I have a application that does a certain experiment 1000 times (multi-threaded, so that multiple experiments are done at the same time). Every experiment needs appr. 50.000 random.random() calls.
What is the best approach to get this really random. I could copy a random object to every experiment and do than a jumpahead of 50.000 * expid. The documentation suggests that jumpahead(1) already scrambles the state, but is that really true?
Or is there another way to do this in 'the best way'?
(No, the random numbers are not used for security, but for a metropolis hasting algorithm. The only requirement is that the experiments are independent, not whether the random sequence is somehow predictable or so)
I could copy a random object to every experiment and do than a jumpahead of 50.000 * expid.
Approximately correct. Each thread gets their own Random instance.
Seed all of them to the same seed value. Use a constant to test, use /dev/random when you "run for the record".
Edit. Outside Python and in older implementations, use jumpahead( 50000 * expid ) to avoid the situation where two generators wind up with a parallel sequences of values. In any reasonably current (post 2.3) Python, jumpahead is no longer linear and using expid is sufficient to scramble the state.
You can't simply do jumpahead(1) in each thread, since that will assure they are synchronized. Use jumpahead( expid ) to assure each thread is distinctly scrambled.
The documentation suggests that jumpahead(1) already scrambles the state, but is that really true?
Yes, jumpahead does indeed "scramble" the state. Recall that for a given seed you get one -- long -- but fixed sequence of pseudo-random numbers. You're jumping ahead in this sequence. To pass randomness tests, you must get all your values from this one sequence.
Edit. Once upon a time, jumpahead(1) was limited. Now jumpahead(1) really does a larger scrambling. The scrambling, however, is deterministic. You can't simply do jumpahead(1) in each thread.
If you have multiple generators with different seeds, you violate the "one sequence from one seed" assumption and your numbers aren't going to be as random as if you get them from a single sequence.
If you only jumphead 1, you'll may be getting parallel sequences which will might be similar. [This similarity might not be detectable; theoretically, there's a similarity.]
When you jumpahead 50,000, you assure that you follow the 1-sequence-1-seed premise. You also assure that you won't have adjacent sequences of numbers in two experiments.
Finally, you also have repeatability. For a given seed, you get consistent results.
Same jumpahead: Not Good.
>>> y=random.Random( 1 )
>>> z=random.Random( 1 )
>>> y.jumpahead(1)
>>> z.jumpahead(1)
>>> [ y.random() for i in range(5) ]
[0.99510321786951772, 0.92436920169905545, 0.21932404923057958, 0.20867489035315723, 0.91525579001682567]
>>> [ z.random() for i in range(5) ]
[0.99510321786951772, 0.92436920169905545, 0.21932404923057958, 0.20867489035315723, 0.91525579001682567]
You shouldn't use that function. There is no proof it can work on Mersenne Twister generator. Indeed, it was removed from Python 3 for that reason.
For more information about generation of pseudo-random numbers on parallel environments, see this article from David Hill.
jumpahead(1) is indeed sufficient (and identical to jumpahead(50000) or any other such call, in the current implementation of random -- I believe that came in at the same time as the Mersenne Twister based implementation). So use whatever argument fits in well with your programs' logic. (Do use a separate random.Random instance per thread for thread-safety purposes of course, as your question already hints).
(random module generated numbers are not meant to be cryptographically strong, so it's a good thing that you're not using for security purposes;-).
Per the random module docs at python.org:
"You can instantiate your own instances of Random to get generators that don’t share state."
And there's also a relevant-looking note on jumpahead, as you mention. But the guarantees there are kind of vague. If the calls to OS-provided randomness aren't so expensive as to dominate your running time, I'd skip all the subtlety and do something like:
randoms = [random.Random(os.urandom(4)) for _ in range(num_expts)]
If num_expts is ~1000, then you're unlikely to have any collisions in your seed (birthday paradox says you need about 65000 experiments before there's a >50% probability that you have a collision). If this isn't good enough for you or if the number of experiments is more like 100k instead of 1k, then I think it's reasonable to follow this up with
for idx, r in enumerate(randoms):
r.jumpahead(idx)
Note that I don't think it will work to just make your seed longer (os.urandom(8), for example), since the random docs state that the seed must be hashable, and so on a 32-bit platform you're only going to get at most 32 bits (4 bytes) of useful entropy in your seed.
