def unrag_tensor(x: tf.RaggedTensor, max_size: int, axis: int) -> tf.Tensor:
"""Converts a ragged tensor to a full tensor by padding to a maximum size.
This function is useful for converting ragged tensors to a fixed size when one or
more of the dimensions are of variable length.
Args:
x: Ragged tensor to convert.
max_size: Maximum size of the axis to pad.
axis: Axis of `x` to pad to `max_size`. This must specify ragged dimensions.
If more than one axis is specified, `max_size` must be of the same length as
`axis`.
Returns:
A padded version of `x`. Padding will use the equivalent of NaNs in the tensor's
native dtype.
This will replace the shape of the specified `axis` with `max_size`, leaving the
remaining dimensions set to the bounding shape of the ragged tensor.
"""
bounding_shape = x.bounding_shape()
axis = tf.cast(axis, tf.int64)
axis = axis % len(x.shape) # Handle negative indices.
axis = tf.reshape(axis, [-1, 1]) # Ensure (n, 1) shape for indexing.
max_size = tf.cast(max_size, bounding_shape.dtype)
max_size = tf.reshape(max_size, [-1]) # Ensure (n,) shape for indexing.
shape = tf.tensor_scatter_nd_update(bounding_shape, axis, max_size)
return x.to_tensor(default_value=tf.cast(np.NaN, x.dtype), shape=shape)
def _ragged_as_leaf_node(
ragged_tensor: tf.RaggedTensor, is_repeated: bool,
reference_ragged_tensor: tf.RaggedTensor,
options: calculate_options.Options) -> prensor.LeafNodeTensor:
"""Creates a ragged tensor as a leaf node."""
assertions = []
size_dim = tf.compat.dimension_at_index(ragged_tensor.shape, 0).value
reference_size_dim = tf.compat.dimension_at_index(
reference_ragged_tensor.shape, 0).value
if (size_dim is not None and reference_size_dim is not None):
if size_dim != reference_size_dim:
raise ValueError("Returned ragged tensor is not the right size.")
elif options.ragged_checks:
assertions.append(
tf.assert_equal(ragged_tensor.nrows(),
reference_ragged_tensor.nrows()))
if not is_repeated:
rowids = ragged_tensor.value_rowids()
if options.ragged_checks:
assertions.append(
tf.compat.v1.assert_positive(rowids[1:] - rowids[:-1]))
if assertions:
with tf.control_dependencies(assertions):
parent_index = ragged_tensor.value_rowids()
return prensor.LeafNodeTensor(parent_index, ragged_tensor.values,
is_repeated)
else:
parent_index = ragged_tensor.value_rowids()
return prensor.LeafNodeTensor(parent_index, ragged_tensor.values,
is_repeated)
def call(self, texts: tf.Tensor,
substrings: tf.RaggedTensor) -> tf.RaggedTensor:
texts = tf.strings.regex_replace(texts,
pattern=self.special_chars,
rewrite='')
texts = tf.strings.strip(
tf.strings.regex_replace(texts, pattern=r'\s{2,}', rewrite=' '))
substrings = tf.strings.regex_replace(substrings,
pattern=self.special_chars,
rewrite='')
substrings = tf.strings.strip(
tf.strings.regex_replace(substrings,
pattern=r'\s{2,}',
rewrite=' '))
pre = r'.*(\s|^)'
post = r'(\s|$).*'
ragged_texts = tf.RaggedTensor.from_row_lengths(
values=tf.repeat(texts, repeats=substrings.row_lengths()),
row_lengths=substrings.row_lengths())
return tf.ragged.map_flat_values(
self.find_match, ragged_texts,
tf.strings.join([pre, substrings, post]))
def update_state(self,
y_true: tf.RaggedTensor,
y_pred: tf.RaggedTensor,
sample_weight: Optional[tf.Tensor] = None) -> None:
"""Computes edit distance for two RaggedTensors"""
assert isinstance(y_true, tf.RaggedTensor) and isinstance(
y_pred, tf.RaggedTensor)
edit_distances = tf.edit_distance(y_pred.to_sparse(),
y_true.to_sparse(),
normalize=True)
return super().update_state(edit_distances, sample_weight)
def pad_sequence_left(sequences_batch: tf.RaggedTensor, mask: bool):
