Function Differences with torch.nn.functional.normalize

torch.nn.functional.normalize

torch.nn.functional.normalize(
    input,
    p=2,
    dim=1,
    eps=1e-12,
    out=None
)

For more information, see torch.nn.functional.normalize.

mindspore.ops.L2Normalize

class mindspore.ops.L2Normalize(
    axis=0,
    epsilon=1e-4
)(input_x)

For more information, see mindspore.ops.L2Normalize.

Differences

PyTorch: Supports using the LP paradigm by specifying the parameter p.

MindSpore：Only L2 paradigm is supported.

Code Example

import mindspore as ms
import mindspore.ops as ops
import torch
import numpy as np

# In MindSpore, you can directly pass data into the function, and the default dimension is 0.
l2_normalize = ops.L2Normalize()
input_x = ms.Tensor(np.array([1.0, 2.0, 3.0]), ms.float32)
output = l2_normalize(input_x)
print(output)
# Out：
# [0.2673 0.5345 0.8018]

# In torch, parameter p should be set to determine it is a lp normalization, and the default dimension is 1.
input_x = torch.tensor(np.array([1.0, 2.0, 3.0]))
outputL2 = torch.nn.functional.normalize(input=input_x, p=2, dim=0)
outputL3 = torch.nn.functional.normalize(input=input_x, p=3, dim=0)
print(outputL2)
print(outputL3)
# Out：
# tensor([0.2673, 0.5345, 0.8018], dtype=torch.float64)
# tensor([0.3029, 0.6057, 0.9086], dtype=torch.float64)