实践：PyTorch 数据集随机值

小咪咪 2022-09-12 03:00 125阅读 0赞

点击上方“**机器学习与生成对抗网络**”，关注星标

获取有趣、好玩的前沿干货！

作者 Elvanth@知乎  
https://zhuanlan.zhihu.com/p/377155682  
文仅交流，侵删

一个快捷的解决方案：

def worker_init_fn(worker_id):
        worker_seed = torch.initial_seed() % 2**32
        np.random.seed(worker_seed)
        random.seed(worker_seed)
    
    ds = DataLoader(ds, 10, shuffle=False, num_workers=4, worker_init_fn=worker_init_fn)

## 01 关于pytorch数据集随机种子的基本认识 ##

在pytorch中random、torch.random等随机值产生方法一般没有问题，只有少数工人运行也可以保障其不同的最终值.

np.random.seed 会出现问题的原因是，当多处理采用 fork 方式产生子进程时，numpy 不会对不同的子进程产生不同的随机值.

换言之，当没有多处理使用时，numpy 不会出现随机种子的不同的问题；实验代码的可复现性要求**一个是工人种子** ,即工人内包括numpy，random，torch.random所有的随机表现；**另一个是Base** ,即程序运行后的初始随机值，其可以通过以下两种方式产生

1.  torch.manual\_seed(base\_seed)
2.  由特定的seed generator设置

generator = torch. Generator()
    g.manual_seed(base_seed)
    DataLoader(dataset, ..., generator=generator)

使用spawn模式可以斩断以上所有烦恼.

## 02 直接在网上搜这个问题会得到什么答案 ##

参考很多的解决方案时，往往会提出以下功能：

def worker_init_fn(worker_id):
        np.random.seed(np.random.get_state()[1][0] + worker_id)

让我们看看它的输出结果：  
（第0，3列是索引，第1，4列是np.random的结果，第2，5列是random.randint的结果）

epoch 0
    tensor([[    0,  5125, 13588,     0, 15905, 23182],
            [    1,  7204, 19825,     0, 13653, 25225]])
    tensor([[    2,  1709, 11504,     0, 12842, 23238],
            [    3,  5715, 14058,     0, 15236, 28033]])
    tensor([[    4,  1062, 11239,     0, 10142, 29869],
            [    5,  6574, 15672,     0, 19623, 25600]])
    ============================================================
    epoch 1
    tensor([[    0,  5125, 18134,     0, 15905, 28990],
            [    1,  7204, 13206,     0, 13653, 25106]])
    tensor([[    2,  1709, 15512,     0, 12842, 29703],
            [    3,  5715, 14201,     0, 15236, 27696]])
    tensor([[    4,  1062, 13994,     0, 10142, 23411],
            [    5,  6574, 18532,     0, 19623, 21744]])
    ============================================================

假设上述方案对一个时代内可以防止不同的工人出现随机值相同的情况，但**不同的时代之间，其最终的随机种子仍然是不变的。**

## 03 那应该如何解决 ##

### 来自pytorch官方的解决方案： ###

https://github.com/pytorch/pytorch/pull/56488\#issuecomment-825128350

def worker_init_fn(worker_id):
        worker_seed = torch.initial_seed() % 2**32
        np.random.seed(worker_seed)
        random.seed(worker_seed)
     
    ds = DataLoader(ds, 10, shuffle=False, num_workers=4, worker_init_fn=worker_init_fn)

### 来自numpy.random原作者的解决方案： ###

https://github.com/pytorch/pytorch/issues/5059\#issuecomment-817392562

def worker_init_fn(id):
        process_seed = torch.initial_seed()
        # Back out the base_seed so we can use all the bits.
        base_seed = process_seed - id
        ss = np.random.SeedSequence([id, base_seed])
        # More than 128 bits (4 32-bit words) would be overkill.
        np.random.seed(ss.generate_state(4))
     
    ds = DataLoader(ds, 10, shuffle=False, num_workers=4, worker_init_fn=worker_init_fn)

### 一个更简单但不保证正确性的解决方案： ###

def worker_init_fn(worker_id):
        np.random.seed((worker_id + torch.initial_seed()) % np.iinfo(np.int32).max)
    
    ds = DataLoader(ds, 10, shuffle=False, num_workers=4, worker_init_fn=worker_init_fn)

