Suppress LightGBM Warning

[startakovsky]

The following line is replicated throughout the training for each iteration and it doesn't seem to be generating through Python's standard (warnings)[https://docs.python.org/3/library/warnings.html] module:

[LightGBM] [Warning] No further splits with positive gain, best gain: -inf

How can I suppress this warning? It is generated here. Also, if possible, can you tell me the meaning of this warning?

[guolinke]

it means:

1. the num_leaves is too large, you can set it to a smaller value
2. the min_data is too large
3. your data is hard to fit

[startakovsky]

thanks, how do you suppress these warnings and keep reporting the validation metrics using verbose_eval?

[guolinke]

for sklearn interface, you can set verbose=-1 when defining the model (not in fit).
for lgb.train interface, you can set verbose=-1 in param dict.

[prashant-kikani]

@guolinke what abour lgb.cv ? Can I suppress this warning in lgb.cv?
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf

[guolinke]

I think it can use in cv as well

[prashant-kikani]

For cv, no parameter like "verbose". "verbose_eval" is there. But setting it to -1 doesn't solve the problem. I temporarily solved it by decreasing the num_leaves parameter.

[guolinke]

set it in param dict, not the function arguments.

[hoihui]

setting 'verbose' or 'verbosity' to -1 in the param dict solves this problem for lgb.train,
but it does not help for lgb.cv, or for lgb.train with continued training (i.e. with init_model)

[guolinke]

ping @StrikerRUS

[StrikerRUS]

@guolinke I confirm that with init_model setting verbose=-1 doesn't work. However, in cv verbose=-1 works well for me.

UPD:
It seems that problem persists only in case of both init_model and valid_sets are specified:

Logs are shown:

import lightgbm as lgb
from sklearn.datasets import load_boston
from sklearn.model_selection import train_test_split

X, y = load_boston(True)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.5, random_state=42)
params = {'verbose': -1}

lgb_train = lgb.Dataset(X_train, y_train, params=params, free_raw_data=False)
lgb_eval = lgb.Dataset(X_test, y_test, params=params, free_raw_data=False)
init_gbm = lgb.train(params, lgb_eval)
gbm = lgb.train(params, lgb_train,
                valid_sets=lgb_eval,
                verbose_eval=False,
                init_model=init_gbm)

No logs:

import lightgbm as lgb
from sklearn.datasets import load_boston
from sklearn.model_selection import train_test_split

X, y = load_boston(True)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.5, random_state=42)
params = {'verbose': -1}

lgb_train = lgb.Dataset(X_train, y_train, params=params, free_raw_data=False)
lgb_eval = lgb.Dataset(X_test, y_test, params=params, free_raw_data=False)
init_gbm = lgb.train(params, lgb_eval)
gbm = lgb.train(params, lgb_train,
#                 valid_sets=lgb_eval,
                verbose_eval=False,
                init_model=init_gbm)

Also, instance of valid_sets is matter (the same as training set):
no logs:

import lightgbm as lgb
from sklearn.datasets import load_boston
from sklearn.model_selection import train_test_split

X, y = load_boston(True)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.5, random_state=42)
params = {'verbose': -1}

lgb_train = lgb.Dataset(X_train, y_train, params=params, free_raw_data=False)
lgb_eval = lgb.Dataset(X_test, y_test, params=params, free_raw_data=False)
init_gbm = lgb.train(params, lgb_eval)
gbm = lgb.train(params, lgb_train,
                valid_sets=lgb_train,  # <-------
                verbose_eval=False,
                init_model=init_gbm)

[guolinke]

@StrikerRUS
what logs are shown ?
Including the training information?

I think we need to pass verbose parameter when creating _InnerPredictor:
https://github.com/Microsoft/LightGBM/blob/master/python-package/lightgbm/engine.py#L113 .

[StrikerRUS]

@guolinke

what logs are shown ?

Only warnings:

[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf
[LightGBM] [Warning] No further splits with positive gain, best gain: -inf

In the example init_model is a Booster, so this line?
https://github.com/Microsoft/LightGBM/blob/c86fe61b7e1fa15d0cdd8e69beef269660ec88c8/python-package/lightgbm/engine.py#L115

[goldentom42]

Hi there, I copy the content of issue #1486 as requested by @StrikerRUS

Environment info

Operating System: Linux and windows (can't test on Apple)
C++/Python/R version: latest version (I believe Kaggle uses the latest master branch), occurs also on 2.1.0

Reproducible examples

https://www.kaggle.com/ogrellier/lighgbm-with-selected-features

The problem only occurs with lgb.train (LGBClassifier does not exhibit the same issue) and only if eval_sets argument is provided.

Let me know if you need any further info.
Thanks, Olivier

[goldentom42]

Just to recap :

• Sklearn API works correctly
• lgb.train has the problem if an eval_sets is provided whatever verbose, verbosity or verbose_eval

Hope this helps.

[pseudotensor]

[LightGBM] [Warning] boosting is set=gbdt, boosting_type=gbdt will be ignored. Current value: boosting=gbdt
[LightGBM] [Warning] num_threads is set=4, nthread=-1 will be ignored. Current value: num_threads=4

With the sklearn API I still get during fit the above types of messages even of False is passed to fit() and verbose=-1 passed as dict into the initial class part of a dict.

Note this is even though I am not passing (ever) boosting_type or nthread.

But if I do model.get_params() I see those extra 2 parameters listed there even though I never passed them. So I assume sklearn API is adding them and then lgbm complains.

[StrikerRUS]

@pseudotensor These parameters are regular arguments of the constructor, you should use them instead of aliases in params.

[pseudotensor]

yes thanks, just bit odd these main sklearn ones are listed as aliases

[StrikerRUS]

I will really appreciate if everyone here will test the fix proposed in #1628 (verbose branch) and report here whether it helped in your case.

@goldentom42 Speaking personally about your case, passing params={'verbose': -1} in Dataset constructor (UPD: and removing silent=True) should help even without the fix.

[goldentom42]

Sorry but I'm not sure I understand your statement: I should remove silent=True in the lgb.Dataset() call and set verbose: - 1 when calling lgb.train ?

[StrikerRUS]

@goldentom42 Please see the example:

lgb_train = lgb.Dataset(X_train, y_train, params={'verbose': -1}, free_raw_data=False)
lgb_eval = lgb.Dataset(X_test, y_test, params={'verbose': -1}, free_raw_data=False)
gbm = lgb.train({'verbose': -1}, lgb_train,
                valid_sets=lgb_eval,
                verbose_eval=False)

verbose=-1 in both Dataset constructor and train function.

[goldentom42]

Hi @StrikerRUS, tested LightGBM on Kaggle (they would normally have the latest version) and I don't see the warnings anymore with verbose : -1 in params.

On LightGBM 2.1.2, setting verbose to -1 in both Dataset and lightgbm params make warnings disappear.

Hope this helps.

[StrikerRUS]

@goldentom42 Thanks for your reply! It seems to be that Kaggle uses the latest available release at PyPI, not the master branch:
https://github.com/Kaggle/docker-python/blob/cd1e6ac7d076775af2a5bfbcb65bfd98ea6629de/Dockerfile#L78

We are going to merge the fix into the master branch and close this issue. So, the latest code should work fine with either silent=True argument, or 'verbose': -1 parameter.
However, feel free to report any cases which left uncovered by the fix.

[joseortiz3]

I still get
[LightGBM] [Warning] num_threads is set=1, n_jobs=-1 will be ignored. Current value: num_threads=1
because of conflicting options. Not sure why certain parameters override other parameters and warn you about it.

[guolinke]

@joseortiz3 We use alphabet order to override the parameters, therefore, num_therads will override n_jobs

[saksim]

thanks for the message in this issue.

In a word, Could we understand it as the model would be run OK although such message printed out.

[iggisv9t]

verbose=-1. I also tried all mentioned above and all that I found in docs. It's too annoying.