def test_pack_dataset_no_eos(self, use_custom_ops):
x = [{
"inputs": [7, 8, 5],
"targets": [3, 9]
}, {
"inputs": [8, 4, 9, 3],
"targets": [4]
}]
ds = create_default_dataset(x)
packed_ds = dataset.pack_dataset(ds,
length={
"inputs": 8,
"targets": 5
},
use_custom_ops=use_custom_ops)
# Packing still works without the eos.
expected = [{
"inputs": [7, 8, 5, 8, 4, 9, 3, 0],
"inputs_segmentation": [1, 1, 1, 2, 2, 2, 2, 0],
"inputs_position": [0, 1, 2, 0, 1, 2, 3, 0],
"targets": [3, 9, 4, 0, 0],
"targets_position": [0, 1, 0, 0, 0],
"targets_segmentation": [1, 1, 2, 0, 0],
}]
self.assert_dataset(packed_ds, expected, {
"inputs": tf.int32,
"targets": tf.int32
})
def test_pack_dataset(self, use_custom_ops):
x = [{
"inputs": [7, 8, 5, 1],
"targets": [3, 9, 1],
"idx": [0]
}, {
"inputs": [8, 4, 9, 3, 1],
"targets": [4, 1],
"idx": [1]
}]
ds = create_default_dataset(x,
feature_names=("inputs", "targets", "idx"))
packed_ds = dataset.pack_dataset(ds,
length={
"inputs": 10,
"targets": 7
},
keys=("inputs", "targets"),
use_custom_ops=use_custom_ops)
expected = [{
"inputs": [7, 8, 5, 1, 8, 4, 9, 3, 1, 0],
"inputs_segmentation": [1, 1, 1, 1, 2, 2, 2, 2, 2, 0],
"inputs_position": [0, 1, 2, 3, 0, 1, 2, 3, 4, 0],
"targets": [3, 9, 1, 4, 1, 0, 0],
"targets_position": [0, 1, 2, 0, 1, 0, 0],
"targets_segmentation": [1, 1, 1, 2, 2, 0, 0],
}]
self.assert_dataset(packed_ds, expected, {
"inputs": tf.int32,
"targets": tf.int32
})
def pack_or_pad_ll(dataset,
length,
pack=True,
feature_keys=None,
ensure_eos=False,
shift_decoder_output=False,
target_prefix_attributes=None,
tokenizer=None):
"""Creates a 'packed' version of a dataset or pads examples with zeros.
If pack=True, then multiple examples concatenated to form one combined
example with the given length.
If pack=False, then examples are padded with zeros to 'length'.
Args:
dataset: a tf.data.Dataset
length: an integer or a dict from feature-key to integer
pack: a boolean, whether to pack (True) or pad (False).
feature_keys: (optional) list of strings, the feature names to limit
packing or padding to. Packing will filter out other features whereas
padding will pass them through unchanged. Defaults to all features.
ensure_eos: a boolean, whether to replace the final token with EOS=1 if it
is not PAD=0.
Returns:
a tf.data.Dataset where all features have fixed shape [length].
"""
feature_keys = feature_keys or list(dataset.output_shapes.keys())
if shift_decoder_output:
left_pad_amts = [
len(tokenizer.encode(target_prefix_attribute)) - 1
for target_prefix_attribute in target_prefix_attributes
]
dataset = shift_decoder_output_fn(dataset,
left_pad_amts=left_pad_amts,
feature_keys=feature_keys)
if pack:
dataset = pack_dataset(dataset, length=length, keys=feature_keys)
# Pad/trim length of each example to length.
dataset = trim_and_pad_dataset(dataset,
length=length,
feature_keys=feature_keys)
if ensure_eos:
dataset = ensure_dataset_eos_ll(dataset, feature_keys)
return dataset
def test_pack_dataset_long_seq(self, use_custom_ops):
x = [{
"inputs": [7, 8, 5, 6, 9, 4, 1],
"targets": [3, 9, 1]
}, {
"inputs": [8, 4, 9, 3, 5, 7, 9, 1],
"targets": [4, 1]
}]
ds = create_default_dataset(x)
packed_ds = dataset.pack_dataset(ds,
length={
"inputs": 7,
"targets": 3
},
use_custom_ops=use_custom_ops)
expected = [
{
"inputs": [7, 8, 5, 6, 9, 4, 1],
"inputs_segmentation": [1, 1, 1, 1, 1, 1, 1],
"inputs_position": [0, 1, 2, 3, 4, 5, 6],
"targets": [3, 9, 1],
"targets_position": [0, 1, 2],
"targets_segmentation": [1, 1, 1],
},
{
# EOS is trimmed
"inputs": [8, 4, 9, 3, 5, 7, 9],
"inputs_segmentation": [1, 1, 1, 1, 1, 1, 1],
"inputs_position": [0, 1, 2, 3, 4, 5, 6],
"targets": [4, 1, 0],
"targets_position": [0, 1, 0],
"targets_segmentation": [1, 1, 0],
}
]
self.assert_dataset(packed_ds, expected, {
"inputs": tf.int32,
"targets": tf.int32
})
