of type float32
of possibly normalized observation values.
- indicator: A one-hot encoding of whether a value in obs_values comes from
observation_values or is just filled in to be 0. A Tensor of
shape [batch_size, max_sequence_length, vocab_size] and type
float32.
"""
with tf.variable_scope('input'):
sequence_feature_map = {
k: tf.sparse_reorder(s) if isinstance(s, tf.SparseTensor) else s
for k, s in sequence_feature_map.items()
}
# Filter out invalid values.
# For invalid observation values we do this through a sparse retain.
# This makes sure that the invalid values will not be considered in the
# normalization.
observation_values = sequence_feature_map[feature_value]
observation_code_sparse = sequence_feature_map[categorical_seq_feature]
# Future work: Create a flag for the missing value indicator.
valid_values = tf.abs(observation_values.values - 9999999.0) > TOLERANCE
# apply input dropout
if input_keep_prob < 1.0:
random_tensor = input_keep_prob
random_tensor += tf.random_uniform(tf.shape(observation_values.values))
# 0. if [input_keep_prob, 1.0) and 1. if [1.0, 1.0 + input_keep_prob)
dropout_mask = tf.floor(random_tensor)
if mode == tf.estimator.ModeKeys.TRAIN:
valid_values = tf.to_float(valid_values) * dropout_mask
valid_values = valid_values > 0.5
sequence_feature_map[feature_value] = tf.sparse_retain(
observation_values, valid_values)
sequence_feature_map[categorical_seq_feature] = tf.sparse_retain(
observation_code_sparse, valid_values)
# 1. Construct the sequence of observation values to feed into the RNN
# and their indicator.
# We assign each observation code an id from 0 to vocab_size-1. At each
# timestep we will lookup the id for the observation code and take the value
# of the lab test and a construct a vector with all zeros but the id-th
# position is set to the lab test value.
obs_code = sequence_feature_map[categorical_seq_feature]
obs_code_dense_ids = contrib_lookup.index_table_from_tensor(
tuple(categorical_values), num_oov_buckets=0,
name='vocab_lookup').lookup(obs_code.values)
obs_code_sparse = tf.SparseTensor(
values=obs_code_dense_ids,
indices=obs_code.indices,
dense_shape=obs_code.dense_shape)
obs_code_sparse = tf.sparse_reorder(obs_code_sparse)
observation_values = sequence_feature_map[feature_value]
observation_values = tf.sparse_reorder(observation_values)
vocab_size = len(categorical_values)
obs_values, indicator = combine_observation_code_and_values(
obs_code_sparse, observation_values, vocab_size, mode, normalize,
momentum, min_value, max_value)
# 2. We compute the diff_delta_time as additional sequence feature.
# Note, the LSTM is very sensitive to how you encode time.
delta_time = sequence_feature_map['deltaTime']
diff_delta_time = tf.concat(
[delta_time[:, :1, :], delta_time[:, :-1, :]], axis=1) - delta_time
diff_delta_time = tf.to_float(diff_delta_time) / (60.0 * 60.0)
return (diff_delta_time, obs_values, indicator)
def get_label_ids(label_tensor, labels_names):
"""Convert a string tensor of string label names into an int32 tensor of label indices (same shape)."""
label_lookup = lookup.index_table_from_tensor(labels_names, default_value=0)
label_ids = label_lookup.lookup(label_tensor)
return tf.cast(label_ids, tf.int32)
def _lookup_key(key, key_vocab):
table = lookup.index_table_from_tensor(key_vocab, default_value=-1)
key_indices = table.lookup(key)
with tf.control_dependencies([tf.assert_non_negative(key_indices)]):
return tf.identity(key_indices)
def create_sprites_dataset(characters, actions, directions, channels=3,
length=8, shuffle=False, fake_data=False):
"""Creates a tf.data pipeline for the sprites dataset.
Args:
characters: A list of (skin, hair, top, pants) tuples containing
relative paths to the sprite png image for each attribute.
actions: A list of Actions.
directions: A list of Directions.
channels: Number of image channels to yield.
length: Desired length of the sequences.
shuffle: Whether or not to shuffle the characters and sequences
start frame.
fake_data: Boolean for whether or not to yield synthetic data.
