def _PruneSparseTensor(unpruned, pruned_pattern):
"""Helper function to prune COO sparse tensor.
Given two sparse tensors 'unpruned' and 'pruned_pattern', generates another
sparse tensor with indices and values fron 'unpruned' only if its indices also
occur in pruned_pattern.
Args:
unpruned: COO matrix with unpruned indices
pruned_pattern: COO matrix with pruned pattern.
TODO(tabakg): This is far from optimal. Consider a C++ implementation.
Returns:
Indices, values, and dense_shape of the pruned matrix.
"""
pruned_indices = sparse_ops.sparse_reshape(pruned_pattern,
shape=(-1, )).indices[..., 0]
unpruned_indices = sparse_ops.sparse_reshape(unpruned,
shape=(-1, )).indices[..., 0]
best_match = array_ops.searchsorted(unpruned_indices, pruned_indices)
keep_indices = array_ops.gather(
best_match,
array_ops.where(
math_ops.equal(array_ops.gather(unpruned_indices, best_match),
pruned_indices)))
return (array_ops.gather_nd(unpruned.indices, keep_indices),
array_ops.gather_nd(unpruned.values,
keep_indices), pruned_pattern.dense_shape)
def testFeedPartialShapes(self):
with self.session(use_gpu=False):
# Incorporate new rank into shape information if known
sp_input = self._SparseTensorPlaceholder()
sp_output = sparse_ops.sparse_reshape(sp_input, [2, 3, 5])
self.assertListEqual(sp_output.indices.get_shape().as_list(),
[None, 3])
self.assertListEqual(sp_output.dense_shape.get_shape().as_list(),
[3])
# Incorporate known shape information about input indices in output
# indices
sp_input = self._SparseTensorPlaceholder()
sp_input.indices.set_shape([5, None])
sp_output = sparse_ops.sparse_reshape(sp_input, [2, 3, 5])
self.assertListEqual(sp_output.indices.get_shape().as_list(),
[5, 3])
self.assertListEqual(sp_output.dense_shape.get_shape().as_list(),
[3])
# Even if new_shape has no shape information, we know the ranks of
# output indices and shape
sp_input = self._SparseTensorPlaceholder()
sp_input.indices.set_shape([5, None])
new_shape = array_ops.placeholder(dtypes.int64)
sp_output = sparse_ops.sparse_reshape(sp_input, new_shape)
self.assertListEqual(sp_output.indices.get_shape().as_list(),
[5, None])
self.assertListEqual(sp_output.dense_shape.get_shape().as_list(),
[None])
def _get_sparse_tensors(self, inputs, weight_collections=None,
trainable=None):
sparse_tensors = self.categorical_column._get_sparse_tensors(inputs)
id_tensor = sparse_tensors.id_tensor
weight_tensor = sparse_tensors.weight_tensor
# Expands final dimension, so that embeddings are not combined during
# embedding lookup.
check_id_rank = check_ops.assert_equal(
array_ops.rank(id_tensor), 2,
data=[
'Column {} expected ID tensor of rank 2. '.format(self.name),
'id_tensor shape: ', array_ops.shape(id_tensor)])
with ops.control_dependencies([check_id_rank]):
id_tensor = sparse_ops.sparse_reshape(
id_tensor,
shape=array_ops.concat([id_tensor.dense_shape, [1]], axis=0))
if weight_tensor is not None:
check_weight_rank = check_ops.assert_equal(
array_ops.rank(weight_tensor), 2,
data=[
'Column {} expected weight tensor of rank 2.'.format(self.name),
'weight_tensor shape:', array_ops.shape(weight_tensor)])
with ops.control_dependencies([check_weight_rank]):
weight_tensor = sparse_ops.sparse_reshape(
weight_tensor,
shape=array_ops.concat([weight_tensor.dense_shape, [1]], axis=0))
return fc._CategoricalColumn.IdWeightPair(id_tensor, weight_tensor)
def _get_sparse_tensors(self, inputs, weight_collections=None,
trainable=None):
sparse_tensors = self.categorical_column._get_sparse_tensors(inputs)
id_tensor = sparse_tensors.id_tensor
weight_tensor = sparse_tensors.weight_tensor
# Expands final dimension, so that embeddings are not combined during
# embedding lookup.
check_id_rank = check_ops.assert_equal(
array_ops.rank(id_tensor), 2,
data=[
'Column {} expected ID tensor of rank 2. '.format(self.name),
'id_tensor shape: ', array_ops.shape(id_tensor)])
with ops.control_dependencies([check_id_rank]):
id_tensor = sparse_ops.sparse_reshape(
id_tensor,
shape=array_ops.concat([id_tensor.dense_shape, [1]], axis=0))
if weight_tensor is not None:
check_weight_rank = check_ops.assert_equal(
array_ops.rank(weight_tensor), 2,
data=[
'Column {} expected weight tensor of rank 2.'.format(self.name),
'weight_tensor shape:', array_ops.shape(weight_tensor)])
with ops.control_dependencies([check_weight_rank]):
weight_tensor = sparse_ops.sparse_reshape(
weight_tensor,
shape=array_ops.concat([weight_tensor.dense_shape, [1]], axis=0))
return fc._CategoricalColumn.IdWeightPair(id_tensor, weight_tensor)
def testFeedMismatchedSizesWithInferredDim(self):
with self.session() as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(sp_input, [4, -1])
with self.assertRaisesOpError("requested shape requires a multiple"):
sess.run(sp_output, {sp_input: input_val})
def testFeedDenseReshapeSemantics(self):
with self.session() as sess:
