Edward Loper e2ff7f453e Extend the ragged version of tf.gather to support batch_dims and axis args.
PiperOrigin-RevId: 299158220
Change-Id: I8cac49a4e2bac64c867c0997aa8f829dc569eec4
2020-03-05 12:03:50 -08:00

501 lines
21 KiB
Python

# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Gather operations for RaggedTensors."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import indexed_slices
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor_shape
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import gen_ragged_array_ops
from tensorflow.python.ops import math_ops
from tensorflow.python.ops.ragged import ragged_array_ops
from tensorflow.python.ops.ragged import ragged_math_ops
from tensorflow.python.ops.ragged import ragged_tensor
#===============================================================================
# ragged_gather
#===============================================================================
def gather(params,
indices,
validate_indices=None,
axis=None,
batch_dims=0,
name=None):
"""Gathers ragged slices from `params` axis `0` according to `indices`.
See `tf.gather` for full documentation. (This version has the same API
as `tf.gather`, but supports ragged `params` and `indices`.)
Examples:
>>> params = tf.constant(['a', 'b', 'c', 'd', 'e'])
>>> indices = tf.constant([3, 1, 2, 1, 0])
>>> ragged_params = tf.ragged.constant([['a', 'b', 'c'], ['d'], [], ['e']])
>>> ragged_indices = tf.ragged.constant([[3, 1, 2], [1], [], [0]])
>>> tf.gather(params, ragged_indices)
<tf.RaggedTensor [[b'd', b'b', b'c'], [b'b'], [], [b'a']]>
>>> tf.gather(ragged_params, indices)
<tf.RaggedTensor [[b'e'], [b'd'], [], [b'd'], [b'a', b'b', b'c']]>
>>> tf.gather(ragged_params, ragged_indices)
<tf.RaggedTensor [[[b'e'], [b'd'], []], [[b'd']], [], [[b'a', b'b', b'c']]]>
Args:
params: The potentially ragged tensor from which to gather values. Must be
at least rank 1.
indices: The potentially ragged tensor indicating which values to gather.
Must have dtype `int32` or `int64`. Values must be in the range `[0,
params.shape[0]]`.
validate_indices: Ignored.
axis: The axis in `params` to gather `indices` from.
batch_dims: The number of batch dimensions.
name: A name for the operation (optional).
Returns:
A `RaggedTensor`, where `output.dtype=params.dtype` and
`output.shape=indices.shape + params.shape[1:]` and
`output.ragged_rank=indices.shape.ndims + params.ragged_rank`.
Raises:
ValueError: If indices.shape.ndims is not known statically.
"""
del validate_indices
with ops.name_scope(name, 'RaggedGather', [params, indices]):
params = ragged_tensor.convert_to_tensor_or_ragged_tensor(
params, name='params')
indices = ragged_tensor.convert_to_tensor_or_ragged_tensor(
indices, name='indices')
params, indices = ragged_tensor.match_row_splits_dtypes(params, indices)
if batch_dims != indices.shape.rank:
batch_dims = array_ops.get_positive_axis(
batch_dims,
indices.shape.rank,
axis_name='batch_dims',
ndims_name='rank(indices)')
if params.shape.rank is not None and batch_dims >= params.shape.rank:
raise ValueError('batch_dims must be less than rank(params)')
if axis is None:
axis = batch_dims
axis = array_ops.get_positive_axis(
axis, params.shape.rank, ndims_name='rank(params)')
if axis < batch_dims:
raise ValueError('axis must be greater than or equal to batch_dims')
if indices.shape.rank is not None:
if not 0 <= batch_dims <= indices.shape.rank:
raise ValueError(
'batch_dims=%s must be between 0 and rank(indices)=%s' %
(batch_dims, indices.shape.rank))
return _gather(params, indices, axis, batch_dims)
def _gather(params, indices, axis, batch_dims):
"""Helper that implements the body for ragged gather().
