Function Differences with tf.math.erf

tf.math.erf

tf.math.erf(x, name=None) -> Tensor

For more information, see tf.math.erf.

mindspore.ops.erf

mindspore.ops.erf(x) -> Tensor

For more information, see mindspore.ops.erf.

Differences

TensorFlow: Compute the Gaussian error function for x element-wise i.e. \( \operatorname{erf}(x)=\frac{2}{\sqrt{\pi}} \int_{0}^{x} e^{-t^{2}} d t \) .

MindSpore: MindSpore API basically implements the same function as TensorFlow, but there are differences in the size of the supported dimensions.

Categories	Subcategories	TensorFlow	MindSpore	Differences
Parameters	Parameter 1	x	x	Same function, difference in size of supported dimensions
	Parameter 2	name	-	Not involved

Code Example 1

TensorFlow does not limit the dimension of x, while MindSpore supports dimensions of x that must be less than 8. When the dimension of x is less than 8, the two APIs achieve the same function and have the same usage.

# TensorFlow
import tensorflow as tf
import numpy as np

x_ = np.ones((1, 1, 1, 1, 1, 1, 1))
x = tf.convert_to_tensor(x_, dtype=tf.float32)
out = tf.math.erf(x).numpy()
print(out)
# [[[[[[[0.8427007]]]]]]]

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

x_ = np.ones((1, 1, 1, 1, 1, 1, 1))
x = Tensor(x_, mindspore.float32)
out = ops.erf(x)
print(out)
# [[[[[[[0.8427007]]]]]]]

Code Example 2

When the dimension of x is more than or equal to 8, the same function can be achieved by API group sum. Use ops.reshape to reduce the dimension of x to 1, then call ops.erf to compute it, and finally use ops.reshape again to up-dimension the obtained result according to the original dimension of x.

# TensorFlow
import tensorflow as tf
import numpy as np

x_ = np.ones((1, 1, 1, 1, 1, 1, 1, 1))
x = tf.convert_to_tensor(x_, dtype=tf.float32)
out = tf.math.erf(x).numpy()
print(out)
# [[[[[[[[0.8427007]]]]]]]]

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

x_ = np.ones((1, 1, 1, 1, 1, 1, 1, 1))
x = Tensor(x_, mindspore.float32)
x_reshaped = ops.reshape(x, (-1,))
out_temp = ops.erf(x_reshaped)
out = ops.reshape(out_temp, x.shape)
print(out)
# [[[[[[[[0.8427007]]]]]]]]