# Compute a random rank-5 initial shape and new shape, randomly sparsify
# it, and check that the output of SparseReshape has the same semantics
# as a dense reshape.
factors = np.array([2] * 4 + [3] * 4 + [5] * 4) # 810k total elements
orig_rank = np.random.randint(2, 7)
orig_map = np.random.randint(orig_rank, size=factors.shape)
orig_shape = [np.prod(factors[orig_map == d]) for d in range(orig_rank)]
new_rank = np.random.randint(2, 7)
new_map = np.random.randint(new_rank, size=factors.shape)
new_shape = [np.prod(factors[new_map == d]) for d in range(new_rank)]
orig_dense = np.random.uniform(size=orig_shape)
orig_indices = np.transpose(np.nonzero(orig_dense < 0.5))
orig_values = orig_dense[orig_dense < 0.5]
new_dense = np.reshape(orig_dense, new_shape)
new_indices = np.transpose(np.nonzero(new_dense < 0.5))
new_values = new_dense[new_dense < 0.5]
sp_input = self._SparseTensorPlaceholder()
input_val = sparse_tensor.SparseTensorValue(orig_indices, orig_values,
orig_shape)
sp_output = sparse_ops.sparse_reshape(sp_input, new_shape)
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices, new_indices)
self.assertAllEqual(output_val.values, new_values)
self.assertAllEqual(output_val.dense_shape, new_shape)
def dense_labels_to_sparse(dense, length):
"""Convert dense labels with sequence lengths to sparse tensor.
Args:
dense: tensor of shape [batch, max_length]
length: int tensor of shape [batch] The length of each sequence in dense.
Returns:
tf.SparseTensor with values only for the valid elements of sequences.
"""
flat_values = array_ops.reshape(dense, [-1])
flat_indices = math_ops.range(
array_ops.shape(flat_values, out_type=dtypes.int64)[0])
mask = array_ops.sequence_mask(length, maxlen=array_ops.shape(dense)[1])
flat_mask = array_ops.reshape(mask, [-1])
indices = array_ops.expand_dims(
array_ops.boolean_mask(flat_indices, flat_mask), 1)
values = array_ops.boolean_mask(flat_values, flat_mask)
sparse = sparse_tensor.SparseTensor(
indices=indices,
values=math_ops.cast(values, dtypes.int32),
dense_shape=array_ops.shape(flat_values, out_type=dtypes.int64))
reshaped = sparse_ops.sparse_reshape(sparse, array_ops.shape(dense))
max_length = math_ops.reduce_max(length)
return sparse_tensor.SparseTensor(
indices=reshaped.indices,
values=reshaped.values,
dense_shape=[
math_ops.cast(reshaped.dense_shape[0], dtypes.int64),
math_ops.cast(max_length, dtypes.int64)
])
def tensors_to_item(self, keys_to_tensors):
"""Maps the given dictionary of tensors to a concatenated list of keypoints.
Args:
keys_to_tensors: a mapping of TF-Example keys to parsed tensors.
Returns:
[time, num_keypoints, 2] tensor of keypoint coordinates, in order [y, x].
Whether the tensor is a SparseTensor or a dense Tensor is determined
by the return_dense parameter. Empty positions in the sparse tensor
are filled with -1.0 values.
"""
coordinates = []
for key in self._full_keys:
value = keys_to_tensors[key]
expanded_dims = array_ops.concat([
math_ops.to_int64(array_ops.shape(value)),
constant_op.constant([1], dtype=dtypes.int64)
], 0)
coordinate = sparse_ops.sparse_reshape(value, expanded_dims)
coordinates.append(coordinate)
keypoints = sparse_ops.sparse_concat(2, coordinates)
if self._return_dense:
keypoints = sparse_ops.sparse_tensor_to_dense(
keypoints, default_value=self._default_value)
return keypoints
def testFeedDenseReshapeSemantics(self):
with self.session(use_gpu=False) as sess:
# Compute a random rank-5 initial shape and new shape, randomly sparsify
# it, and check that the output of SparseReshape has the same semantics
# as a dense reshape.
factors = np.array([2] * 4 + [3] * 4 + [5] * 4) # 810k total elements
orig_rank = np.random.randint(2, 7)
orig_map = np.random.randint(orig_rank, size=factors.shape)
orig_shape = [np.prod(factors[orig_map == d]) for d in range(orig_rank)]
new_rank = np.random.randint(2, 7)
new_map = np.random.randint(new_rank, size=factors.shape)
new_shape = [np.prod(factors[new_map == d]) for d in range(new_rank)]
orig_dense = np.random.uniform(size=orig_shape)
orig_indices = np.transpose(np.nonzero(orig_dense < 0.5))
orig_values = orig_dense[orig_dense < 0.5]
new_dense = np.reshape(orig_dense, new_shape)
new_indices = np.transpose(np.nonzero(new_dense < 0.5))
new_values = new_dense[new_dense < 0.5]
sp_input = self._SparseTensorPlaceholder()
input_val = sparse_tensor.SparseTensorValue(orig_indices, orig_values,
orig_shape)
sp_output = sparse_ops.sparse_reshape(sp_input, new_shape)
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices, new_indices)
self.assertAllEqual(output_val.values, new_values)
self.assertAllEqual(output_val.dense_shape, new_shape)
def testFeedMultipleInferredDims(self):
with self.session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(sp_input, [4, -1, -1])
with self.assertRaisesOpError("only one output dimension may be -1"):
sess.run(sp_output, {sp_input: input_val})
def testFeedMismatchedSizesWithInferredDim(self):
with self.session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(sp_input, [4, -1])
with self.assertRaisesOpError("requested shape requires a multiple"):
sess.run(sp_output, {sp_input: input_val})
def _SparseReduceSumSparseGrad(op, unused_output_indices_grad, out_grad,
unused_output_shape_grad):
"""
Args:
op: the SparseReorder op
unused_output_indices_grad: the incoming gradients of the output indices
out_grad: the incoming gradients of the output values
Returns:
Gradient for each of the 4 input tensors:
(input_indices, input_values, input_shape, reduction_axes)
The gradients for input_indices, reduction_axes and input_shape is None.