Assumes that `params` and `indices` have been converted to tensors or
ragged tensors, and that `axis` and `batch_dims` have been normalized to
be positive. (So these conversions & normalizations can be skipped in
recursive calls to _gather).
Args:
params: The tensor from which to gather values.
indices: The indices of values to gather.
axis: The axis in `params` to gather `indices` from.
batch_dims: The number of batch dimensions.
Returns:
A potentially ragged tensor.
"""
params_is_ragged = ragged_tensor.is_ragged(params)
indices_is_ragged = ragged_tensor.is_ragged(indices)
if not (params_is_ragged or indices_is_ragged):
return array_ops.gather(params, indices, axis=axis, batch_dims=batch_dims)
if batch_dims > 0:
return _batch_gather(params, indices, axis, batch_dims)
if axis > 0:
return _axis_gather(params, indices, axis)
if indices_is_ragged:
return indices.with_values(_gather(params, indices.values, 0, 0))
if indices.shape.ndims is None:
raise ValueError('rank(indices) must be known statically')
out_ragged_rank = indices.shape.ndims + len(params.nested_row_splits) - 1
result = gen_ragged_array_ops.ragged_gather(
indices=indices,
params_dense_values=params.flat_values,
params_nested_splits=params.nested_row_splits,
OUTPUT_RAGGED_RANK=out_ragged_rank)
result = ragged_tensor.RaggedTensor.from_nested_row_splits(
result.output_dense_values, result.output_nested_splits, validate=False)
# Inject uniform_row_lengths into the result RaggedTensors for dimensions
# corresponding to dense outer dimensions of `indices`.
# TODO(edloper): Change this to construct the result using RowPartition
# objects instead, so we don't need to modify private variables.
if indices.shape.ndims > 1:
target = result
indices_shape = array_ops.shape(indices, out_type=params.row_splits.dtype)
shape_cumprod = math_ops.cumprod(indices_shape)
for dim in range(indices.shape.ndims - 1):
# pylint: disable=protected-access
target._cached_nrows = shape_cumprod[dim]
target._uniform_row_length = indices_shape[dim + 1]
target = target.values
return result
def _batch_gather(params, indices, axis, batch_dims):
"""Helper that implements the body for ragged gather() when batch_dims>0.
Args:
params: The tensor from which to gather values.
indices: The indices of values to gather.
axis: The axis in `params` to gather `indices` from.
batch_dims: The number of batch dimensions.
Returns:
A potentially ragged tensor.
"""
# Perform static checks that `params` and `indices` have compatible batch
# dimensions. Note: we do not perform *runtime* checks that `params` and
# `indices` actually have the same row-splits (because we wish to avoid the
# runtime cost of those checks). If `params` and `indices` are
# incompatible, the resulting `RaggedTensor` may be nonsensical.
if not params.shape[:batch_dims].is_compatible_with(
indices.shape[:batch_dims]):
raise ValueError('batch shape from indices %s does not match params '
'shape %s' % (indices.shape[:batch_dims], params.shape))
if batch_dims > 1:
# Convert params & indices to ragged tensors.
if not isinstance(params, ragged_tensor.RaggedTensor):
if indices.uniform_row_length is None:
raise ValueError(
'batch shape from indices does not match params shape: ragged '
'indices dimension corresponds to uniform params dimension')
params = ragged_tensor.RaggedTensor.from_tensor(
params, ragged_rank=1, row_splits_dtype=indices.row_splits.dtype)
if not isinstance(indices, ragged_tensor.RaggedTensor):
if params.uniform_row_length is None:
raise ValueError(
'batch shape from indices does not match params shape: ragged '
'params dimension corresponds to uniform indices dimension')
indices = ragged_tensor.RaggedTensor.from_tensor(
indices, ragged_rank=1, row_splits_dtype=params.row_splits.dtype)
