比较与torch.randperm的功能差异

torch.randperm

class torch.randperm(
    n,
    out=None,
    dtype=torch.int64,
    layout=torch.strided,
    device=None,
    requires_grad=False
)

更多内容详见 torch.randperm。

mindspore.ops.Randperm

class mindspore.ops.Randperm(
    max_length=1,
    pad=-1,
    dtype=mstype.int32
)(n)

更多内容详见 mindspore.ops.Randperm。

使用方式

PyTorch: 返回从0到n-1的整数的随机排列。

MindSpore: 生成从0到n-1的n个随机样本，不重复。如果max_length大于n，最后的max_length-n个元素将用参数值pad填充。

代码示例

import torch
import mindspore.ops as ops
import mindspore as ms

# MindSpore
# The result of every execution is different because this operator will generate n random samples.
randperm = ops.Randperm(max_length=30, pad=-1)
n = ms.Tensor([20], dtype=ms.int32)
output = randperm(n)
print(output)
# Out:
# [15 6 11 19 14 16 9 5 13 18 4 10 8 0 17 2 1 12 3 7
#  -1 -1 -1 -1 -1 -1 -1 -1 -1 -1]

# PyTorch
torch.randperm(30)
# Out:
# tensor([ 1, 25, 20,  0, 26, 16, 21, 27, 12,  7,  8, 15, 14, 23,  4,  3, 17, 11,
#          9, 13,  5,  6,  2, 28, 19, 22, 24, 10, 29, 18])