""" Pad sequences with zeros on left side """
# Truncate rows to have at most `settings.SEQUENCE_LENGTH` items
sequences_batch = sequences_batch[:,-settings.settings.sequence_length:]
if mask:
# Add one to indices, to reserve 0 index for padding
sequences_batch += 1
pad_row_lengths = settings.settings.sequence_length - sequences_batch.row_lengths()
pad_values = tf.zeros( [(settings.settings.sequence_length * sequences_batch.nrows()) - tf.size(sequences_batch, tf.int64)],
sequences_batch.dtype)
padding = tf.RaggedTensor.from_row_lengths(pad_values, pad_row_lengths)
return tf.concat([padding, sequences_batch], axis=1).to_tensor()
def raged_lists_batch_to_multihot(ragged_lists_batch: tf.RaggedTensor, multihot_dim: int) -> tf.Tensor:
""" Maps a batch of label indices to a batch of multi-hot ones """
# TODO: Seems tf.one_hot supports ragged tensors, so try to remove to_tensor call
t = ragged_lists_batch.to_tensor(-1) # Default value = -1 -> one_hot will not assign any one
t = tf.one_hot( t , multihot_dim )
t = tf.reduce_max( t , axis=1 )
return t
def __init__(self,
hierarchical_histogram: tf.RaggedTensor,
use_efficient: bool = False):
"""Initializer for `HierarchicalHistogramDecoder`.
`use_efficient` decides whether to use the accuracy optimization trick from
the paper ["Efficient Use of Differentially Private Binary Trees. James
Honaker".](https://privacytools.seas.harvard.edu/files/privacytools/files/honaker.pdf)
for differentially private hierarchical histogram. The optimization trick
leverages redudant information in the hierarchical histogram to optimize the
accuracy of node queries.
Args:
hierarchical_histogram: A `tf.RaggedTensor` for the hierarchical
histogram.
use_efficient: A boolean indicating the usage of the efficient tree
aggregation algorithm.
"""
self._hierarchical_histogram = hierarchical_histogram.to_list()
_check_hierarchical_histogram_shape(self._hierarchical_histogram)
if len(self._hierarchical_histogram) == 1:
self._arity = 2
else:
self._arity = int(
len(self._hierarchical_histogram[1]) /
len(self._hierarchical_histogram[0]))
self._size = len(hierarchical_histogram[-1])
self._num_layers = math.ceil(math.log(self._size, self._arity)) + 1
self._use_efficient = use_efficient
def embedding_pooling(tensor: tf.RaggedTensor, combiner: str = "sqrtn"):
tensor_sum = tf.math.reduce_sum(tensor, axis=1)
if combiner == "sum":
return tensor_sum
row_lengths = tf.expand_dims(tensor.row_lengths(axis=1), axis=1)
row_lengths = tf.math.maximum(tf.ones_like(row_lengths), row_lengths)
row_lengths = tf.cast(row_lengths, dtype=tf.float32)
if combiner == "mean":
return tensor_sum / row_lengths
if combiner == "sqrtn":
return tensor_sum / tf.math.sqrt(row_lengths)
def pad_sequence_right( sequences_batch: tf.RaggedTensor, mask: bool) -> tf.Tensor:
""" Pad sequences with zeros on right side """
# Avoid sequences larger than sequence_length: Get last sequence_length of each sequence
sequences_batch = sequences_batch[:,-settings.settings.sequence_length:]
if mask:
# Add one to indices, to reserve 0 index for padding
sequences_batch += 1
# Convert to dense, padding zeros to the right
sequences_batch = sequences_batch.to_tensor(0, shape=[None, settings.settings.sequence_length])
return sequences_batch
def _run_model_filter_empty_sequences(self,
batch_item_indices: tf.RaggedTensor,
batch_customer_indices, n_results):
# Check if there are empty sequences
sequences_lenghts = batch_item_indices.row_lengths()
non_empty_seq_count = tf.math.count_nonzero(sequences_lenghts)
n_sequences = tf.shape(sequences_lenghts, tf.int64)[0]
#print(">>>", non_empty_seq_count, n_results)
