def LoadDataset(self, file_pattern):
file_pattern_glob = py_utils.ShardedFilePatternToGlob(file_pattern)
dataset = tf.data.Dataset.list_files(file_pattern_glob, shuffle=False)
def MakeExample(data):
return py_utils.NestedMap(data=data, source_id=tf.constant(0))
return dataset.map(MakeExample, deterministic=True)
def LoadDataset(self, file_pattern):
file_pattern_glob = py_utils.ShardedFilePatternToGlob(file_pattern)
filenames = sorted(tf.io.gfile.glob(file_pattern_glob))
dataset = tf.data.Dataset.from_tensor_slices(filenames)
def MakeExample(data):
return py_utils.NestedMap(data=data, source_id=tf.constant(0))
return dataset.map(MakeExample, deterministic=True)
def GetDataset(self):
p = self.params
if not p.file_pattern:
raise ValueError('A file pattern must be provided.')
file_pattern_glob = py_utils.ShardedFilePatternToGlob(p.file_pattern)
dataset = tf.data.Dataset.list_files(
file_pattern_glob,
shuffle=not self.cluster.require_sequential_input_order)
dataset = tf.data.TextLineDataset(
dataset,
num_parallel_reads=(1 if self.cluster.in_unit_test else
tf.data.experimental.AUTOTUNE))
if not self.cluster.require_sequential_input_order:
dataset = dataset.shuffle(p.shuffle_buffer_size,
reshuffle_each_iteration=True)
dataset = dataset.repeat()
return dataset
