Flaky e2e test for the XGBoost model with the 'gblinear' booster

[izeigerman]

Context from @StrikerRUS:

Now Go is failing (refer to #200 (comment)):

=================================== FAILURES ===================================
_ test_e2e[xgboost_XGBClassifier - go_lang - train_model_classification_binary2] _
estimator = XGBClassifier(base_score=0.6, booster='gblinear', colsample_bylevel=None,
              colsample_bynode=None, colsamp...ambda=0, scale_pos_weight=1, subsample=None,
              tree_method=None, validate_parameters=False, verbosity=None)
executor_cls = <class 'tests.e2e.executors.go.GoExecutor'>

...

expected=[0.04761511 0.9523849 ], actual=[0.047615, 0.952385]
expected=[0.06296992 0.9370301 ], actual=[0.06297, 0.93703]
expected=[0.12447995 0.87552005], actual=[0.124479, 0.875521]
expected=[0.0757848 0.9242152], actual=[0.075784, 0.924216]
expected=[0.8092151  0.19078489], actual=[0.809212, 0.190788]

BTW, in attempts to check my guess from #200 (comment), I found that coefs in gblinear are also float32:
https://github.com/dmlc/xgboost/blob/67d267f9da3b15a6e5a8393afae9be921a4e224b/src/gbm/gblinear_model.h#L110

https://github.com/dmlc/xgboost/blob/67d267f9da3b15a6e5a8393afae9be921a4e224b/src/gbm/gblinear_model.h#L120

https://github.com/dmlc/xgboost/blob/67d267f9da3b15a6e5a8393afae9be921a4e224b/src/gbm/gblinear_model.h#L82

https://github.com/dmlc/xgboost/blob/67d267f9da3b15a6e5a8393afae9be921a4e224b/src/gbm/gblinear_model.h#L91

and from #188 (comment) we know that bst_float is actually float
https://github.com/dmlc/xgboost/blob/8d06878bf9b778db68ae98f68d99a3557c7ea885/include/xgboost/base.h#L110-L111

Created dmlc/xgboost#5634.

[izeigerman]

Duplicating my question here:
@StrikerRUS excellent observation 👍 Do you think we can pull off the same trick as we did for thresholds? Basically cast weights and coefs to float32 on our end. Or is it different this time?

[StrikerRUS]

I think we can wait for a while for any reply in dmlc/xgboost#5634. Let say two or three days, and only then take any actions.

[StrikerRUS]

For the record. Two new cases (both Powershell):

expected=[0.04631048 0.9536895 ], actual=[0.0463101577391968, 0.953689842260803]
expected=[0.06417668 0.9358233 ], actual=[0.064176231780856, 0.935823768219144]
expected=[0.12548977 0.8745102 ], actual=[0.125489171473375, 0.874510828526625]
expected=[0.07494283 0.9250572 ], actual=[0.0749422009903284, 0.925057799009672]
expected=[0.80788165 0.19211833], actual=[0.807878724641367, 0.192121275358633]

expected=[0.04594815 0.95405185], actual=[0.0459477380325012, 0.954052261967499]
expected=[0.06591994 0.93408006], actual=[0.0659193724496036, 0.934080627550396]
expected=[0.13377541 0.8662246 ], actual=[0.133774447739148, 0.866225552260852]
expected=[0.07744116 0.92255884], actual=[0.0774401457484807, 0.922559854251519]
expected=[0.79560804 0.20439194], actual=[0.795604680327838, 0.204395319672162]

Side question: I wonder, why values are so non-deterministic? Did we forget to set random_state somewhere?..

[izeigerman]

Hm, good question. I briefly look into the code and couldn't spot any obvious sources of non-determinism.

[StrikerRUS]

Probably we've encountered this issue dmlc/xgboost#5298 (comment) and #188 was not enough to fix issues like we have in generated examples:

m2cgen/generated_code_examples/python/regression/lightgbm.py

Line 8 in 4869e00

if (input[12]) > (14.395000000000001):

m2cgen/generated_code_examples/python/regression/lightgbm.py

Line 13 in 4869e00

if (input[12]) > (19.830000000000002):

m2cgen/generated_code_examples/haskell/regression/random_forest.hs

Line 15 in 4869e00

25.320000000000004

But anyway, we should get a reliable repro first.

[StrikerRUS]

Have no time to read this right now, but looks like related to our issue: numpy/numpy#9941.