Add support for np.random.Generator #6566
Conversation
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #6566 +/- ##
=======================================
Coverage 97.79% 97.79%
=======================================
Files 1126 1128 +2
Lines 95784 95825 +41
=======================================
+ Hits 93670 93715 +45
+ Misses 2114 2110 -4 ☔ View full report in Codecov by Sentry. |
sorry about the delay, I will do the review this afternoon. |
There was a problem hiding this comment.
Instead of introducing new type we should consider defining and extending RANDOM_STATE_OR_SEED_LIKE and provide parsers that would express it as either a RandomState or Generator as needed. Eventually we may completely transition to Generator-s and deprecate/remove the parser to RandomState.
cirq-core/cirq/value/prng.py
Outdated
| def parse_prng( | ||
| prng_or_seed: PRNG_OR_SEED_LIKE, | ||
| ) -> Union[np.random.Generator, np.random.RandomState, _CUSTOM_PRNG_T]: |
There was a problem hiding this comment.
The return type of different types tends to be a code smell. Such returned value is less useful for type checking. In addition, Generator and RandomState (not to mention _CUSTOM_PRNG_T) have different APIs so the parse_prng caller would still need to do some isinstance check to ascertain the actual type and figure what methods can be called.
I would propose an alternative approach:
(1) convert the RANDOM_STATE_OR_SEED_LIKE type from Any to the Union of numpy types that can be converted to RandomState and the np.random.Generator type. Hopefully this can be done without too much hassle with typechecks, because the current Any type skips them completely.
(2) extend parse_random_state to accept a Generator object and convert it to RandomState.
Generator-s have bit_generator attribute that can be used to create RandomState.
(3) add method parse_random_generator to the cirq.value.random_state module which would take RANDOM_STATE_OR_SEED_LIKE argument and convert it to a Generator object.
np.random.RandomState() has a _bit_generator attribute that can be used for creating a Generator.
If in some configurations the _bit_generator is not present, we can just use RandomState.randint to get a seed for the np.random.default_rng()
With these steps in place, we can keep all the existing interfaces that take RANDOM_STATE_OR_SEED_LIKE and just start replacing its interpretation from parse_random_state to parse_random_generator as needed.
This would also avoid bifurcation between RANDOM_STATE_OR_SEED_LIKE and PRNG_OR_SEED_LIKE types that may need several major releases to clear up.
|
@pavoljuhas I updated parse_random_state to accept a generator and turn it into a randomstate and parse_prng to always return a generator. however I think that we do need new type |
| If is an integer or None, turns into a `np.random.Generator` seeded with that value. | ||
| If is an instance of `np.random.Generator` or a subclass of it, return as is. | ||
| If is an instance of `np.random.RandomState` or has a `randint` method, returns |
There was a problem hiding this comment.
| If is an integer or None, turns into a `np.random.Generator` seeded with that value. | |
| If is an instance of `np.random.Generator` or a subclass of it, return as is. | |
| If is an instance of `np.random.RandomState` or has a `randint` method, returns | |
| If an integer or None, turns into a `np.random.Generator` seeded with that value. | |
| If an instance of `np.random.Generator` or a subclass of it, return as is. | |
| If an instance of `np.random.RandomState` or has a `randint` method, returns |
|
|
||
|
|
||
| def parse_prng( | ||
| prng_or_seed: Union[PRNG_OR_SEED_LIKE, RANDOM_STATE_OR_SEED_LIKE] |
There was a problem hiding this comment.
Can we just have only the PRNG_OR_SEED_LIKE type?
RANDOM_STATE_OR_SEED_LIKE is Any so it turns off type checking of the argument.
| if prng_or_seed is None or isinstance(prng_or_seed, numbers.Integral): | ||
| return np.random.default_rng(prng_or_seed if prng_or_seed is None else int(prng_or_seed)) |
There was a problem hiding this comment.
Can we return a singleton Generator object for None?
The None arg is going to be frequently used as a default for optional arguments.
Singleton would prevent creation of potentially large number of Generator objects.
| if prng_or_seed is None or isinstance(prng_or_seed, numbers.Integral): | ||
| return np.random.default_rng(prng_or_seed if prng_or_seed is None else int(prng_or_seed)) | ||
| if isinstance(prng_or_seed, np.random.RandomState): | ||
| return np.random.default_rng(prng_or_seed.randint(2**31)) |
There was a problem hiding this comment.
We can reuse the bit generator for a more genuine conversion.
| return np.random.default_rng(prng_or_seed.randint(2**31)) | |
| return np.random.default_rng(prng_or_seed._bit_generator) |
| randint = getattr(prng_or_seed, "randint", None) | ||
| if randint is not None: | ||
| return np.random.default_rng(randint(2**31)) | ||
| raise TypeError(f"{prng_or_seed} can't be converted to a pseudorandom number generator") |
There was a problem hiding this comment.