"""
# sp_indices = op.inputs[0]
# vals_shape = array_ops.shape(op.inputs[1])
sp_shape = op.inputs[2]
out_shape = op.outputs[2]
output_shape_kept_dims = math_ops.to_int64(
math_ops.reduced_shape(sp_shape, op.inputs[3]))
sp_grad = sparse_tensor.SparseTensor(op.outputs[0], out_grad, out_shape)
sp_grad = sparse_ops.sparse_reshape(sp_grad, output_shape_kept_dims)
#TODO: replace hardcoded 16 with inferred dimension size from sp_shape
# sp_tile = sparse_ops.sparse_concat(2, [sp_grad]*math_ops.to_int64(sp_shape)[2]) #tile gradients along 3rd axis
sp_tile = sparse_ops.sparse_concat(2, [sp_grad] *
128) #tile gradients along 3rd axis
# (sparse_indices, sparse_values, sparse_shape, reduction_axes)
return (None, sp_tile._values, None, None)
def testFeedMultipleInferredDims(self):
with self.session() as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(sp_input, [4, -1, -1])
with self.assertRaisesOpError("only one output dimension may be -1"):
sess.run(sp_output, {sp_input: input_val})
def dense_labels_to_sparse(dense, length):
"""Convert dense labels with sequence lengths to sparse tensor.
Args:
dense: tensor of shape [batch, max_length]
length: int tensor of shape [batch]
The length of each sequence in dense.
Returns:
tf.SparseTensor with values only for the valid elements of sequences.
"""
flat_values = array_ops.reshape(dense, [-1])
flat_indices = math_ops.range(
array_ops.shape(flat_values, out_type=dtypes.int64)[0])
mask = array_ops.sequence_mask(length, maxlen=array_ops.shape(dense)[1])
flat_mask = array_ops.reshape(mask, [-1])
indices = array_ops.expand_dims(
array_ops.boolean_mask(flat_indices, flat_mask), 1)
values = array_ops.boolean_mask(flat_values, flat_mask)
sparse = sparse_tensor.SparseTensor(
indices=indices, values=math_ops.cast(values, dtypes.int32),
dense_shape=array_ops.shape(flat_values, out_type=dtypes.int64))
reshaped = sparse_ops.sparse_reshape(sparse, array_ops.shape(dense))
max_length = math_ops.reduce_max(length)
return sparse_tensor.SparseTensor(
indices=reshaped.indices,
values=reshaped.values,
dense_shape=[
math_ops.cast(reshaped.dense_shape[0], dtypes.int64),
math_ops.cast(max_length, dtypes.int64)])
def batch_reduce_fn(state, value):
padded_value = sparse_tensor.SparseTensor(
indices=value.indices, values=value.values, dense_shape=padded_shape)
reshaped_value = sparse_ops.sparse_reshape(
padded_value,
array_ops.concat(
[np.array([1], dtype=np.int64), padded_value.dense_shape], 0))
return sparse_ops.sparse_concat(0, [state, reshaped_value])
def testFeedMismatchedSizes(self):
with self.session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(sp_input, [4, 7])
with self.assertRaisesOpError(
"Input to reshape is a tensor with 30 dense values"):
sess.run(sp_output, {sp_input: input_val})
def batch_reduce_fn(state, value):
padded_value = sparse_tensor.SparseTensor(
indices=value.indices, values=value.values, dense_shape=padded_shape)
reshaped_value = sparse_ops.sparse_reshape(
padded_value,
array_ops.concat(
[np.array([1], dtype=np.int64), padded_value.dense_shape], 0))
return sparse_ops.sparse_concat(0, [state, reshaped_value])
def testFeedMismatchedSizes(self):
with self.session() as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(sp_input, [4, 7])
with self.assertRaisesOpError(
"Input to reshape is a tensor with 30 dense values"):
sess.run(sp_output, {sp_input: input_val})
def testSameShape(self):
with self.session(use_gpu=False) as sess:
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(input_val, [5, 6])
output_val = sess.run(sp_output)
self.assertAllEqual(output_val.indices, input_val.indices)
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.dense_shape, input_val.dense_shape)
def testSameShape(self):
with self.session() as sess:
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(input_val, [5, 6])
output_val = self.evaluate(sp_output)
self.assertAllEqual(output_val.indices, input_val.indices)
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.dense_shape, input_val.dense_shape)
def testPropagatesFullyKnownDenseShapeWhenShapePartiallyKnown(self):
sp_input = sparse_tensor.SparseTensor.from_value(
self._SparseTensorValue_2x3x4())
self.assertAllEqual((2, 3, 4), sp_input.shape)
sp_output = sparse_ops.sparse_reshape(
sp_input, shape=array_ops.concat(
(constant_op.constant([2], dtype=dtypes.int64),
array_ops.placeholder(dtype=dtypes.int64, shape=[1])),
axis=0))
self.assertAllEqual((2, 3 * 4), sp_output.shape)
def testFeedSameShape(self):
with self.session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(sp_input, [5, 6])
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices, input_val.indices)
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.dense_shape, input_val.dense_shape)
def testFeedSameShapeWithInferredDim(self):
with self.session() as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(sp_input, [-1, 6])
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices, input_val.indices)
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.dense_shape, input_val.dense_shape)
def testFeedSameShape(self):
with self.test_session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(sp_input, [5, 6])
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices, input_val.indices)
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.dense_shape, input_val.dense_shape)
def testUpRank(self):
with self.session(use_gpu=False) as sess:
