Question

我有一个脚本，正在使用Keras初始化和训练神经网络，因此我在测试和优化代码时为100％可再现的训练结果初始化了随机种子。同时，我还一直在使用importlib.reload（）函数重新加载我在开发过程中要更改的所有自定义模块。但是问题是我在ipython会话中第一次调用脚本时的随机顺序可能与随后的时间不同。我找到了一个解决方案（请参阅下文），但看起来笨拙而笨拙。我想知道是否有一种更有效或更Python的方式。还是我应该完全不同地处理随机种子？

import sys
modulesUnderDevelopment = ['common_config', 'cnn_tools', 'prep_dataset'] 
moduleShorthand = ['cc', 'ct', 'pd']
reloadTracker = dict()
#record which modules have not yet been imported, before we do any importing.  
for moduleName, shorthand in zip(modulesUnderDevelopment, moduleShorthand):
    reloadTracker[moduleName] = {'imported':moduleName in sys.modules, 'shorthand': shorthand}

#import modules.  Anything that's already been imported won't be imported by these lines,
# so any randomness that occurs during import will NOT occur if the module was 
# already imported.  HOWEVER, if the module has not been imported, any randomness 
# in the module import will be executed.
print('imports are beginning')
import common_config as cc
import cnn_tools as ct
import prep_dataset as pd
print('imports are ending')

import importlib
#now reload the key modules that I'm optimizing and regularly changing.  These reloads will
# ALWAYS happen, so the randomness in them will always be executed.  
print('reloads are beginning')
for key, item in reloadTracker.items():
    if not item['imported']:
        continue
    if key == 'common_config':
        importlib.reload(cc)
    elif key == 'cnn_tools':
        importlib.reload(ct)
    elif key == 'prep_dataset':
        importlib.reload(pd)
print('reloads are complete')

同时，我的模块包含如下代码：

## prep_dataset ##
import numpy as np
print('starting prep_dataset -- {}'.format(np.random.random()))

## cnn_tools ##
import numpy as np
print('starting cnn_tools -- {}'.format(np.random.random()))

我第一次启动ipython并运行此脚本时，将打印以下内容：

imports are beginning
starting cnn_tools -- 0.5507979025745755
Using TensorFlow backend.
starting prep_dataset -- 0.7081478226181048
imports are ending
reloads are beginning
reloads are complete

随后，它将显示以下内容：

imports are beginning
imports are ending
reloads are beginning
starting cnn_tools -- 0.5507979025745755
starting prep_dataset -- 0.7081478226181048
reloads are complete

所以这解决了我的问题-随机数流两次都相同。但是，它看起来确实很笨拙……相反，如果我不经历所有这些麻烦，仅在导入后重新加载，我第一次运行脚本时会得到不同的结果-流中有2个额外的随机数。

starting cnn_tools -- 0.5507979025745755
Using TensorFlow backend.
starting prep_dataset -- 0.7081478226181048
reloads are beginning
starting cnn_tools -- 0.2909047389129443
starting prep_dataset -- 0.510827605197663
reloads are complete

如何仅在当前执行的脚本不只是加载模块的情况下重新加载模块

0 个答案: