Differences with torch.distributions.laplace.Laplace

torch.distributions.laplace.Laplace

torch.distributions.laplace.Laplace(loc, scale) -> Class Instance

For more information, see torch.distributions.laplace.Laplace.

mindspore.ops.standard_laplace

mindspore.ops.standard_laplace(shape, seed=None) -> Tensor

For more information, see mindspore.ops.standard_laplace.

Differences

PyTorch: Create a Laplace distribution instance and call the sample interface of the instance to generate random values that match the Laplace distribution.

MindSpore: Generates random numbers that match the standard Laplace (mean=0, lambda=1) distribution. When loc=0, scale=1 in PyTorch and the sample function input shape is the same as MindSpore, the two APIs achieve the same function.

Categories	Subcategories	PyTorch	MindSpore	Differences
Parameters	Parameter 1	loc	-	MindSpore does not have this parameter and implements loc=0 in PyTorch by default
	Parameter 2	scale	-	MindSpore does not have this parameter and implements scale=1 in PyTorch by default
	Parameter 3	-	shape	This parameter in PyTorch is passed in when the sample interface is called
	Parameter 4	-	seed	Random seeds for the operator layer. PyTorch does not have this parameter

Code Example

Each randomly generated value in PyTorch occupies one dimension, so the innermost layer of the shape passed in MindSpore adds a dimension of length 1, and the two APIs achieve the same function.

# PyTorch
import torch

m = torch.distributions.laplace.Laplace(torch.tensor([0.0]), torch.tensor([1.0]))
shape = (4, 4)
sample = m.sample(shape)
print(tuple(sample.shape))
# (4, 4, 1)

# MindSpore
import mindspore
from mindspore import ops

shape = (4, 4, 1)
output = ops.standard_laplace(shape)
result = output.shape
print(result)
# (4, 4, 1)