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(input_val, [2, 3, 5])
output_val = sess.run(sp_output)
self.assertAllEqual(output_val.indices,
np.array([[0, 0, 0], [0, 1, 1], [0, 1, 4], [0, 2, 0],
[1, 1, 0], [1, 1, 1]]))
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.dense_shape, [2, 3, 5])
def testUpRank(self):
with self.session() as sess:
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(input_val, [2, 3, 5])
output_val = self.evaluate(sp_output)
self.assertAllEqual(output_val.indices,
np.array([[0, 0, 0], [0, 1, 1], [0, 1, 4], [0, 2, 0],
[1, 1, 0], [1, 1, 1]]))
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.dense_shape, [2, 3, 5])
def call(self, inputs):
input_shape = array_ops.stack(
(math_ops.reduce_prod(array_ops.shape(inputs)[:-1]),
self.kernel.shape[0]))
output_shape = array_ops.concat(
(array_ops.shape(inputs)[:-1], [self.kernel.shape[1]]), -1)
x = sparse_ops.sparse_reshape(inputs, input_shape)
return array_ops.reshape(
self.activation(
sparse_ops.sparse_tensor_dense_matmul(x, self.kernel) + self.bias),
output_shape)
def testWorksWellWithTfShape(self):
with self.session() as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
shape = array_ops.shape(sp_input) # tf.shape generates int32 output
sp_output = sparse_ops.sparse_reshape(sp_input, shape)
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices, input_val.indices)
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.dense_shape, input_val.dense_shape)
def testWorksWellWithTfShape(self):
with self.session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
shape = array_ops.shape(sp_input) # tf.shape generates int32 output
sp_output = sparse_ops.sparse_reshape(sp_input, shape)
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices, input_val.indices)
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.dense_shape, input_val.dense_shape)
def testFeedUpRankWithInferredDim(self):
with self.session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(sp_input, [2, -1, 5])
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices,
np.array([[0, 0, 0], [0, 1, 1], [0, 1, 4], [0, 2, 0],
[1, 1, 0], [1, 1, 1]]))
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.dense_shape, [2, 3, 5])
def testFeedDownRankWithInferredDim(self):
with self.session() as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_2x3x4()
sp_output = sparse_ops.sparse_reshape(sp_input, [6, -1])
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices,
np.array([[0, 1], [1, 0], [1, 2], [3, 3], [4, 1],
[4, 3], [5, 2]]))
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.dense_shape, [6, 4])
def testFeedUpRankWithInferredDim(self):
with self.session() as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(sp_input, [2, -1, 5])
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices,
np.array([[0, 0, 0], [0, 1, 1], [0, 1, 4], [0, 2, 0],
[1, 1, 0], [1, 1, 1]]))
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.dense_shape, [2, 3, 5])
def testFeedNewShapeSameRank(self):
with self.session() as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(sp_input, [3, 10])
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices,
np.array([[0, 0], [0, 6], [0, 9], [1, 0], [2, 0],
[2, 1]]))
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.dense_shape, [3, 10])
def testFeedDownRankWithInferredDim(self):
with self.session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_2x3x4()
sp_output = sparse_ops.sparse_reshape(sp_input, [6, -1])
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices,
np.array([[0, 1], [1, 0], [1, 2], [3, 3], [4, 1],
[4, 3], [5, 2]]))
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.dense_shape, [6, 4])
def _maybe_reshape_input_tensor(tensor, column_name, output_rank):
"""Reshape the input tensor by the following rule.
1. If `output_rank > input_rank + 1`, raise a `ValueError`.
2. If `output_rank == input_rank + 1`, expand the tensor by one dimension.
3. If `output_rank == input_rank`, do nothing.
4. If `output_rank < input_rank`, flatten the inner dimensions of the tensor.
Args:
tensor: A Tensor or SparseTensor to be reshaped.
column_name: A string name of the feature column for the tensor.
output_rank: the desired rank of the tensor.
Returns:
A reshaped Tensor or SparseTensor.
Raises:
ValueError: if `output_rank > input_rank + 1` for the input tensor.
"""
input_rank = tensor.get_shape().ndims
if input_rank is None and isinstance(tensor,
sparse_tensor_py.SparseTensor):
# Try to get the rank of a sparse tensor by its dense_shape's shape.
input_rank = tensor.dense_shape.get_shape().as_list()[0]
if input_rank is None:
raise ValueError(
'Error while processing column {}. Rank of input Tensor '
'can not be None.'.format(column_name))
if output_rank > input_rank + 1:
raise ValueError(
'Error while processing column {}. Rank of input Tensor '
'({}) should be the same as output_rank ({}). For '
'example, sequence data should typically be 3 '
'dimensional (rank 3) while non-sequence data is '
'typically 2 dimensional (rank 2).'.format(column_name, input_rank,
output_rank))
elif output_rank == input_rank + 1:
# Expand the tensor's shape by 1 dimension.
if isinstance(tensor, sparse_tensor_py.SparseTensor):
output_shape = array_ops.concat([tensor.dense_shape, [1]], 0)
return sparse_ops.sparse_reshape(tensor, output_shape)
else:
reshaped = array_ops.expand_dims(tensor, -1)
