[Bug] LightGBMError: bin size 257 cannot run on GPU

[rohan-gt]

I'm getting the following error while running the latest LightGBM GPU using these params:

params = {
            'device_type': 'gpu',
            'gpu_device_id': 0,
            'gpu_platform_id': 0,
            'gpu_use_dp': 'false',
            'max_bin': 255
}

on Google Colab using this Kaggle dataset: https://www.kaggle.com/c/ieee-fraud-detection. I'm dropping all the categorical variables:

LightGBMError: bin size 257 cannot run on GPU

[hengzhe-zhang]

I also have the same problem. Is there any way to solve this problem?

[guolinke]

sorry for missing this issue.
The max-bin actually cannot limit the number of bins for categorical feature.
there are two workarounds:

1. use the categorical encodings, converting categorical features to numerical ones.
2. split one categorical feature to multi categorical features, and make sure the number of categories in each splitted feature smaller than 256.

[pseudotensor]

I hit this randomly for no reason with categorical_features as explicitly empty. Has nothing to do with that. The test that hit this normally has passed 1000 times before.

File "/opt/h2oai/dai/cuda-10.0/lib/python3.6/site-packages/lightgbm_gpu/sklearn.py", line 794, in fit
    categorical_feature=categorical_feature, callbacks=callbacks, init_model=init_model)
  File "/opt/h2oai/dai/cuda-10.0/lib/python3.6/site-packages/lightgbm_gpu/sklearn.py", line 637, in fit
    callbacks=callbacks, init_model=init_model)
  File "/opt/h2oai/dai/cuda-10.0/lib/python3.6/site-packages/lightgbm_gpu/engine.py", line 230, in train
    booster = Booster(params=params, train_set=train_set)
  File "/opt/h2oai/dai/cuda-10.0/lib/python3.6/site-packages/lightgbm_gpu/basic.py", line 2104, in __init__
    ctypes.byref(self.handle)))
  File "/opt/h2oai/dai/cuda-10.0/lib/python3.6/site-packages/lightgbm_gpu/basic.py", line 52, in _safe_call
    raise LightGBMError(_LIB.LGBM_GetLastError().decode('utf-8'))
lightgbm.basic.LightGBMError: bin size 257 cannot run on GPU

The number of bins was 255 and there are no categorical features as explicitly chosen.

[George3d6]

Same issue happened for me:

  File "/usr/local/lib/python3.6/dist-packages/lightgbm/engine.py", line 228, in train
    booster = Booster(params=params, train_set=train_set)
  File "/usr/local/lib/python3.6/dist-packages/lightgbm/basic.py", line 2237, in __init__
    ctypes.byref(self.handle)))
  File "/usr/local/lib/python3.6/dist-packages/lightgbm/basic.py", line 110, in _safe_call
    raise LightGBMError(_LIB.LGBM_GetLastError().decode('utf-8'))
lightgbm.basic.LightGBMError: bin size 257 cannot run on GPU

[George3d6]

Some of the values are categorical in my case but not as many as 257 different ones, combined with @pseudotensor comment, I assume this is something else.

[lewis-morris]

I am also getting the same error.

lightgbm.basic.LightGBMError: bin size 257 cannot run on GPU

[MAxx8371]

What causes this error？Is the bin_size of a categorical feature bigger than the max_bin that causes the error? Or it is because the memory is not enough. And the model can work on CPU. Thank you!

[jiluojiluo]

lightgbm.basic.LightGBMError: bin size 407 cannot run on GPU

[jiluojiluo]

lightgbm.basic.LightGBMError: bin size 407 cannot run on GPU

this is a bug for lightGBM run on GPU,when use CPU,it is OK. SO ,LGBM on GPU need improve.

[ChiHangChen]

Same error encountered, any update?

[aforadi]

Same here.

lightgbm.basic.LightGBMError: bin size 670 cannot run on GPU

[CVPaul]

It seems that I've identified the cause of the error: The calculation method for num_total_bin used during Exclusive Feature Bundling

LightGBM/src/io/dataset.cpp

Line 134 in 665c473

(bin_mappers[fidx]->GetDefaultBin() == 0 ? -1 : 0);

doesn't align completely with the way num_total_bin is calculated during the creation of a FeatureGroup

LightGBM/include/LightGBM/feature_group.h

Line 68 in 665c473

if (bin_mappers_[i]->GetMostFreqBin() == 0) {

As a result, the max_bin_per_group (=256) is working during Bundling, but it is not working when creating the FeatureGroup. When I replaced the GetDefaultBin() at dataset.cpp#L134 with GetMostFreqBin(), the issue was resolved. I had tested with the case reported here: #4082

[XQ-UT]

Same issue here. Can we prioritize the fixing MR?

[damvantai]

I also encountered the same error in this situation in fold 3th (total 5 fold), when my category characteristic had many NAN values.
Can do value np.NAN, np.int in feature category -> max_bin
I have use labelencoder convert feature category -> maxbin

But after i use
df_train[cat_cols] = df_train[cat_cols].astype(str)
df_train[cat_cols] = df_train[cat_cols].astype("category")
then the training process will good!

or can setup
"min_data_in_bin": 256, or higher