Function Differences with tf.raw_ops.SparseToDense

tf.raw_ops.SparseToDense

tf.raw_ops.SparseToDense(
    sparse_indices,
    output_shape,
    sparse_values,
    default_value,
    validate_indices=True,
    name=None
) -> Tensor

For more information, see tf.raw_ops.SparseToDense.

mindspore.ops.SparseToDense

class mindspore.ops.SparseToDense(
    indices,
    values,
    sparse_shape
) -> Tensor

For more information, see mindspore.ops.SparseToDense.

Differences

TensorFlow: SparseToDense converts a sparse representation of a Tensor to a dense Tensor.

MindSpore: MindSpore API basically implements the same functions as TensorFlow. TensorFlow default_value parameter can specify a default padding value, while MindSpore does not have this parameter, but it can be implemented by calling SparseToDense twice. The principle is to first create a temporary dense Tensor using SparseToDense, then add a default_value consistent with TensorFlow to each element, and finally subtract the default_value from the position set by the indices to get the target result.

Categories	Subcategories	TensorFlow	MindSpore	Differences
Parameters	Parameter 1	sparse_indices	indices	Same function, different parameter names
	Parameter 2	output_shape	sparse_shape	Same function, different parameter names
	Parameter 3	sparse_values	values	Same function, different parameter names
	Parameter 4	default_value	-	MindSpore does not have this parameter, but you can call SparseToDense twice to achieve the same function
	Parameter 5	validate_indices	-	Not involved
	Parameter 6	name	-	Not involved

Code Example 1

The two APIs achieve the same function when 0 is padded by default.

# TensorFlow
import tensorflow as tf

indices = tf.constant([[0, 1], [1, 2], [2, 3]], dtype=tf.int64)
values = tf.constant([1, 2, 3], dtype=tf.float32)
shape = tf.constant([3, 4], dtype=tf.int64)
sparse_tensor = tf.sparse.SparseTensor(indices=indices, values=values, dense_shape=shape)

default_value = tf.constant(0, dtype=tf.float32)
out = tf.raw_ops.SparseToDense(sparse_indices=sparse_tensor.indices,
                                       output_shape=sparse_tensor.dense_shape,
                                       sparse_values=sparse_tensor.values,
                                       default_value=default_value)
print(out)
# tf.Tensor(
# [[0. 1. 0. 0.]
#  [0. 0. 2. 0.]
#  [0. 0. 0. 3.]], shape=(3, 4), dtype=float32)

# MindSpore
import mindspore
import mindspore.ops as ops
from mindspore import Tensor

indices = Tensor([[0, 1], [1, 2], [2, 3]], dtype=mindspore.int64)
values = Tensor([1, 2, 3], dtype=mindspore.float32)
sparse_shape = (3, 4)
out = ops.SparseToDense()(indices, values, sparse_shape)
print(out)
# [[0. 1. 0. 0.]
#  [0. 0. 2. 0.]
#  [0. 0. 0. 3.]]

Code Example 2

TensorFlow default_value parameter can specify a default padding value, while MindSpore implements functions with two calls to SparseToDense. First create a temporary dense Tensor using SparseToDense, then add a default_value consistent with TensorFlow to each element, subtract the default_value from the element at the location specified by indices.

# TensorFlow
import tensorflow as tf

indices = tf.constant([[0, 1], [1, 2], [2, 3]], dtype=tf.int64)
values = tf.constant([1, 2, 3], dtype=tf.float32)
shape = tf.constant([3, 4], dtype=tf.int64)
sparse_tensor = tf.sparse.SparseTensor(indices=indices, values=values, dense_shape=shape)

default_value = tf.constant(5, dtype=tf.float32)
dense_tensor = tf.raw_ops.SparseToDense(sparse_indices=sparse_tensor.indices,
                                       output_shape=sparse_tensor.dense_shape,
                                       sparse_values=sparse_tensor.values,
                                       default_value=default_value)
print(dense_tensor)
# tf.Tensor(
# [[5. 1. 5. 5.]
#  [5. 5. 2. 5.]
#  [5. 5. 5. 3.]], shape=(3, 4), dtype=float32)

# MindSpore
import mindspore
import mindspore.ops as ops
from mindspore import Tensor

indices = Tensor([[0, 1], [1, 2], [2, 3]], dtype=mindspore.int64)
values = Tensor([1, 2, 3], dtype=mindspore.float32)
sparse_shape = (3, 4)
default_value = 5

out_plus_default = ops.SparseToDense()(indices, values, sparse_shape) + default_value

values = Tensor([5, 5, 5], dtype=mindspore.float32)
temp = ops.SparseToDense()(indices, values, sparse_shape)
out = out_plus_default - temp
print(out)
# [[5. 1. 5. 5.]
#  [5. 5. 2. 5.]
#  [5. 5. 5. 3.]]