# Try to calculate the new shape.
static_shape = tensor.get_shape()
if static_shape is not None and static_shape.dims is not None:
reshaped.set_shape(static_shape.as_list() + [1])
return reshaped
elif output_rank < input_rank:
return layers._inner_flatten(tensor, output_rank) # pylint: disable=protected-access
else:
return tensor
def testFeedNewShapeSameRank(self):
with self.session(use_gpu=False) as sess:
sp_input = self._SparseTensorPlaceholder()
input_val = self._SparseTensorValue_5x6()
sp_output = sparse_ops.sparse_reshape(sp_input, [3, 10])
output_val = sess.run(sp_output, {sp_input: input_val})
self.assertAllEqual(output_val.indices,
np.array([[0, 0], [0, 6], [0, 9], [1, 0], [2, 0],
[2, 1]]))
self.assertAllEqual(output_val.values, input_val.values)
self.assertAllEqual(output_val.dense_shape, [3, 10])
def _MakeAndReshapeTensor(self, tensor_class, original_shape, target_shape):
if tensor_class == "sparse":
ind = np.zeros([0, len(original_shape)]).astype(np.int64)
val = np.array([]).astype(np.float64)
shape = np.array(original_shape).astype(np.int64)
sp_input = sparse_tensor.SparseTensorValue(ind, val, shape)
sp_output = self.evaluate(
sparse_ops.sparse_reshape(sp_input, target_shape))
return sp_output.dense_shape
else:
dense_input = array_ops.zeros(original_shape)
dense_output = self.evaluate(array_ops.reshape(dense_input, target_shape))
return dense_output.shape
def _maybe_reshape_input_tensor(tensor, column_name, output_rank):
"""Reshape the input tensor by the following rule.
1. If `output_rank > input_rank + 1`, raise a `ValueError`.
2. If `output_rank == input_rank + 1`, expand the tensor by one dimension.
3. If `output_rank == input_rank`, do nothing.
4. If `output_rank < input_rank`, flatten the inner dimensions of the tensor.
Args:
tensor: A Tensor or SparseTensor to be reshaped.
column_name: A string name of the feature column for the tensor.
output_rank: the desired rank of the tensor.
Returns:
A reshaped Tensor or SparseTensor.
Raises:
ValueError: if `output_rank > input_rank + 1` for the input tensor.
"""
input_rank = tensor.get_shape().ndims
if input_rank is None and isinstance(tensor, sparse_tensor_py.SparseTensor):
# Try to get the rank of a sparse tensor by its dense_shape's shape.
input_rank = tensor.dense_shape.get_shape().as_list()[0]
if input_rank is None:
raise ValueError('Error while processing column {}. Rank of input Tensor '
'can not be None.'.format(column_name))
if output_rank > input_rank + 1:
raise ValueError('Error while processing column {}. Rank of input Tensor '
'({}) should be the same as output_rank ({}). For '
'example, sequence data should typically be 3 '
'dimensional (rank 3) while non-sequence data is '
'typically 2 dimensional (rank 2).'.format(
column_name, input_rank, output_rank))
elif output_rank == input_rank + 1:
# Expand the tensor's shape by 1 dimension.
if isinstance(tensor, sparse_tensor_py.SparseTensor):
output_shape = array_ops.concat([tensor.dense_shape, [1]], 0)
return sparse_ops.sparse_reshape(tensor, output_shape)
else:
reshaped = array_ops.expand_dims(tensor, -1)
# Try to calculate the new shape.
static_shape = tensor.get_shape()
if static_shape is not None and static_shape.dims is not None:
reshaped.set_shape(static_shape.as_list() + [1])
return reshaped
elif output_rank < input_rank:
return layers._inner_flatten(tensor, output_rank) # pylint: disable=protected-access
else:
return tensor
def testFeedPartialShapes(self):
with self.session(use_gpu=False):
# Incorporate new rank into shape information if known
sp_input = self._SparseTensorPlaceholder()
sp_output = sparse_ops.sparse_reshape(sp_input, [2, 3, 5])
self.assertListEqual(sp_output.indices.get_shape().as_list(), [None, 3])
self.assertListEqual(sp_output.dense_shape.get_shape().as_list(), [3])
# Incorporate known shape information about input indices in output
# indices
sp_input = self._SparseTensorPlaceholder()
sp_input.indices.set_shape([5, None])
sp_output = sparse_ops.sparse_reshape(sp_input, [2, 3, 5])
self.assertListEqual(sp_output.indices.get_shape().as_list(), [5, 3])
self.assertListEqual(sp_output.dense_shape.get_shape().as_list(), [3])
# Even if new_shape has no shape information, we know the ranks of
# output indices and shape
sp_input = self._SparseTensorPlaceholder()
sp_input.indices.set_shape([5, None])
new_shape = array_ops.placeholder(dtypes.int64)
sp_output = sparse_ops.sparse_reshape(sp_input, new_shape)
self.assertListEqual(sp_output.indices.get_shape().as_list(), [5, None])
self.assertListEqual(sp_output.dense_shape.get_shape().as_list(), [None])
def tensors_to_item(self, keys_to_tensors):
tensor = keys_to_tensors[self._tensor_key]
shape = self._shape
if self._shape_key:
shape = keys_to_tensors[self._shape_key]
if isinstance(shape, ops.SparseTensor):
shape = sparse_ops.sparse_tensor_to_dense(shape)
if isinstance(tensor, ops.SparseTensor):
if shape is not None:
tensor = sparse_ops.sparse_reshape(tensor, shape)
tensor = sparse_ops.sparse_tensor_to_dense(tensor, self._default_value)
else:
if shape is not None:
tensor = array_ops.reshape(tensor, shape)
return tensor
def tensors_to_item(self, keys_to_tensors):