This question piqued my curiosity, so I went and looked at the code implementing the random module. I am definitely not a PRNG expert, but it does seem like slightly differing values of n in jumpahead(n) will lead to markedly different Random instance states. (Always scary to contradict Alex Martelli, but the code does use the value of n when shuffling the random state).
Related
I have an understanding (on a very high level) about the usage of seed to generate the random numbers. so setting a particular seed prior to generating random numbers would result in generating the same numbers at each call.
I am assuming restoring the state by
import random
test123 = random.getstate()
random.setstate(test123)
would also result in the replication on random numbers by some similar process
i am looking for an understanding in their similarities and differences. For example: are setting the seed to some value and restoring the internal state of the generator via random.setstate(random.getstate()) certain methods in every situation for replicating random numbers?
there is not much documented about them that i could find
The seed is a simple initial value that you can pass from which the generator will be initialised. The state on the other hand is the full current internal state of the generator. Yes, setting a seed and setting a state are fundamentally the same thing, in that it allows you to replay a specific sequence of pseudo-random numbers. However, internally the generator doesn't work with simple values like a seed. What the get/setstate methods allow you to do is to programmatically restore a state from a running generator; while a seed value is rather something you'd supply as an argument from a configuration file for example.
Let's talk about use cases: you have some algorithm which involves a random value, and you want to test that algorithm. For it to be testable, it must be reproducible. That means you need to be able to control the random element within your algorithm. That's where PRNGs come in, they're predictable and repeatable (unlike true RNGs). You would write some test cases which specify a simple seed to set the PRNG into a specific state. Seeds are short readable values you can easily write into test cases.
Now, if you want to interrupt your test case at some point to inspect it, and then perhaps take a step back and rerun one specific step (whether manually or programmatically), you'll want to get the specific state of the PRNG to be able to reset it to that specific state later. Now, there's no getseed method on the PRNG because it doesn't internally work with values like the seed you first supplied; but it has a getstate method which serves the same purpose, its return value is just more complex.
Is there any Python complexity reference? In cppreference, for example, for many functions (such as std::array::size or std::array::fill) there's a complexity section which describes their running complexity, in terms of linear in the size of the container or constant.
I would expect the same information to appear in the python website, perhaps, at least for the CPython implementation. For example, in the list reference, in list.insert I would expect to see complexity: linear; I know this case (and many other container-related operations) is covered here, but many other cases are not. Here are a few examples:
What is the complexity of tuple.__le__? It seems like when comparing two tuples of size n, k, the complexity is about O(min(n,k)) (however, for small n's it looks different).
What is the complexity of random.shuffle? It appears to be O(n). It also appears that the complexity of random.randint is O(1).
What is the complexity of the __format__ method of strings? It appears to be linear in the size of the input string; however, it also grows when the number of relevant arguments grow (compare ("{0}"*100000).format(*(("abc",)*100000)) with ("{}"*100000).format(*(("abc",)*100000))).
I'm aware that (a) each of these questions may be answered by itself, (b) one may look at the code of these modules (even though some are written in C), and (c) StackExchange is not a python mailing list for user requests. So: this is not a doc-feature request, just a question of two parts:
Do you know if such a resource exists?
If not, do you know what is the place to ask for such, or can you suggest why I don't need such?
CPython is pretty good about its algorithms, and the time complexity of an operation is usually just the best you would expect of a good standard library.
For example:
Tuple ordering has to be O(min(n,m)), because it works by comparing element-wise.
random.shuffle is O(n), because that's the complexity of the modern Fisher–Yates shuffle.
.format I imagine is linear, since it only requires one scan through the template string. As for the difference you see, CPython might just be clever enough to cache the same format code used twice.
The docs do mention time complexity, but generally only when it's not what you would expect — for example, because a deque is implemented with a doubly-linked list, it's explicitly mentioned as having O(n) for indexing in the middle.
