Function Differences with tf.math.argmin

tf.math.argmin

tf.math.argmin(
    input,
    axis=None,
    output_type=tf.dtypes.int64,
    name=None,
) -> Tensor

For more information, see tf.math.argmin.

mindspore.ops.argmin

mindspore.ops.argmin(x, axis=None, keepdims=False) -> Tensor

For more information, see mindspore.ops.argmin.

Differences

TensorFlow: Return the index of the smallest value of the Tensor along the given dimension. The return value type defaults to tf.int64, and the default is to return the index of the smallest value when axis is 0.

MindSpore: MindSpore API basically implements the same function as TensorFlow. The default return value type is ms.int32, and the default is to return the index of the smallest value when axis is -1.

Categories	Subcategories	TensorFlow	MindSpore	Differences
Input	Single input	input	x	Both are input Tensor. Both do not support zero-dimensional tensor. TensorFlow supports Tensor type and Numpy.ndarray type input, while MindSpore only supports Tensor type input
Parameters	Parameter 1	axis	axis	Same function, different parameter names, different default values
	Parameter 2	output_type	-	Specify the output type. MindSpore does not have this parameter
	Parameter 3	name	-	Not involved
	Parameter 4	-	keepdims	TensorFlow does not have this parameter. MindSpore parameter keepdims is set to True to keep the aggregated dimensions and set to 1

Code Example 1

Whne TensorFlow argmin operator does not explicitly give the axis parameter, the computation result is the index of the minimum value when axis is 0 by default, while MindSpore returns the index of the smallest value when axis is -1 by default. Therefore, in order to get the same result, the mindspore.ops.argmin operator axis is assigned to 0 before the calculation, and to ensure that the output types are the same, use mindspore.ops.Cast operator to convert the result of MindSpore to mindspore.int64.

# TensorFlow
import tensorflow as tf
import numpy as np

x = np.arange(2*3*4).reshape(2, 3, 4).astype(np.float32)
tf_argmin = tf.math.argmin
tf_output = tf.math.argmin(tf.constant(x))
tf_out_np = tf_output.numpy()
print(tf_out_np)
# [[0 0 0 0]
#  [0 0 0 0]
#  [0 0 0 0]]

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

x = np.arange(2*3*4).reshape(2,3,4).astype(np.float32)
axis = 0
ms_argmin = mindspore.ops.argmin
ms_output = ms_argmin(Tensor(x), axis)
ms_cast = mindspore.ops.Cast()
ms_output = ms_cast(ms_output, mindspore.int64)
ms_out_np = ms_output.asnumpy()
print(ms_out_np)
# [[0 0 0 0]
#  [0 0 0 0]
#  [0 0 0 0]]

Code Example 2

The way that TensorFlow and MindSpore parameters are passed in does not affect the function.

# TensorFlow
import tensorflow as tf
import numpy as np

x = np.arange(2*3*4).reshape(2, 3, 4).astype(np.float32)
tf_argmin = tf.math.argmin
axis = 2
tf_output = tf.math.argmin(tf.constant(x), axis)
tf_out_np = tf_output.numpy()
print(tf_out_np)
# [[0 0 0]
#  [0 0 0]]

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

x = np.arange(2*3*4).reshape(2,3,4).astype(np.float32)
axis = 2
ms_argmin = mindspore.ops.argmin
ms_output = ms_argmin(Tensor(x), axis)
ms_cast = mindspore.ops.Cast()
ms_output = ms_cast(ms_output, mindspore.int64)
ms_out_np = ms_output.asnumpy()
print(ms_out_np)
# [[0 0 0]
#  [0 0 0]]

Code Example 3

The TensorFlow parameter output_type is used to specify the output data type, and the default is tf.int64. The default value of the MindSpore parameter output_type is mindspore.int32. To ensure that the two output types are the same, use the mindspore.ops.Cast operator to convert the MindSpore into mindspore.int64. The TensorFlow parameter name is used to define the name of the executed operation and does not affect the result, while MindSpore does not have this parameter.

# TensorFlow
import tensorflow as tf
import numpy as np

x = np.arange(2*3*4).reshape(2, 3, 4).astype(np.float32)
tf_argmin = tf.math.argmin
axis = 1
tf_output = tf.math.argmin(tf.constant(x), axis, name="tf_output")
tf_out_np = tf_output.numpy()
print(tf_out_np)
# [[0 0 0 0]
#  [0 0 0 0]]

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

x = np.arange(2*3*4).reshape(2,3,4).astype(np.float32)
axis = 1
ms_argmin = mindspore.ops.argmin
ms_output = ms_argmin(Tensor(x), axis)
ms_cast = mindspore.ops.Cast()
ms_output = ms_cast(ms_output, mindspore.int64)
ms_out_np = ms_output.asnumpy()
print(ms_out_np)
# [[0 0 0 0]
#  [0 0 0 0]]