if non_empty_seq_count == 0:
# All sequences are empty
label_predictions = tf.zeros([n_sequences, n_results],
dtype=tf.string)
probs_predictions = tf.zeros([n_sequences, n_results],
dtype=tf.float32)
return (label_predictions, probs_predictions)
elif non_empty_seq_count >= n_sequences:
# There are no empty sequences. Run the model with the full batch
return self._run_model_and_postprocess(batch_item_indices,
batch_customer_indices,
n_results)
else:
# There are some empty sequences
# Model will fail if a sequence is empty, and it seems it's the expected behaviour: Do not feed empty sequences
# Get non empty sequences mask
non_empty_mask = tf.math.greater(sequences_lenghts, 0)
# Get non empty sequences
non_empty_sequences: tf.RaggedTensor = tf.ragged.boolean_mask(
batch_item_indices, non_empty_mask)
non_empty_customers = tf.boolean_mask(batch_customer_indices,
non_empty_mask)
# Run model
label_predictions, probs_predictions = self._run_model_and_postprocess(
non_empty_sequences, non_empty_customers, n_results)
# Merge real predictions with empty predictions for empty sequences:
indices = tf.where(non_empty_mask)
final_shape = [n_sequences, n_results]
label_predictions = tf.scatter_nd(indices, label_predictions,
final_shape)
#print(label_predictions)
probs_predictions = tf.scatter_nd(indices, probs_predictions,
final_shape)
#print(probs_predictions)
return (label_predictions, probs_predictions)
def __init__(
self,
num_states: int,
data: tf.Tensor,
indices: tf.Tensor,
indptr: tf.Tensor,
states: tf.RaggedTensor,
ref_ids: tf.RaggedTensor,
ref_ids_lookup: Dict[str, int],
) -> None:
self._num_states = tf.Variable([num_states], trainable=False)
self._data = tf.Variable(data, trainable=False)
self._indices = tf.Variable(indices, trainable=False)
self._indptr = tf.Variable(indptr, trainable=False)
self._states = tf.Variable(states.to_tensor(), trainable=False)
self._ref_ids = tf.Variable(ref_ids.to_tensor(), trainable=False)
keys = tf.convert_to_tensor(list(ref_ids_lookup.keys()),
dtype=tf.string)
values = tf.convert_to_tensor(list(ref_ids_lookup.values()),
dtype=tf.int32)
initializer = tf.lookup.KeyValueTensorInitializer(keys, values)
self._ref_ids_lookup = tf.lookup.StaticHashTable(initializer,
default_value=0)
def multi_behavior_embedding_pooling(tensor: tf.RaggedTensor,
combiner: str = "sqrtn"):
if len(tensor.shape) == 3:
tensor = tf.expand_dims(
tensor, axis=2) # batch_size * behavior_count * 1 * embedding_size
tensor_sum = tf.math.reduce_sum(
tensor, axis=2) # batch_size * behavior_count * embedding_size
if combiner == "sum":
return tensor_sum
row_lengths = tf.expand_dims(tensor.row_lengths(axis=2), axis=2)
row_lengths = tf.math.maximum(tf.ones_like(row_lengths), row_lengths)
row_lengths = tf.cast(row_lengths, dtype=tf.float32)
if combiner == "mean":
return tensor_sum / row_lengths
if combiner == "sqrtn":
return tensor_sum / tf.math.sqrt(row_lengths)
def _truncate_row_lengths(ragged_tensor: tf.RaggedTensor,
new_lengths: tf.Tensor) -> tf.RaggedTensor:
"""Truncates the rows of `ragged_tensor` to the given row lengths."""
new_lengths = tf.broadcast_to(new_lengths,
ragged_tensor.bounding_shape()[0:1])
def fn(x):
row, new_length = x
return row[0:new_length]
fn_dtype = tf.RaggedTensorSpec(dtype=ragged_tensor.dtype,
ragged_rank=ragged_tensor.ragged_rank - 1)
result = tf.map_fn(fn, (ragged_tensor, new_lengths), dtype=fn_dtype)
# Work around broken shape propagation: without this, result has unknown rank.
flat_values_shape = [None] * ragged_tensor.flat_values.shape.rank
result = result.with_flat_values(
tf.ensure_shape(result.flat_values, flat_values_shape))
return result
