LightGBMError: GPU Tree Learner was not enabled in this build (ubuntu 18.04 - anaconda - jupyter notebook env)

[diego-florez]

How you are using LightGBM?

LightGBM component: python-api -- sklear-api -- lightgbm.LGBMClassifier

Environment info

ubuntu 18.04 -- anaconda3 -- python3.7 -- jupyter notebook

Operating System: Ubuntu 18.04

CPU/GPU model: NVIDIA-SMI 390.138

C++ compiler version: gcc version 7.5.0

CMake version: cmake version 3.10.2

Python version: Python 3.7.4

LightGBM version or commit hash: LightGBM 3.0.0

Error message and / or logs

LightGBMError: GPU Tree Learner was not enabled in this build. Please recompile with CMake option -DUSE_GPU=1

I want to use LightGBM classifier with gpu on jupyter notebook and I didn't find a way to do it
I know there is a closed issue#2222 solving the problem for windows env. I have followed it as well as followed the installation guide https://lightgbm.readthedocs.io/en/latest/GPU-Tutorial.html, however I was unable to solve the problem

[StrikerRUS]

@diego-florez

I have followed it as well as followed the installation guide https://lightgbm.readthedocs.io/en/latest/GPU-Tutorial.html, ...

Have you tried this installation guide?
https://lightgbm.readthedocs.io/en/latest/Installation-Guide.html#id17

however I was unable to solve the problem

Please be more precise. What is the problem? What steps did you do? Post your commands and corresponding logs.

[diego-florez]

Hi @StrikerRUS

• problem:
  LightGBMError: GPU Tree Learner was not enabled in this build. Please recompile with CMake option -DUSE_GPU=1

• steps:
  picture bellow
  
  Then on the jupyter nb I set the model params like this --> clf = RandomizedSearchCV(lgb.LGBMClassifier(max_depth=-1, random_state=None, silent=True, n_estimators=200 , device="gpu", gpu_platform_id = 0, gpu_device_id = 0)

[StrikerRUS]

OK. Seems that you have successfully compiled dynamic library. Now you should install Python wrapper.

pip uninstall lightgbm  # remove old installation that have no GPU support
cd ../python-package
python setup.py install --precompile

Please pay attention to the second line: fatal: ... It is possible that you are compiling some old version of the library. Remove LightGBM folder before git clone command.

[diego-florez]

Thanks, I solved the problem following your instructions, but now I have a problem with bin size --> LightGBMError: bin size 369 cannot run on GPU
I tried in params --> clf = RandomizedSearchCV(lgb.LGBMClassifier(max_depth=-1, random_state=None, silent=True, n_estimators=200, max_bin = 63, device="gpu", gpu_platform_id = 0, gpu_device_id = 0)

However I have the same error

Should I open another issue? As issue #3310 is solved

[guolinke]

@diego-florez did you use categorical features? if yes, you can try to disable them.

[diego-florez]

@guolinke yes, 11/18 are categorical, but they are set as type "categorical" and I tried with the following params categorical_feature="auto" (as it is recommended), categorical_feature=my_list_of_cat_vars (as it is recommended), ignore_column=my_list_of_cat_vars (as you have said -- not sure at all if this is the param you meant)

And the same error persist with the 3 of them

[StrikerRUS]

@diego-florez Are you loading your data from a file? Because ignore_column parameter only works in that case:

Note: works only in case of loading data directly from file
https://lightgbm.readthedocs.io/en/latest/Parameters.html#ignore_column

[diego-florez]

the training data is a pandas dataframe previously loaded in the notebook, there is not any other solution for the problem? Maybe change categorical variables to one hot encoding?

[guolinke]

@diego-florez can you try to use label encoding for categorical features, and don't set 'categorical_feature'. Then LightGBM will treat them as numerical features.

[naveen-marthala]

OK. Seems that you have successfully compiled dynamic library. Now you should install Python wrapper.

pip uninstall lightgbm  # remove old installation that have no GPU support
cd ../python-package
python setup.py install --precompile

Please pay attention to the second line: fatal: ... It is possible that you are compiling some old version of the library. Remove LightGBM folder before git clone command.

I have done exactly what was shown in the post you replied to, in this reply.
And also did what you wrote in this reply. I can now use lighgbm with gpu.

The issue is that there's no increase in the computation speed. and the GPU is barley being used(I see usage as low as 0.08GB). How do I speed up the training? my data set sizes are, train=(247570, 70) and validation=(13031, 70). here is all my code and my hyperparameters:

from lightgbm import LGBMClassifier

lgb_test_on_gpu = LGBMClassifier(learning_rate=0.01, n_estimators=10_000, num_leaves=127, n_jobs=-1,
                                 device = 'gpu', gpu_platform_id = 0, gpu_device_id = 0,
                                 colsample_bytree=0.75, subsample=0.75, subsample_freq=5, random_state=1, objective='multi_logloss')
lgb_test_on_gpu.fit(X=strat_train_df.drop(columns='damage_grade', axis=1), y=strat_train_df.loc[:,'damage_grade'],
                    eval_set=[(strat_valid_df.drop(columns='damage_grade', axis=1), strat_valid_df.loc[:,'damage_grade'])],
                    verbose=100, early_stopping_rounds=1_000)

[StrikerRUS]

@naveen-9697 Please refer to #768 and some benchmarks in https://github.com/Laurae2/ml-perf/issues and https://github.com/szilard/GBM-perf/issues for the problem of small GPU utilization.

[StrikerRUS]

Closing this issue because the original problem has been solved.

[github-actions]

This issue has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.