Returns:
A tf.data.Dataset yielding (seq, skin label index, hair label index,
top label index, pants label index, action label index, skin label
name, hair label_name, top label name, pants label name, action
label name) tuples.
"""
if fake_data:
dummy_image = tf.random.normal([HEIGHT, WIDTH, CHANNELS])
else:
basedir = download_sprites()
action_names = [action.name for action in actions]
action_metadata = [(action.start_row, action.frames) for action in actions]
direction_rows = [direction.row_offset for direction in directions]
chars = tf.data.Dataset.from_tensor_slices(characters)
act_names = tf.data.Dataset.from_tensor_slices(action_names).repeat()
acts_metadata = tf.data.Dataset.from_tensor_slices(action_metadata).repeat()
dir_rows = tf.data.Dataset.from_tensor_slices(direction_rows).repeat()
if shuffle:
chars = chars.shuffle(len(characters))
dataset = tf.data.Dataset.zip((chars, act_names, acts_metadata, dir_rows))
skin_table = contrib_lookup.index_table_from_tensor(sorted(SKIN_COLORS))
hair_table = contrib_lookup.index_table_from_tensor(sorted(HAIRSTYLES))
top_table = contrib_lookup.index_table_from_tensor(sorted(TOPS))
pants_table = contrib_lookup.index_table_from_tensor(sorted(PANTS))
action_table = contrib_lookup.index_table_from_tensor(sorted(action_names))
def process_example(attrs, act_name, act_metadata, dir_row_offset):
"""Processes a dataset row."""
skin_name = attrs[0]
hair_name = attrs[1]
top_name = attrs[2]
pants_name = attrs[3]
if fake_data:
char = dummy_image
else:
skin = read_image(basedir + os.sep + skin_name)
hair = read_image(basedir + os.sep + hair_name)
top = read_image(basedir + os.sep + top_name)
pants = read_image(basedir + os.sep + pants_name)
char = create_character(skin, hair, top, pants)
if shuffle:
seq = create_random_seq(char, act_metadata, dir_row_offset, length)
else:
seq = create_seq(char, act_metadata, dir_row_offset, length)
seq = seq[..., :channels] # limit output channels
skin_idx = skin_table.lookup(skin_name)
hair_idx = hair_table.lookup(hair_name)
top_idx = top_table.lookup(top_name)
pants_idx = pants_table.lookup(pants_name)
act_idx = action_table.lookup(act_name)
return (seq, skin_idx, hair_idx, top_idx, pants_idx, act_idx,
skin_name, hair_name, top_name, pants_name, act_name)
dataset = dataset.map(process_example)
return dataset
def preprocessing_fn(inputs):
table = lookup.index_table_from_tensor(['a', 'b'])
integerized = table.lookup(inputs['x'])
return {'integerized': integerized}
def convert_label(label):
table = lookup.index_table_from_tensor(['>50K', '<=50K'])
return table.lookup(label)
def init_predict_graph(self):
"""
init predict model graph
:return:
"""
# split 1-D String dense Tensor to words SparseTensor
self.input_sentences = tf.placeholder(dtype=tf.string,
shape=[None],
name='input_sentences')
sparse_words = tf.string_split(self.input_sentences, delimiter=' ')
# slice SparseTensor
valid_indices = tf.less(sparse_words.indices,
tf.constant([self.num_steps], dtype=tf.int64))
valid_indices = tf.reshape(
tf.split(valid_indices, [1, 1], axis=1)[1], [-1])
valid_sparse_words = tf.sparse_retain(sparse_words, valid_indices)
excess_indices = tf.greater_equal(
sparse_words.indices, tf.constant([self.num_steps],
dtype=tf.int64))
excess_indices = tf.reshape(
tf.split(excess_indices, [1, 1], axis=1)[1], [-1])
excess_sparse_words = tf.sparse_retain(sparse_words, excess_indices)
# compute sentences lengths
int_values = tf.ones(shape=tf.shape(valid_sparse_words.values),
dtype=tf.int64)
int_valid_sparse_words = tf.SparseTensor(
indices=valid_sparse_words.indices,
values=int_values,
dense_shape=valid_sparse_words.dense_shape)
input_sentences_lengths = tf.sparse_reduce_sum(int_valid_sparse_words,
