Pytorch张量根据正态分布进行随机初始化

最新推荐文章于 2024-01-21 22:53:32 发布

飞机火车巴雷特

最新推荐文章于 2024-01-21 22:53:32 发布

阅读量1.1k

点赞数 2

分类专栏： Python相关文章标签： pytorch

本文链接：https://blog.csdn.net/qq_36158230/article/details/123436523

版权

Python相关专栏收录该内容

61 篇文章 1 订阅

订阅专栏

一、需求

对于张量 $x$ ，我们希望能够对其进行随机初始化，且 $x\sim \mathcal{N}(\mu ,\sigma ^{2})$ 。

二、实现

import torch

# 1. 张量根据标准正态分布(mean=0, std=1)进行随机初始化
a = torch.randn(size=(3, 5))
b = torch.randn(size=(2, 4, 4))
# 随机初始化的张量的mean接近于0且std接近于1(由于数据量太少，所以误差会比较大)
print(a, a.mean().item(), a.std().item())
print(b, b.mean().item(), b.std().item())

# 2. 张量根据一般正态分布(mean=mu, std=sigma)进行随机初始化
mu = 2
sigma = 6
c = mu + sigma * torch.randn(size=(3, 5))
d = mu + sigma * torch.randn(size=(2, 4, 4))
# 随机初始化的张量的mean接近于2且std接近于6(由于数据量太少，所以误差会比较大)
print(c, c.mean().item(), c.std().item())
print(d, d.mean().item(), d.std().item())

三、效果(程序执行一次输出结果)

tensor([[-0.1577, -1.1384,  0.3573, -0.3869,  2.2071],
        [-0.2356,  1.7997,  1.6409, -1.3254, -0.7604],
        [ 1.6190, -0.5193,  0.1736, -0.2998, -0.5120]]) 0.1641494482755661 1.1231410503387451
tensor([[[ 0.0238,  0.0624,  1.2002, -0.0963],
         [-0.4617, -0.4194,  0.1404, -0.0497],
         [-0.8695, -1.7731,  1.2776,  1.2151],
         [ 0.6929, -0.7230, -2.0728, -0.6245]],

        [[-0.6560, -0.0338,  1.0010,  0.7865],
         [ 0.2711,  0.9265,  0.3155, -1.9558],
         [-1.9122,  0.6705, -0.7552,  1.1139],
         [-0.2056, -0.8133,  1.4233,  1.5125]]]) -0.024651043117046356 1.0221213102340698
tensor([[  0.4474,   2.9793,   7.1230, -11.5415,  -0.0688],
        [ 13.4291,   4.3203,  -4.9895,   9.2665,   1.7634],
        [  8.6169,   5.1654,  -5.6114,   1.5580,  -2.8541]]) 1.9735734462738037 6.494339942932129
tensor([[[ 6.1913, 12.8018,  7.7551, -0.4036],
         [-9.9380,  1.2322,  7.5808,  4.9575],
         [-3.1395, -8.1253,  5.8119, 14.8277],
         [ 2.8647,  7.0317,  0.9330,  0.8613]],

        [[ 1.7921, -4.0716, -7.3107, -6.9644],
         [ 7.0471, 20.2164,  0.4942,  6.4316],
         [ 2.8614,  5.3229,  6.7565, 10.2164],
         [ 3.1384,  5.6582,  1.5364,  1.6309]]]) 3.312441110610962 6.568889617919922