Nit - maybe state the actual class here ?
| raise TypeError(f"{prng_or_seed} can't be converted to a pseudorandom number generator") | |
| raise TypeError(f"{prng_or_seed} cannot be converted to the numpy.random.Generator") |
| def _sample(prng): | ||
| return tuple(prng.random(10)) |
There was a problem hiding this comment.
I don't think we need this. One output from random() is enough to check if 2 generators are at the same seed.
| # An `np.random.Generator` or a seed. | ||
| group_inputs: List[Union[int, np.random.Generator]] = [42, np.random.default_rng(42)] | ||
| group: List[np.random.Generator] = [cirq.value.parse_prng(s) for s in group_inputs] | ||
| eq.add_equality_group(*[_sample(g) for g in group]) |
There was a problem hiding this comment.
Let us not check cross-group inequality. Following the test_parse_random_state style is a bit more readable
| # An `np.random.Generator` or a seed. | |
| group_inputs: List[Union[int, np.random.Generator]] = [42, np.random.default_rng(42)] | |
| group: List[np.random.Generator] = [cirq.value.parse_prng(s) for s in group_inputs] | |
| eq.add_equality_group(*[_sample(g) for g in group]) | |
| # An `np.random.Generator` or a seed. | |
| prngs = [ | |
| cirq.value.parse_prng(42), | |
| cirq.value.parse_prng(np.int32(42)), | |
| cirq.value.parse_prng(np.random.default_rng(42)), | |
| ] | |
| vals = [prng.random() for prng in prngs] | |
| eq = cirq.testing.EqualsTester() | |
| eq.add_equality_group(*vals) |
|
|
||
| # A None seed. | ||
| prng = cirq.value.parse_prng(None) | ||
| eq.add_equality_group(_sample(prng)) |
There was a problem hiding this comment.
This is a noop check for a single value. Perhaps replace with
assert prng is cirq.value.parse_prng(None)
if you are OK with the previous suggestion to have a singleton generator for None.
| # RandomState PRNG. | ||
| prng = cirq.value.parse_prng(np.random.RandomState(42)) | ||
| eq.add_equality_group(_sample(prng)) |
There was a problem hiding this comment.
We can check reproducibility here -
| # RandomState PRNG. | |
| prng = cirq.value.parse_prng(np.random.RandomState(42)) | |
| eq.add_equality_group(_sample(prng)) | |
| # RandomState PRNG. | |
| prngs = [ | |
| cirq.value.parse_prng(np.random.RandomState(42)), | |
| cirq.value.parse_prng(np.random.RandomState(42)), | |
| ] | |
| vals = [prng.random() for prng in prngs] | |
| eq = cirq.testing.EqualsTester() | |
| eq.add_equality_group(*vals) |
| # np.random module | ||
| prng = cirq.value.parse_prng(np.random) | ||
| eq.add_equality_group(_sample(prng)) |
There was a problem hiding this comment.
We should not support creation of generator from a module, not a good practice.
The use of np.random module was causing pickle havoc in #3717.
I don't quite see a need for it, users can pass None for a default generator.
I'd be open to throwing TypeError for a module argument.
| # np.random module | |
| prng = cirq.value.parse_prng(np.random) | |
| eq.add_equality_group(_sample(prng)) |
This PR addresses #6531
I don't modify support for old
RANDOM_STATE_OR_SEED_LIKEinstead I create a newPRNG_OR_SEED_LIKEtype.The reasons for that are:
RANDOM_STATE_OR_SEED_LIKEimplies that the result will be a RandomState. This is enforced in the code by forcing a a cast toRandomStatein theparse_random_statefunctionCirq/cirq-core/cirq/value/random_state.py
Line 60 in e1b03ef
parse_random_statedefaults tonp.randomwhen the seed isNonewhich is problematic when usign multiprocessing/multithreading since the internal state of the np.random module becomes a shared state that will make the results of experiments correlated, it will also negatively impact the performance since the threads/processes will be blocked on write operations to the internal state of the module.Cirq/cirq-core/cirq/value/random_state.py
Line 56 in e1b03ef
np.randomis also problematic in other ways as demonstrated in Can't use cirq.Simulator() in a multiprocessing closure (unable to pickle) #3717RANDOM_STATE_OR_SEED_LIKEtype is just an alias ofAnywhich makes using@overloadto specify the return type impossible.The solution
PRNG_OR_SEED_LIKEhas union ofNone, int, np.random.RandomState, np.random.Generator, _CUSTOM_PRNG_T, the first two are types of seeds that will always give annp.random.Generator, the rest are objects that will be returned as is. and using@overloadwe specifiy the correct return type.deprecating
RANDOM_STATE_OR_SEED_LIKEatm is not possible due to how widely it's used. instead we should start to prefer usingPRNG_OR_SEED_LIKEand maybe deprecateRANDOM_STATE_OR_SEED_LIKEatcirq 2.0and remove it bycirq 3.0