axis=1)
# sparse to dense
default_padding_word = self.data_utils._START_VOCAB[0]
words = tf.sparse_to_dense(
sparse_indices=valid_sparse_words.indices,
output_shape=[valid_sparse_words.dense_shape[0], self.num_steps],
sparse_values=valid_sparse_words.values,
default_value=default_padding_word)
# dict words to ids
with open(os.path.join(self.vocab_path, 'words_vocab.txt'),
encoding='utf-8',
mode='rt') as data_file:
words_table_list = [
line.strip() for line in data_file if line.strip()
]
words_table_tensor = tf.constant(words_table_list, dtype=tf.string)
words_table = lookup.index_table_from_tensor(
mapping=words_table_tensor,
default_value=self.data_utils._START_VOCAB_ID[3])
# words_table = lookup.index_table_from_file(os.path.join(vocab_path, 'words_vocab.txt'), default_value=3)
words_ids = words_table.lookup(words)
# blstm model predict
with tf.variable_scope('model', reuse=None):
logits = self.sequence_labeling_model.inference(
words_ids,
input_sentences_lengths,
self.num_classes,
is_training=False)
if self.use_crf:
logits = tf.reshape(logits,
shape=[-1, self.num_steps, self.num_classes])
transition_params = tf.get_variable(
"transitions", [self.num_classes, self.num_classes])
input_sentences_lengths = tf.to_int32(input_sentences_lengths)
predict_labels_ids, sequence_scores = crf.crf_decode(
logits, transition_params, input_sentences_lengths)
predict_labels_ids = tf.to_int64(predict_labels_ids)
sequence_scores = tf.reshape(sequence_scores, shape=[-1, 1])
normalized_sequence_scores = self.tensorflow_utils.score_normalize(
sequence_scores)
predict_scores = tf.matmul(
normalized_sequence_scores,
tf.ones(shape=[1, self.num_steps], dtype=tf.float32))
else:
props = tf.nn.softmax(logits)
max_prop_values, max_prop_indices = tf.nn.top_k(props, k=1)
predict_labels_ids = tf.reshape(max_prop_indices,
shape=[-1, self.num_steps])
predict_labels_ids = tf.to_int64(predict_labels_ids)
predict_scores = tf.reshape(max_prop_values,
shape=[-1, self.num_steps])
predict_scores = tf.as_string(predict_scores, precision=3)
# dict ids to labels
with open(os.path.join(self.vocab_path, 'labels_vocab.txt'),
encoding='utf-8',
mode='rt') as data_file:
labels_table_list = [
line.strip() for line in data_file if line.strip()
]
labels_table_tensor = tf.constant(labels_table_list, dtype=tf.string)
labels_table = lookup.index_to_string_table_from_tensor(
mapping=labels_table_tensor, default_value=self.default_label)
# labels_table = lookup.index_to_string_table_from_file(os.path.join(vocab_path, 'labels_vocab.txt'), default_value='O')
predict_labels = labels_table.lookup(predict_labels_ids)
sparse_predict_labels = self.tensorflow_utils.sparse_concat(
predict_labels, valid_sparse_words, excess_sparse_words,
self.default_label)
sparse_predict_scores = self.tensorflow_utils.sparse_concat(
predict_scores, valid_sparse_words, excess_sparse_words, '0.0')
self.format_predict_labels = self.tensorflow_utils.sparse_string_join(
sparse_predict_labels, 'predict_labels')
self.format_predict_scores = self.tensorflow_utils.sparse_string_join(
sparse_predict_scores, 'predict_scores')
saver = tf.train.Saver()
tables_init_op = tf.tables_initializer()
self.sess = tf.Session()
self.sess.run(tables_init_op)
ckpt = tf.train.get_checkpoint_state(self.checkpoint_path)
if ckpt and ckpt.model_checkpoint_path:
print('read model from {}'.format(ckpt.model_checkpoint_path))
saver.restore(self.sess, ckpt.model_checkpoint_path)
else:
print('No checkpoint file found at %s' % self.checkpoint_path)
return
def decode_target(target_string):
table = lookup.index_table_from_tensor(tf.constant(commons.TARGET_LABELS))
return table.lookup(target_string)