tensor = keys_to_tensors[self._tensor_key]
shape = self._shape
if self._shape_key:
shape = keys_to_tensors[self._shape_key]
if isinstance(shape, ops.SparseTensor):
shape = sparse_ops.sparse_tensor_to_dense(shape)
if isinstance(tensor, ops.SparseTensor):
if shape is not None:
tensor = sparse_ops.sparse_reshape(tensor, shape)
tensor = sparse_ops.sparse_tensor_to_dense(tensor,
self._default_value)
else:
if shape is not None:
tensor = array_ops.reshape(tensor, shape)
return tensor
def tensors_to_item(self, keys_to_tensors):
tensor = keys_to_tensors[self._tensor_key]
shape = self._shape
if self._shape_keys:
shape_dims = []
for k in self._shape_keys:
shape_dim = keys_to_tensors[k]
if isinstance(shape_dim, ops.SparseTensor):
shape_dim = sparse_ops.sparse_tensor_to_dense(shape_dim)
shape_dims.append(shape_dim)
shape = array_ops.squeeze(array_ops.pack(shape_dims))
if isinstance(tensor, ops.SparseTensor):
if shape is not None:
tensor = sparse_ops.sparse_reshape(tensor, shape)
tensor = sparse_ops.sparse_tensor_to_dense(tensor, self._default_value)
else:
if shape is not None:
tensor = array_ops.reshape(tensor, shape)
return tensor
def tensors_to_item(self, keys_to_tensors):
tensor = keys_to_tensors[self._tensor_key]
shape = self._shape
if self._shape_keys:
shape_dims = []
for k in self._shape_keys:
shape_dim = keys_to_tensors[k]
if isinstance(shape_dim, sparse_tensor.SparseTensor):
shape_dim = sparse_ops.sparse_tensor_to_dense(shape_dim)
shape_dims.append(shape_dim)
shape = array_ops.reshape(array_ops.stack(shape_dims), [-1])
if isinstance(tensor, sparse_tensor.SparseTensor):
if shape is not None:
tensor = sparse_ops.sparse_reshape(tensor, shape)
tensor = sparse_ops.sparse_tensor_to_dense(tensor, self._default_value)
else:
if shape is not None:
tensor = array_ops.reshape(tensor, shape)
return tensor
def reshape_first_ndims(tensor, first_ndims, new_shape):
"""Reshapes the first n dims of the input `tensor` to `new shape`.
Args:
tensor: The input `Tensor`.
first_ndims: A int denoting the first n dims.
new_shape: A list of int representing the new shape.
Returns:
A reshaped `Tensor`.
"""
assert tensor.get_shape().ndims is None or tensor.get_shape(
).ndims >= first_ndims, (
'Tensor shape is less than {} dims.'.format(first_ndims))
new_shape = array_ops.concat(
[new_shape, array_ops.shape(tensor)[first_ndims:]], 0)
if isinstance(tensor, sparse_tensor.SparseTensor):
return sparse_ops.sparse_reshape(tensor, new_shape)
return array_ops.reshape(tensor, new_shape)
def _structuredRaggedSparseElement(self, structure, shapes, dtype,
padded_shape):
if structure is None:
dense_shape = np.maximum(np.amax(shapes, axis=0), padded_shape)
values = []
for shape in shapes:
dense_to_sparse = self._make_dense_to_sparse_fn(len(shape) == 0) # pylint: disable=g-explicit-length-test
sparse = dense_to_sparse(array_ops.zeros(shape, dtype=dtype))
padded_sparse = sparse_tensor.SparseTensor(sparse.indices,
sparse.values, dense_shape)
reshaped_sparse = sparse_ops.sparse_reshape(
padded_sparse,
array_ops.concat([np.array([1], dtype=np.int64), dense_shape], 0))
values.append(reshaped_sparse)
return sparse_ops.sparse_concat(0, values)
else:
return tuple([
self._structuredRaggedSparseElement(substructure, shapes, dtype,
padded_shape)
for substructure in structure
])
def safe_embedding_lookup_sparse(embedding_weights,
sparse_ids,
sparse_weights=None,
combiner=None,
default_id=None,
name=None,
partition_strategy="div",
max_norm=None):
"""Lookup embedding results, accounting for invalid IDs and empty features.
The partitioned embedding in `embedding_weights` must all be the same shape
except for the first dimension. The first dimension is allowed to vary as the
vocabulary size is not necessarily a multiple of `P`. `embedding_weights`
may be a `PartitionedVariable` as returned by using `tf.get_variable()` with a
partitioner.
Invalid IDs (< 0) are pruned from input IDs and weights, as well as any IDs
with non-positive weight. For an entry with no features, the embedding vector
for `default_id` is returned, or the 0-vector if `default_id` is not supplied.
The ids and weights may be multi-dimensional. Embeddings are always aggregated
along the last dimension.
Args:
embedding_weights: A list of `P` float tensors or values representing
partitioned embedding tensors. Alternatively, a `PartitionedVariable`,
created by partitioning along dimension 0. The total unpartitioned
shape should be `[e_0, e_1, ..., e_m]`, where `e_0` represents the
vocab size and `e_1, ..., e_m` are the embedding dimensions.
sparse_ids: `SparseTensor` of shape `[d_0, d_1, ..., d_n]` containing the
ids. `d_0` is typically batch size.
sparse_weights: `SparseTensor` of same shape as `sparse_ids`, containing
float weights corresponding to `sparse_ids`, or `None` if all weights
are be assumed to be 1.0.
combiner: A string specifying how to combine embedding results for each
entry. Currently "mean", "sqrtn" and "sum" are supported, with "mean"
the default.
default_id: The id to use for an entry with no features.
name: A name for this operation (optional).
partition_strategy: A string specifying the partitioning strategy.