# Flatten the two outer batch dimensions into a single batch dimension,
# and recurse.
return params.with_values(
_gather(params.values, indices.values, axis - 1, batch_dims - 1))
if axis > 1:
# Convert an axis dimension into a batch dimension, by adding a dimension
# to `indices`, and tiling it to match `params`. E.g., if `params`
# had shape `[B, P1, P2]`, and `indices` had shape `[B, I1, I2]`, then we
# tile `indices` to have shape `[B, P1, I1, I2]`. That way, we can treat
# the `P1` dimension as a batch dimension.
if not isinstance(indices, ragged_tensor.RaggedTensor):
adjusted_indices = params.with_values(
array_ops.repeat(indices, params.row_lengths(), 0))
else:
if not isinstance(params, ragged_tensor.RaggedTensor):
params = ragged_tensor.RaggedTensor.from_tensor(
params, ragged_rank=1, row_splits_dtype=indices.row_splits.dtype)
adjusted_indices = _gather(
indices,
params.with_values(
array_ops.repeat(
math_ops.range(params.nrows()), params.row_lengths())), 0, 0)
return _batch_gather(params, adjusted_indices, axis, batch_dims + 1)
if indices.shape.rank is None:
raise ValueError('rank(indices) must be known statically')
assert batch_dims == 1
# If params.shape=[B, P1...PN] and indices.shape=[B, I1...IM], then:
#
# output[b, i1...im, p2...pn] =
# params[b, indices[b, i1...im], p2...pn]
#
# We construct `output` by flattening `params`, adjusting the `indices` to
# point into that flattened list, and recursively calling `gather`.
flat_params = _flatten_dims_0_and_1(params)
adjustments = _row_starts(params, indices.dtype) # offset for each batch
# increase adjustments's rank so it broadcasts w/ the outer dim of indices
adjustments = _increase_rank_to(adjustments, indices.shape.ndims)
adjusted_indices = indices + adjustments
return _gather(flat_params, adjusted_indices, axis - 1, 0)
def _axis_gather(params, indices, axis):
"""Helper that implements ragged gather when axis>0 and batch_dims==0.
Args:
params: The tensor from which to gather values.
indices: The indices of values to gather.
axis: The axis in `params` to gather `indices` from.
Returns:
A potentially ragged tensor.
"""
if axis > 1:
if not isinstance(params, ragged_tensor.RaggedTensor):
params = ragged_tensor.RaggedTensor.from_tensor(
params, ragged_rank=1, row_splits_dtype=indices.row_splits.dtype)
# Recurse, using the flattened params (but do not flatten indices).
return params.with_values(_gather(params.values, indices, axis - 1, 0))
if indices.shape.rank is None:
raise ValueError('rank(indices) must be known statically')
if (isinstance(params, ragged_tensor.RaggedTensor) and
params.uniform_row_length is None):
raise ValueError('axis may not be a ragged dimension')
assert axis == 1
# If params.shape=[P1...PN] and indices.shape=[I1...IM], then:
#
# output[p1, i1...im, p3...pn] =
# params[p1, indices[i1...im], p3...pn]
#
# We construct `output` by flattening `params`, adjusting the `indices` to
# have one additional dimension, and to point into that flattened list, and
# recursively calling `gather`.
flat_params = _flatten_dims_0_and_1(params)
adjustments = _row_starts(params, indices.dtype) # offset for each batch
adjustments = _increase_rank_to(adjustments, indices.shape.ndims + 1)
adjusted_indices = indices + adjustments
return _gather(flat_params, adjusted_indices, axis - 1, 0)
def _flatten_dims_0_and_1(t):
"""Returns a copy of `t` with the outer two dimensions merged."""
if isinstance(t, ragged_tensor.RaggedTensor):
return t.values
else:
t_shape = array_ops.shape(t)
return array_ops.reshape(t, array_ops.concat([[-1], t_shape[2:]], axis=0))
def _row_starts(t, dtype):
"""Returns the start indices for the rows in `t`."""
if isinstance(t, ragged_tensor.RaggedTensor):
return math_ops.cast(t.row_starts(), dtype)
else:
t_shape = array_ops.shape(t, out_type=dtype)
return math_ops.range(t_shape[0]) * t_shape[1]
def _increase_rank_to(t, rank):
"""Adds *trailing* size-1 dimensions to `t` until it has the given rank."""