### 04 附上可运行的完整文件 ###

import numpy as np
    import random
    import torch
    
    # np.random.seed(0)
    
    class Transform(object):
        def __init__(self):
            pass
    
        def __call__(self, item = None):
            return [np.random.randint(10000, 20000), random.randint(20000,30000)]
    
    class RandomDataset(object):
        def __init__(self):
            pass
    
        def __getitem__(self, ind):
            item = [ind, np.random.randint(1, 10000), random.randint(10000, 20000), 0]
            tsfm =Transform()(item)
            return np.array(item + tsfm)
        def __len__(self):
            return 20
    
    from torch.utils.data import DataLoader
    
    def worker_init_fn(worker_id):
        np.random.seed(np.random.get_state()[1][0] + worker_id)
    
    ds = RandomDataset()
    ds = DataLoader(ds, 10, shuffle=False, num_workers=4, worker_init_fn=worker_init_fn)
    
    for epoch in range(2):
        print("epoch {}".format(epoch))
        np.random.seed()
        for batch in ds:
            print(batch)

**猜您喜欢：**

**[等你着陆！【GAN生成对抗网络】知识星球！][GAN]  **

**[CVPR 2021 | GAN的说话人驱动、3D人脸论文汇总][CVPR 2021 _ GAN_3D]**

**[CVPR 2021 | 图像转换 今如何？几篇GAN论文][CVPR 2021 _ _ _GAN]**

[**【CVPR 2021】通过GAN提升人脸识别的遗留难题**][CVPR 2021_GAN]

[**CVPR 2021生成对抗网络GAN部分论文汇总**][CVPR 2021_GAN 1]

[**经典GAN不得不读：StyleGAN**][GAN_StyleGAN]

[**最新最全20篇！基于 StyleGAN 改进或应用相关论文**][20_ StyleGAN]

**[超100篇！CVPR 2020最全GAN论文梳理汇总！][100_CVPR 2020_GAN]**

[**附下载 | 《Python进阶》中文版**][_ _Python]

[**附下载 | 经典《Think Python》中文版**][_ _Think Python]

[**附下载 | 《Pytorch模型训练实用教程》**][_ _Pytorch]

[**附下载 | 最新2020李沐《动手学深度学习》**][_ _2020]

**[附下载 | 《可解释的机器学习》中文版][Link 1]**

[**附下载 |《TensorFlow 2.0 深度学习算法实战》**][_TensorFlow 2.0]

**[附下载 | 超100篇！CVPR 2020最全GAN论文梳理汇总！][100_CVPR 2020_GAN]**

**[附下载 |《计算机视觉中的数学方法》分享][Link 2]**

![642460228e904adc8453d17dd6a5524e.png][]

