Support numpy.random.RandomState objects (take 2) #556
Conversation
Codecov Report
@@ Coverage Diff @@
## master #556 +/- ##
===========================================
+ Coverage 85.30% 98.83% +13.53%
===========================================
Files 115 115
Lines 4205 4297 +92
===========================================
+ Hits 3587 4247 +660
+ Misses 618 50 -568
Continue to review full report at Codecov.
|
|
This was the error coming out of all the windows unit tests: |
| @@ -26,6 +26,7 @@ class CatBoostClassifier(Estimator): | |||
| supported_problem_types = [ProblemTypes.BINARY, ProblemTypes.MULTICLASS] | |||
|
|
|||
| def __init__(self, n_estimators=1000, eta=0.03, max_depth=6, bootstrap_type=None, random_state=0): | |||
| random_seed = get_random_seed(random_state, 0, SEED_BOUNDS.max_bound) | |||
There was a problem hiding this comment.
The catboost estimators need random_seed to be between 0 and max int, otherwise they throw an error.
| @@ -19,13 +19,14 @@ class XGBoostClassifier(Estimator): | |||
| supported_problem_types = [ProblemTypes.BINARY, ProblemTypes.MULTICLASS] | |||
|
|
|||
| def __init__(self, eta=0.1, max_depth=3, min_child_weight=1, n_estimators=100, random_state=0): | |||
| random_seed = get_random_seed(random_state, SEED_BOUNDS.min_bound, SEED_BOUNDS.max_bound) | |||
There was a problem hiding this comment.
Like catboost, the xgboost classifier needs random_seed to be an int, not np.random.RandomState. The weird thing is, passing random_state worked fine on linux... so it appears its only on windows that xgboost requires this!
| return random_state.randint(min_bound, max_bound) | ||
| if random_state < min_bound or random_state >= max_bound: | ||
| return random_state % min(abs(min_bound), abs(max_bound)) | ||
| return random_state |
There was a problem hiding this comment.
I was hoping to avoid defining a function like this because it introduces complexity. But having this provides a pattern which we can follow for when we add new pipelines which require integer seeds instead of np.random.RandomState.
There was a problem hiding this comment.
I think this is a good solution!
There was a problem hiding this comment.
Thanks. I still wish it were simpler and therefore easier to understand, but I'm not sure there's a better alternative right now. Let's continue to think about it :)
dsherry
force-pushed
the
ds_revert_347_random_state
branch
from
d60a2e7
to
fbaa89f
Compare
dsherry
force-pushed
the
ds_347_random_state_windows
branch
2 times, most recently
from
a295283
to
33ef6ba
Compare
* Squash random_state work from 347_random_state * Lint * Lint * Changelog * Lint * Test update * Always pass random_state to components * Lint * Fix bug: set random state first. Remove usages of random_state as dict param item in test_pipelines.py * update test for clarity * Fix catboost * Update logreg test * Lint catboost * Update tuner impl to handle random_state * Test changes * Lint * Docs changes * Add unit test for get_random_state * Update test * Remove uncalled code after my changes * Fix tests after rebase * Add unit test coverage for RandomSearchTuner.is_search_space_exhausted * Add unit test coverage for max_time * Add test coverage of get_pipeline when invalid * Lint * Add unit test coverage of when fit/score throws in autobase * Remove duplicate * Lets try that again... got mysterious docs failure
… for different classes
dsherry
force-pushed
the
ds_347_random_state_windows
branch
from
33ef6ba
to
85fae99
Compare
Closes #347 .
A version of #530 which won't break the windows tests.
Will retarget on
masteronce #554 and #555 are merged.Changes
random_stateas a keyword argumentrandom_stateas either anintseed or anumpy.random.RandomStateobjectget_random_statehelper to standardize tonp.random.RandomStateobjectsnp.random.RandomStateobjects to get random seeds, via aget_random_seedmethodSEED_BOUNDSrange constantBuilding off of #441 , opening to test my own changes on top of Angela's work (thank you @angela97lin!)
Note @kmax12 we had discussed sticking with seeds internally, but using
np.random.RandomStateturned out to be the simpler option.