if isinstance(t, ragged_tensor.RaggedTensor):
return t.with_values(_increase_rank_to(t, rank - 1))
else:
old_dims = array_ops.shape(t)
new_dims = array_ops.ones([rank - array_ops.rank(t)], old_dims.dtype)
new_shape = array_ops.concat([old_dims, new_dims], axis=0)
return array_ops.reshape(t, new_shape)
#===============================================================================
# ragged.gather_nd
#===============================================================================
def gather_nd(params, indices, batch_dims=0, name=None):
"""Gather slices from `params` using `n`-dimensional indices.
This operation is similar to `gather`, but it uses the innermost dimension
of `indices` to define a slice into `params`. In particular, if:
* `indices` has shape `[A1...AN, I]`
* `params` has shape `[B1...BM]`
Then:
* `result` has shape `[A1...AN, B_{I+1}...BM]`.
* `result[a1...aN] = params[indices[a1...aN, :]]`
Args:
params: A potentially ragged tensor with shape `[A1...AN, I]`.
indices: A potentially ragged tensor with shape `[B1...BM]`.
batch_dims: Must be zero.
name: A name for the operation (optional).
Returns:
A potentially ragged tensor with shape `[A1...AN, B_{I+1}...BM]`.
#### Examples:
>>> params = tf.ragged.constant(
... [ [ ['000', '001'], ['010' ] ],
... [ ['100' ], ['110', '111', '112'], ['120'] ],
... [ [ ], ['210' ] ] ])
>>> # Gather 2D slices from a 3D tensor
>>> tf.gather_nd(params, [[2], [0]])
<tf.RaggedTensor [[[], [b'210']], [[b'000', b'001'], [b'010']]]>
>>> # Gather 1D slices from a 3D tensor
>>> tf.gather_nd(params, [[2, 1], [0, 0]])
<tf.RaggedTensor [[b'210'], [b'000', b'001']]>
>>> # Gather scalars from a 3D tensor
>>> tf.gather_nd(params, [[0, 0, 1], [1, 1, 2]]).numpy()
array([b'001', b'112'], dtype=object)
"""
if not isinstance(batch_dims, int) or batch_dims != 0:
raise ValueError('batch_dims != 0 is not supported for ragged gather yet.')
if not (ragged_tensor.is_ragged(params) or ragged_tensor.is_ragged(indices)):
return array_ops.gather_nd(params, indices, name)
with ops.name_scope(name, 'RaggedGatherNd', [params, indices]):
params = ragged_tensor.convert_to_tensor_or_ragged_tensor(
params, name='params')
indices = ragged_tensor.convert_to_tensor_or_ragged_tensor(
indices, name='indices')
params, indices = ragged_tensor.match_row_splits_dtypes(params, indices)
indices_shape = indices.shape
indices_ndims = indices_shape.ndims
if indices_ndims is None:
raise ValueError('indices.rank be statically known.')
if indices_ndims == 0:
raise ValueError('indices.rank must be at least 1.')
if (ragged_tensor.is_ragged(indices) and
indices_ndims == indices.ragged_rank + 1):
raise ValueError('The innermost dimension of indices may not be ragged')
# `index_size` is the "n" in "gather_nd" -- i.e., the number of dimensions
# that each index slices into.
index_size = tensor_shape.dimension_value(indices_shape[-1])
if index_size is None:
raise ValueError('indices.shape[-1] must be statically known.')