[GAN]: http://mp.weixin.qq.com/s?__biz=MzU5MTgzNzE0MA==&mid=2247491056&idx=1&sn=535829ad04b094f59b2b38eb85cdbb38&chksm=fe29bf4bc95e365d4f53676d3aa2b80299375e55f9bab3bc7f8ad9405e7ada61d7ac00abb4bf&scene=21#wechat_redirect
[CVPR 2021 _ GAN_3D]: http://mp.weixin.qq.com/s?__biz=MzU5MTgzNzE0MA==&mid=2247492417&idx=1&sn=292be5e5fab52c5df5c94508b898da5a&chksm=fe2a41fac95dc8ec752351699318742fa502dfc945ce072567bb4bf0c59ce83049dc58c1a57d&scene=21#wechat_redirect
[CVPR 2021 _ _ _GAN]: http://mp.weixin.qq.com/s?__biz=MzU5MTgzNzE0MA==&mid=2247491779&idx=1&sn=f0a601ca66c01361709e6af8485f629b&chksm=fe2a4278c95dcb6ebd0ba346a11dc195a0ac4f7830f3a7cf69fcffc7057fa1bf3eaeb9386300&scene=21#wechat_redirect
[CVPR 2021_GAN]: http://mp.weixin.qq.com/s?__biz=MzU5MTgzNzE0MA==&mid=2247491634&idx=1&sn=427cb6bfd6868662f2b5ae1ac37f6072&chksm=fe2a4289c95dcb9f047c18e63b3be9d78e5e75f1d94dc80be34686ce6b4f22cdee2d7cc7f2ae&scene=21#wechat_redirect
[CVPR 2021_GAN 1]: http://mp.weixin.qq.com/s?__biz=MzU5MTgzNzE0MA==&mid=2247490911&idx=1&sn=671926492c3b36e3350db4963f7fcb16&chksm=fe29bfe4c95e36f275332b31b993532b98d52acc76d26c15e2321de4cf52141510bee3ff96af&scene=21#wechat_redirect
[GAN_StyleGAN]: http://mp.weixin.qq.com/s?__biz=MzU5MTgzNzE0MA==&mid=2247490656&idx=2&sn=bbe156fef077f23263b57e506ad43314&chksm=fe29bedbc95e37cd709f583372918478ad8d3cff5309a85f3a940e7c0356050278f89d7e1c79&scene=21#wechat_redirect
[20_ StyleGAN]: http://mp.weixin.qq.com/s?__biz=MzU5MTgzNzE0MA==&mid=2247490498&idx=1&sn=324cfbbde362a728aa5336f4b200f7a3&chksm=fe29b979c95e306f03b134809b16009d06aa0694b03da3368f84625de3700e6a56221d49d1bf&scene=21#wechat_redirect
[100_CVPR 2020_GAN]: http://mp.weixin.qq.com/s?__biz=MzU5MTgzNzE0MA==&mid=2247486455&idx=1&sn=94156f53845cdf07e859e343862be745&chksm=fe29a94cc95e205ae880cb9518cb16a54c0abde69e5a8654f7fedf70a0cc990775265bf28a1c&scene=21#wechat_redirect
[_ _Python]: http://mp.weixin.qq.com/s?__biz=MzU5MTgzNzE0MA==&mid=2247485709&idx=2&sn=55a51c98ab48f393ba305c082ba0e4fe&chksm=fe29abb6c95e22a04d1eb019f964c8ae97b2f3af98c22f730f5d5ed2aea9fdb320b00c688967&scene=21#wechat_redirect
[_ _Think Python]: http://mp.weixin.qq.com/s?__biz=MzU5MTgzNzE0MA==&mid=2247485091&idx=2&sn=f569d5d70c2e6c9cc567fbc5147787a4&chksm=fe29a418c95e2d0e9a3de46e0e4d1b8d4947154d8639418a511a3091300aa31699932046556e&scene=21#wechat_redirect
[_ _Pytorch]: http://mp.weixin.qq.com/s?__biz=MzU5MTgzNzE0MA==&mid=2247485635&idx=2&sn=90608bce32fe353cf3c64b39e3406bc3&chksm=fe29aa78c95e236e380c9fd2ae79917d7451ccdd4803c5e6047a16f921d1b61d18efc79ada7c&scene=21#wechat_redirect
[_ _2020]: http://mp.weixin.qq.com/s?__biz=MzU5MTgzNzE0MA==&mid=2247484959&idx=2&sn=563999cc8d48498beed00270844ceec4&chksm=fe29a4a4c95e2db2aab16bdcefa78795bbb7efbc70b533bb197d73f617558f2f8a523120c6e3&scene=21#wechat_redirect
[Link 1]: http://mp.weixin.qq.com/s?__biz=MzU5MTgzNzE0MA==&mid=2247484929&idx=1&sn=523220858287ed3e6105919703c8e833&chksm=fe29a4bac95e2dac758c4aac10a3aac49eb062bc0829d4ef2ebc8e6d5069d246385c5084a2b4&scene=21#wechat_redirect
[_TensorFlow 2.0]: http://mp.weixin.qq.com/s?__biz=MzU5MTgzNzE0MA==&mid=2247485195&idx=2&sn=e3387067db29d1759755c6c2ea96c870&chksm=fe29a5b0c95e2ca6599d4a91d5d40900566b7525de3dd85259e7a6c173aa26aa7ea3641d1aaf&scene=21#wechat_redirect
[Link 2]: http://mp.weixin.qq.com/s?__biz=MzU5MTgzNzE0MA==&mid=2247485492&idx=2&sn=c4fb135de02cd52f06bc61b0e79476e4&chksm=fe29aa8fc95e23993e057fb0dbe09f2810ea07184a439c173f404a7ab657cadcde3d24efa52a&scene=21#wechat_redirect
[642460228e904adc8453d17dd6a5524e.png]: /images/20220828/aa86054a9992485792c49dca7446e912.png