Currently `"div"` and `"mod"` are supported. Default is `"div"`.
max_norm: If not None, all embeddings are l2-normalized to max_norm before
combining.
Returns:
Dense tensor of shape `[d_0, d_1, ..., d_{n-1}, e_1, ..., e_m]`.
Raises:
ValueError: if `embedding_weights` is empty.
"""
if combiner is None:
logging.warn("The default value of combiner will change from \"mean\" "
"to \"sqrtn\" after 2016/11/01.")
combiner = "mean"
if embedding_weights is None:
raise ValueError("Missing embedding_weights %s." % embedding_weights)
if isinstance(embedding_weights, variables.PartitionedVariable):
embedding_weights = list(embedding_weights) # get underlying Variables.
if not isinstance(embedding_weights, list):
embedding_weights = [embedding_weights]
if len(embedding_weights) < 1:
raise ValueError("Missing embedding_weights %s." % embedding_weights)
dtype = sparse_weights.dtype if sparse_weights is not None else None
if isinstance(embedding_weights, variables.PartitionedVariable):
embedding_weights = list(embedding_weights)
embedding_weights = [
ops.convert_to_tensor(w, dtype=dtype) for w in embedding_weights
]
contrib_tensor_util.assert_same_float_dtype(embedding_weights +
[sparse_weights])
with ops.name_scope(name, "embedding_lookup",
embedding_weights + [sparse_ids,
sparse_weights]) as scope:
# Reshape higher-rank sparse ids and weights to linear segment ids.
original_shape = sparse_ids.dense_shape
original_rank_dim = sparse_ids.dense_shape.get_shape()[0]
original_rank = (
array_ops.size(original_shape)
if original_rank_dim.value is None
else original_rank_dim.value)
sparse_ids = sparse_ops.sparse_reshape(sparse_ids, [
math_ops.reduce_prod(
array_ops.slice(original_shape, [0], [original_rank - 1])),
array_ops.gather(original_shape, original_rank - 1)])
if sparse_weights is not None:
sparse_weights = sparse_tensor.SparseTensor(
sparse_ids.indices,
sparse_weights.values, sparse_ids.dense_shape)
# Prune invalid ids and weights.
sparse_ids, sparse_weights = _prune_invalid_ids(sparse_ids, sparse_weights)
# Fill in dummy values for empty features, if necessary.
sparse_ids, is_row_empty = sparse_ops.sparse_fill_empty_rows(sparse_ids,
default_id or
0)
if sparse_weights is not None:
sparse_weights, _ = sparse_ops.sparse_fill_empty_rows(sparse_weights, 1.0)
result = embedding_ops.embedding_lookup_sparse(
embedding_weights,
sparse_ids,
sparse_weights,
combiner=combiner,
partition_strategy=partition_strategy,
name=None if default_id is None else scope,
max_norm=max_norm)
if default_id is None:
# Broadcast is_row_empty to the same shape as embedding_lookup_result,
# for use in Select.
is_row_empty = array_ops.tile(
array_ops.reshape(is_row_empty, [-1, 1]),
array_ops.stack([1, array_ops.shape(result)[1]]))
result = array_ops.where(is_row_empty,
array_ops.zeros_like(result),
result,
name=scope)
# Reshape back from linear ids back into higher-dimensional dense result.
final_result = array_ops.reshape(
result,
array_ops.concat([
array_ops.slice(
math_ops.cast(original_shape, dtypes.int32), [0],
[original_rank - 1]),
array_ops.slice(array_ops.shape(result), [1], [-1])
], 0))
final_result.set_shape(tensor_shape.unknown_shape(
(original_rank_dim - 1).value).concatenate(result.get_shape()[1:]))
return final_result
def reshape_fn(value):
return sparse_ops.sparse_reshape(
value,
array_ops.concat([np.array([1], dtype=np.int64), value.dense_shape], 0))
def safe_embedding_lookup_sparse(
embedding_weights,
sparse_ids,
sparse_weights=None,
combiner="mean",
default_id=None,
name="safe_embedding_lookup_sparse",
partition_strategy=None, # no used
max_norm=None,
return_trainable=False):
""" Provides a dynamic version of `tf.nn.safe_embedding_lookup_sparse`.
Lookup embedding results, accounting for empty features and invalid weights.
Any IDs will be treated as valid include non-positive IDs.
Invalid weights (<= 0) are pruned from input weights, as well as any IDs
with non-positive weight. For an entry with no features, the embedding vector
for `default_id` is returned, or the 0-vector if `default_id` is not supplied.
The ids and weights may be multi-dimensional. Embeddings are always aggregated
along the last dimension.
Args:
embedding_weights: A single `dynamic_embedding.Variable` instance
representing the complete embedding tensor.
sparse_ids: `SparseTensor` of shape `[d_0, d_1, ..., d_n]` containing the
ids. `d_0` is typically batch size.
sparse_weights: `SparseTensor` of same shape as `sparse_ids`, containing
float weights corresponding to `sparse_ids`, or `None` if all weights are
be assumed to be 1.0.
combiner: A string specifying how to combine embedding results for each
entry. Currently "mean", "sqrtn" and "sum" are supported, with "mean" the
default.
default_id: The id to use for an entry with no features.
name: A name for this operation (optional).
partition_strategy: A string specifying the partitioning strategy. Currently
`"div"` and `"mod"` are supported. Default is `"div"`.
max_norm: If not `None`, all embeddings are l2-normalized to max_norm before
combining.