Would the docs benefit from having time complexity called out everywhere it's appropriate? I'm not sure. The docs generally present builtins by what they should be used for and have implementations optimized for those use cases. Emphasizing time complexity seems like it would either be useless noise or encourage developers to second-guess the Python implementation itself.
We have a very simple program (single-threaded) where we we do a bunch of random sample generation. For this we are using several calls of the numpy random functions (like normal or random_sample). Sometimes the result of one random call determines the number of times another random function is called.
Now I want to set a seed in the beginning s.th. multiple runs of my program should yield the same result. For this I'm using an instance of the numpy class RandomState. While this is the case in the beginning, at some time the results become different and this is why I'm wondering.
When I am doing everything correctly, having no concurrency and thereby a linear call of the functions AND no other random number generator involded, why does it not work?
Okay, David was right. The PRNGs in numpy work correctly. Throughout every minimal example I created, they worked as they are supposed to.
My problem was a different one, but finally I solved it. Do never loop over a dictionary within a deterministic algorithm. It seems that Python orders the items arbitrarily when calling the .item() function for getting in iterator.
So I am not that disappointed that this was this kind of error, because it is a useful reminder of what to think about when trying to do reproducible simulations.
If reproducibility is very important to you, I'm not sure I'd fully trust any PRNG to always produce the same output given the same seed. You might consider capturing the random numbers in one phase, saving them for reuse; then in a second phase, replay the random numbers you've captured. That's the only way to eliminate the possibility of non-reproducibility -- and it solves your current problem too.
I have a rather big program, where I use functions from the random module in different files. I would like to be able to set the random seed once, at one place, to make the program always return the same results. Can that even be achieved in python?
The main python module that is run should import random and call random.seed(n) - this is shared between all other imports of random as long as somewhere else doesn't reset the seed.
zss's comment should be highlighted as an actual answer:
Another thing for people to be careful of: if you're using
numpy.random, then you need to use numpy.random.seed() to set the
seed. Using random.seed() will not set the seed for random numbers
generated from numpy.random. This confused me for a while. -zss
In the beginning of your application call random.seed(x) making sure x is always the same. This will ensure the sequence of pseudo random numbers will be the same during each run of the application.
Jon Clements pretty much answers my question. However it wasn't the real problem:
It turns out, that the reason for my code's randomness was the numpy.linalg SVD because it does not always produce the same results for badly conditioned matrices !!
So be sure to check for that in your code, if you have the same problems!
Building on previous answers: be aware that many constructs can diverge execution paths, even when all seeds are controlled.
I was thinking "well I set my seeds so they're always the same, and I have no changing/external dependencies, therefore the execution path of my code should always be the same", but that's wrong.
The example that bit me was list(set(...)), where the resulting order may differ.
One important caveat is that for python versions earlier than 3.7, Dictionary keys are not deterministic. This can lead to randomness in the program or even a different order in which the random numbers are generated and therefore non-deterministic random numbers. Conclusion update python.
I was also puzzled by the question when reproducing a deep learning project.So I do a toy experiment and share the results with you.
I create two files in a project, which are named test1.py and test2.py respectively. In test1, I set random.seed(10) for the random module and print 10 random numbers for several times. As you can verify, the results are always the same.
What about test2? I do the same way except setting the seed for the random module.The results display differently every time. Howerver, as long as I import test1———even without using it, the results appear the same as in test1.
So the experiment comes the conclusion that if you want to set seed for all files in a project, you need to import the file/module that define and set the seed.
According to Jon's answer, setting random.seed(n), at the beginning of the main program will set the seed globally. Afterward to set seeds of the imported libraries, one can use the output from random.random(). For example,
rng = np.random.default_rng(int(abs(math.log(random.random()))))
tf.random.set_seed(int(abs(math.log(random.random()))))
You can guarantee this pretty easily by using your own random number generator.
Just pick three largish primes (assuming this isn't a cryptography application), and plug them into a, b and c:
a = ((a * b) % c)
This gives a feedback system that produces pretty random data. Note that not all primes work equally well, but if you're just doing a simulation, it shouldn't matter - all you really need for most simulations is a jumble of numbers with a pattern (pseudo-random, remember) complex enough that it doesn't match up in some way with your application.
