Allow seed fix in Sampler
#669
Conversation
Previous approach via xor lead to collisions for pairs of equal seeds
| @@ -638,3 +653,32 @@ def cart_prod(*arrs): | |||
| """ | |||
| grids = np.meshgrid(*arrs, indexing='ij') | |||
| return np.stack(grids, axis=-1).reshape(-1, len(arrs)) | |||
|
|
|||
| def mix_samplers_seeds(left, right): | |||
There was a problem hiding this comment.
Should be a staticmethod of base Sampler class
| name = _get_method_by_alias(name, 'ss') | ||
| self.name = name | ||
| self.state = make_rng(seed) | ||
| self.distr = getattr(ss, self.name)(**kwargs) | ||
|
|
||
| def sample(self, size): |
There was a problem hiding this comment.
rng should be allowed to be passed as optional argument
| @@ -254,7 +254,7 @@ def __init__(self, domain=None, **kwargs): | |||
| self.n_updates = 0 | |||
| self.additional = True | |||
| self.create_id_prefix = False | |||
| self.random_state = None | |||
| self.rng = None | |||
There was a problem hiding this comment.
There is a chain of random_state variables and their SeedSequences in batchflow:
---Pipeline
---Dataset
------Batch
---------inbatch_parallel workers (threads / processes / for-items)
And Research is even one level above that.
Have you read this entire chain of properties before changing it?
| @@ -638,3 +653,32 @@ def cart_prod(*arrs): | |||
| """ | |||
| grids = np.meshgrid(*arrs, indexing='ij') | |||
| return np.stack(grids, axis=-1).reshape(-1, len(arrs)) | |||
|
|
|||
| def mix_samplers_seeds(left, right): | |||
There was a problem hiding this comment.
You are mixing seeds, which is wrong. If I would mix two samplers in the proposed way and generate one random number, it would be the same, irrespectable of how many times each of those two samplers was called before creating the mixture.
You need to mix entropies, and there is a well-established (and used in other places in batchflow) way to do so: np.random.SeedSequence
rng1, rng2
state1 = rng1.bit_generator.state['state']['state']
state2 = rng2.bit_generator.state['state']['state']
seed = np.random.SeedSequence([state1, state2])
rng = np.random.default_rng(seed)While the difference between these two approaches is hard to come by in any realistic example, the latter is the official way to do so.
In either case, the current proposed way to seed the RNG in sampler would not work with batchflow+seismiQB ways to fix the randomization, and the only thing you need to actually fix the seed for make_locations(sampler) in seismiQB is the ability to pass custom rng into Sampler.sample call
No description provided.