Returns:
combined_embeddings:
A dense `Tensor` of shape `[d_0, d_1, ..., d_{n-1}, e_1, ..., e_m]`.
trainable_wrap:
A TrainableWrapper object used to fill the Optimizers `var_list`
Only provided if `return_trainable` is True.
Raises:
ValueError: if `embedding_weights` is empty.
"""
if embedding_weights is None:
raise ValueError("Missing embedding_weights %s." % embedding_weights)
if embedding_weights.key_dtype != sparse_ids.dtype:
raise TypeError(
"embedding_weights.key_dtype should be same with sparse_ids.dtype: "
"{} vs. {}".format(embedding_weights.key_dtype, sparse_ids.dtype))
weights_dtype = sparse_weights.dtype if sparse_weights is not None else None
if weights_dtype and embedding_weights.value_dtype != weights_dtype:
raise TypeError(
"embedding_weights.value_dtype should be same with sparse_weights.dtype"
": {} vs. {}".format(embedding_weights.value_dtype, weights_dtype))
scope = variable_scope.get_variable_scope()
full_name = scope.name + "/" + name if scope.name else name
with ops.name_scope(full_name + "/"):
# Reshape higher-rank sparse ids and weights to linear segment ids.
original_shape = sparse_ids.dense_shape
original_rank_dim = tensor_shape.dimension_value(
sparse_ids.dense_shape.get_shape()[0])
original_rank = (array_ops.size(original_shape)
if original_rank_dim is None else original_rank_dim)
sparse_ids = sparse_ops.sparse_reshape(sparse_ids, [
math_ops.reduce_prod(
array_ops.slice(original_shape, [0], [original_rank - 1])),
array_ops.gather(original_shape, original_rank - 1)
])
if sparse_weights is not None:
sparse_weights = sparse_tensor.SparseTensor(
sparse_ids.indices, sparse_weights.values,
sparse_ids.dense_shape)
# Prune invalid weights.
if combiner != "sum":
sparse_ids, sparse_weights = _prune_invalid_weights(
sparse_ids, sparse_weights)
# Fill in dummy values for empty features, if necessary.
sparse_ids, is_row_empty = sparse_ops.sparse_fill_empty_rows(
sparse_ids, default_id or 0)
if sparse_weights is not None:
sparse_weights, _ = sparse_ops.sparse_fill_empty_rows(
sparse_weights, 1.0)
result, trainable_ = embedding_lookup_sparse(
embedding_weights,
sparse_ids,
sparse_weights,
combiner=combiner,
partition_strategy=partition_strategy,
name=name + "/embedding_lookup_sparse",
max_norm=max_norm,
return_trainable=True)
if default_id is None:
# Broadcast is_row_empty to the same shape as embedding_lookup_result,
# for use in Select.
is_row_empty = array_ops.tile(
array_ops.reshape(is_row_empty, [-1, 1]),
array_ops.stack([1, array_ops.shape(result)[1]]))
result = array_ops.where(is_row_empty,
array_ops.zeros_like(result),
result,
name="where")
# Reshape back from linear ids back into higher-dimensional dense result.
final_result = array_ops.reshape(
result,
array_ops.concat([
array_ops.slice(math_ops.cast(original_shape, dtypes.int32),
[0], [original_rank - 1]),
array_ops.slice(array_ops.shape(result), [1], [-1])
], 0))
final_result.set_shape(
tensor_shape.unknown_shape(
(tensor_shape.Dimension(original_rank_dim) -
1).value).concatenate(result.get_shape()[1:]))
return (final_result, trainable_) if return_trainable else final_result
def testProvideStaticallyMismatchedSizes(self):
input_val = self._SparseTensorValue_5x6()
sp_input = sparse_tensor.SparseTensor.from_value(input_val)
with self.assertRaisesRegexp(ValueError, "Cannot reshape"):
sparse_ops.sparse_reshape(sp_input, [4, 7])
def reshape_fn(value):
return sparse_ops.sparse_reshape(
value,
array_ops.concat([np.array([1], dtype=np.int64), value.dense_shape], 0))
def testStaticShapeInfoPreserved(self):
sp_input = sparse_tensor.SparseTensor.from_value(
self._SparseTensorValue_5x6())
self.assertAllEqual((5, 6), sp_input.get_shape())
sp_output = sparse_ops.sparse_reshape(sp_input, shape=(1, 5, 2, 3))
self.assertAllEqual((1, 5, 2, 3), sp_output.get_shape())
def testStaticShapeInfoPreservedWithInferredDims(self):
sp_input = sparse_tensor.SparseTensor.from_value(
self._SparseTensorValue_2x3x4())
self.assertAllEqual((2, 3, 4), sp_input.get_shape())
sp_output = sparse_ops.sparse_reshape(sp_input, shape=(2, -1))
self.assertAllEqual((2, 3 * 4), sp_output.get_shape())
def testRaisesIfMoreThanOneInferredDim(self):
sp_input = sparse_tensor.SparseTensor.from_value(
self._SparseTensorValue_2x3x4())
with self.assertRaisesRegexp(ValueError, "At most one dimension can"):
sparse_ops.sparse_reshape(sp_input, shape=(-1, 2, -1))
def testRaisesIfInferredShapeNotPossible(self):
sp_input = sparse_tensor.SparseTensor.from_value(
self._SparseTensorValue_2x3x4())
with self.assertRaisesRegexp(ValueError, "Cannot reshape"):
sparse_ops.sparse_reshape(sp_input, shape=(-1, 7))
