比较与torch.flatten的功能差异

torch.flatten

torch.flatten(
    input,
    start_dim=0,
    end_dim=-1
)

更多内容详见torch.flatten。

mindspore.ops.Flatten

class mindspore.ops.Flatten(*args, **kwargs)(input_x)

更多内容详见mindspore.ops.Flatten。

使用方式

PyTorch：支持指定维度对元素进行展开。

MindSpore：仅支持保留第0维元素，对其余维度的元素进行展开。

代码示例

import mindspore
from mindspore import Tensor
import mindspore.ops as ops
import torch
import numpy as np

# In MindSpore, only the 0th dimension will be reserved and the rest will be flattened.
input_tensor = Tensor(np.ones(shape=[1, 2, 3, 4]), mindspore.float32)
flatten = ops.Flatten()
output = flatten(input_tensor)
print(output.shape)
# Out：
# (1, 24)

# In torch, the dimension to reserve will be specified and the rest will be flattened.
input_tensor = torch.Tensor(np.ones(shape=[1, 2, 3, 4]))
output1 = torch.flatten(input=input_tensor, start_dim=1)
print(output1.shape)
# Out：
# torch.Size([1, 24])

input_tensor = torch.Tensor(np.ones(shape=[1, 2, 3, 4]))
output2 = torch.flatten(input=input_tensor, start_dim=2)
print(output2.shape)
# Out：
# torch.Size([1, 2, 12])