# If `indices` has more than 2 dimensions, then recurse. If `indices` is
# dense, then we convert it to ragged before recursing, and then convert
# the result back to `dense` if appropriate.
if indices_ndims > 2:
indices_is_dense = not ragged_tensor.is_ragged(indices)
if indices_is_dense:
indices = ragged_tensor.RaggedTensor.from_tensor(
indices, ragged_rank=indices_ndims - 2,
row_splits_dtype=params.row_splits.dtype)
result = indices.with_flat_values(gather_nd(params, indices.flat_values))
if (indices_is_dense and ragged_tensor.is_ragged(result) and
result.ragged_rank == indices_ndims - 2):
result = ragged_tensor.RaggedTensor.to_tensor(result)
return result
# indices_ndims <= 2, and the innermost dimension of indices may not be
# ragged, so `indices` must not be ragged.
assert not ragged_tensor.is_ragged(indices)
assert ragged_tensor.is_ragged(params)
# Handle corner case: An empty index tuple selects the entire `params`
# value. So if `index_size` is zero, then tile `params`.
if index_size == 0:
params_ndims = params.ragged_rank + array_ops.rank(params.flat_values)
for dim in range(indices_ndims - 1):
params = ragged_array_ops.expand_dims(params, axis=0)
multiples = array_ops.concat([
array_ops.shape(indices)[:-1],
array_ops.ones([params_ndims], dtypes.int32)
],
axis=0)
return ragged_array_ops.tile(params, multiples)
# When index_size=1, we can just flatten the index tuples and use gather.
elif index_size == 1:
flattened_index_tuples = array_ops.reshape(indices, [-1])
return gather(params, flattened_index_tuples)
# Otherwise, params is a RaggedTensor, and indices is a 1D or 2D Tensor.
# Flatten both the index tuples and the params, such that the flattened
# index tuples point to the correct values in the flattened params; and
# then use ragged.gather on the flattened index tuples & params.
else:
indices = math_ops.cast(indices, params.row_splits.dtype)
# Flatten the outermost 2 dimensions of the index tuples & params.
flattened_index_tuples = array_ops.gather(params.row_splits,
indices[..., 0])
flattened_index_tuples += indices[..., 1]
flattened_params = params.values
# Flatten any remaining dimensions.
for dim in range(2, index_size):
if not ragged_tensor.is_ragged(flattened_params):
flattened_index_tuples = array_ops.expand_dims(
flattened_index_tuples, axis=1)
flattened_index_tuples = array_ops.concat(
[flattened_index_tuples, indices[..., dim:]], axis=1)
return array_ops.gather_nd(flattened_params, flattened_index_tuples)
flattened_index_tuples = array_ops.gather(
flattened_params.row_starts(), flattened_index_tuples)
flattened_index_tuples += indices[..., dim]
flattened_params = flattened_params.values
# Gather using the flattened index tuples and params.
return gather(flattened_params, flattened_index_tuples)
#===============================================================================
# Gradient for the RaggedGather kernel
#===============================================================================
@ops.RegisterGradient('RaggedGather')
def _ragged_gather_grad(op, *grads):
"""Gradient for RaggedGather op."""
param_nested_splits = op.inputs[:-2]
param_inner_values = op.inputs[-2]
indices = op.inputs[-1]
grad_inner_values = grads[-1]
# For each row in `params`, find the range of values in `params.inner_values`
# that is covered by that row. In particular, the values in row `i` are
# `param_inner_values[combined_splits[i]:combined_splits[i+1]`.
combined_splits = param_nested_splits[0]
for row_splits in param_nested_splits[1:]:
combined_splits = array_ops.gather(row_splits, combined_splits)
# The outer dimensions of `indices` correspond 1:1 with the outer dimensions
# of `ragged_grad` that are encoded by `grad_nested_splits`. Thus, the
# flattened `indices` correspond 1:1 with `grad_inner_values`.
flat_indices = array_ops.reshape(indices, [-1])
# Build an IndexedSlices where the values are taken from `flat_grad`.
grad_indices = ragged_math_ops.range(
array_ops.gather(combined_splits, flat_indices),
array_ops.gather(combined_splits[1:], flat_indices)).values
param_inner_values_grad = indexed_slices.IndexedSlices(
values=grad_inner_values, indices=grad_indices,
dense_shape=array_ops.shape(param_inner_values))
return [None for _ in param_nested_splits] + [param_inner_values_grad, None]