Knuth talks about this.
I have a big script in Python. I inspired myself in other people's code so I ended up using the numpy.random module for some things (for example for creating an array of random numbers taken from a binomial distribution) and in other places I use the module random.random.
Can someone please tell me the major differences between the two?
Looking at the doc webpage for each of the two it seems to me that numpy.random just has more methods, but I am unclear about how the generation of the random numbers is different.
The reason why I am asking is because I need to seed my main program for debugging purposes. But it doesn't work unless I use the same random number generator in all the modules that I am importing, is this correct?
Also, I read here, in another post, a discussion about NOT using numpy.random.seed(), but I didn't really understand why this was such a bad idea. I would really appreciate if someone explain me why this is the case.
You have made many correct observations already!
Unless you'd like to seed both of the random generators, it's probably simpler in the long run to choose one generator or the other. But if you do need to use both, then yes, you'll also need to seed them both, because they generate random numbers independently of each other.
For numpy.random.seed(), the main difficulty is that it is not thread-safe - that is, it's not safe to use if you have many different threads of execution, because it's not guaranteed to work if two different threads are executing the function at the same time. If you're not using threads, and if you can reasonably expect that you won't need to rewrite your program this way in the future, numpy.random.seed() should be fine. If there's any reason to suspect that you may need threads in the future, it's much safer in the long run to do as suggested, and to make a local instance of the numpy.random.Random class. As far as I can tell, random.random.seed() is thread-safe (or at least, I haven't found any evidence to the contrary).
The numpy.random library contains a few extra probability distributions commonly used in scientific research, as well as a couple of convenience functions for generating arrays of random data. The random.random library is a little more lightweight, and should be fine if you're not doing scientific research or other kinds of work in statistics.
Otherwise, they both use the Mersenne twister sequence to generate their random numbers, and they're both completely deterministic - that is, if you know a few key bits of information, it's possible to predict with absolute certainty what number will come next. For this reason, neither numpy.random nor random.random is suitable for any serious cryptographic uses. But because the sequence is so very very long, both are fine for generating random numbers in cases where you aren't worried about people trying to reverse-engineer your data. This is also the reason for the necessity to seed the random value - if you start in the same place each time, you'll always get the same sequence of random numbers!
As a side note, if you do need cryptographic level randomness, you should use the secrets module, or something like Crypto.Random if you're using a Python version earlier than Python 3.6.
From Python for Data Analysis, the module numpy.random supplements the Python random with functions for efficiently generating whole arrays of sample values from many kinds of probability distributions.
By contrast, Python's built-in random module only samples one value at a time, while numpy.random can generate very large sample faster. Using IPython magic function %timeit one can see which module performs faster:
In [1]: from random import normalvariate
In [2]: N = 1000000
In [3]: %timeit samples = [normalvariate(0, 1) for _ in xrange(N)]
1 loop, best of 3: 963 ms per loop
In [4]: %timeit np.random.normal(size=N)
10 loops, best of 3: 38.5 ms per loop
The source of the seed and the distribution profile used are going to affect the outputs - if you are looking for cryptgraphic randomness, seeding from os.urandom() will get nearly real random bytes from device chatter (ie ethernet or disk) (ie /dev/random on BSD)
this will avoid you giving a seed and so generating determinisitic random numbers. However the random calls then allow you to fit the numbers to a distribution (what I call scientific random ness - eventually all you want is a bell curve distribution of random numbers, numpy is best at delviering this.
SO yes, stick with one generator, but decide what random you want - random, but defitniely from a distrubtuion curve, or as random as you can get without a quantum device.
It surprised me the randint(a, b) method exists in both numpy.random and random, but they have different behaviors for the upper bound.
random.randint(a, b) returns a random integer N such that a <= N <= b. Alias for randrange(a, b+1). It has b inclusive. random documentation
However if you call numpy.random.randint(a, b), it will return low(inclusive) to high(exclusive). Numpy documentation