mindspore.ops.UnsortedSegmentMax

class mindspore.ops.UnsortedSegmentMax

沿分段计算输入Tensor的最大值。

更多参考详见 mindspore.ops.unsorted_segment_max()。

输入：

• input_x (Tensor) - shape： \((x_1, x_2, ..., x_R)\) 。 数据类型支持float16、float32或int32。

• segment_ids (Tensor) - shape为 \((x_1)\) 的一维张量，值必须是非负数。 数据类型支持int32。

• num_segments (int) - 分段的数量。

输出：

Tensor，若 num_segments 值为 N ，则shape为 \((N, x_2, ..., x_R)\) 。

支持平台：

Ascend GPU CPU

样例：

>>> # case 1: Only have two num_segments, where is 0 and 1, and segment_ids=[0, 1, 1]
>>> # num_segments = 2 indicates that there are two types of segment_id,
>>> # the first number '0' in [0, 1, 1] indicates input_x[0],
>>> # the second number '1' in [0, 1, 1] indicates input_x[1],
>>> # the third number '1' in [0, 1, 1] indicates input_x[2],
>>> # input_x[0], which is [1, 2, 3] will not be compared to other segment_id.
>>> # Only the same segment_id will be compared.
>>> from mindspore import Tensor
>>> from mindspore import ops
>>> import numpy as np
>>> input_x = Tensor(np.array([[1, 2, 3], [4, 5, 6], [4, 2, 1]]).astype(np.float32))
>>> segment_ids = Tensor(np.array([0, 1, 1]).astype(np.int32))
>>> num_segments = 2
>>> unsorted_segment_max = ops.UnsortedSegmentMax()
>>> output = unsorted_segment_max(input_x, segment_ids, num_segments)
>>> print(output)
[[1. 2. 3.]
 [4. 5. 6.]]
>>>
>>> # case 2: The segment_ids=[0, 0, 1, 1].
>>> # [1, 2, 3] will compare with [4, 2, 0],
>>> # and [4, 5, 6] will compare with [4, 2, 1].
>>> input_x = Tensor(np.array([[1, 2, 3], [4, 2, 0], [4, 5, 6], [4, 2, 1]]).astype(np.float32))
>>> segment_ids = Tensor(np.array([0, 0, 1, 1]).astype(np.int32))
>>> num_segments = 2
>>> unsorted_segment_max = ops.UnsortedSegmentMax()
>>> output = unsorted_segment_max(input_x, segment_ids, num_segments)
>>> print(input_x.shape)
    (4, 3)
>>> print(output)
    [[4. 2. 3.]
     [4. 5. 6.]]
>>> # case 3: If the input_x have three dimensions even more, what will happen?
>>> # The shape of input_x is (2, 4, 3),
>>> # and the length of segment_ids should be the same as the first dimension of input_x.
>>> # Because the segment_ids are different, input_x[0] will not be compared to input_x[1].
>>> input_x = Tensor(np.array([[[1, 2, 3], [4, 2, 0], [4, 5, 6], [4, 2, 1]],
...                            [[1, 2, 3], [4, 2, 0], [4, 5, 6], [4, 2, 1]]]).astype(np.float32))
>>> segment_ids = Tensor(np.array([0, 1]).astype(np.int32))
>>> num_segments = 2
>>> unsorted_segment_max = ops.UnsortedSegmentMax()
>>> output = unsorted_segment_max(input_x, segment_ids, num_segments)
>>> print(input_x.shape)
    (2, 4, 3)
>>> print(output)
    [[[1. 2. 3.]
      [4. 2. 0.]
      [4. 5. 6.]
      [4. 2. 1.]]
     [[1. 2. 3.]
      [4. 2. 0.]
      [4. 5. 6.]
      [4. 2. 1.]]]
>>> # case 4: It has the same input with the 3rd case.
>>> # Because num_segments is equal to 2, there are two segment_ids, but currently only one 0 is used.
>>> # the segment_id i is absent in the segment_ids, then output[i] will be filled with
>>> # the smallest possible value of the input_x's type.
>>> segment_ids = Tensor(np.array([0, 0]).astype(np.int32))
>>> output = unsorted_segment_max(input_x, segment_ids, num_segments)
>>> print(output)
    [[[ 1.0000000e+00  2.0000000e+00  3.0000000e+00]
      [ 4.0000000e+00  2.0000000e+00  0.0000000e+00]
      [ 4.0000000e+00  5.0000000e+00  6.0000000e+00]
      [ 4.0000000e+00  2.0000000e+00  1.0000000e+00]]
     [[-3.4028235e+38 -3.4028235e+38 -3.4028235e+38]
      [-3.4028235e+38 -3.4028235e+38 -3.4028235e+38]
      [-3.4028235e+38 -3.4028235e+38 -3.4028235e+38]
      [-3.4028235e+38 -3.4028235e+38 -3.4028235e+38]]]