XGboost conversion exception: Unable to guess objective in []

[shadimari]

I am trying to convert the attached XGboost model, which was trained using XGboost v1.3-1 + binary:logistic + early_stopping. The conversion results in an exception due to the fact that the source code assumes it is a multi:* objective function rather than a binary:logistic.

if trees > ntrees > 0:
            state = booster.__getstate__()
            bstate = bytes(state['handle'])
            reg = re.compile(b'(multi:[a-z]{1,15})')
            objs = list(set(reg.findall(bstate)))
            if len(objs) != 1:
                raise RuntimeError(
                    "Unable to guess objective in {}.".format(objs))
            kwargs['num_class'] = trees // ntrees
            kwargs["objective"] = objs[0].decode('ascii')
        else:
            kwargs['num_class'] = 1
            kwargs["objective"] = "binary:logistic"

Source code as follows:

import pickle
import xgboost
from onnxmltools.convert.common import data_types
from onnxmltools.convert import convert_xgboost
from onnxmltools.utils import save_model

xgb_model_loaded = pickle.load(open('xgboost-model', "rb"))
num_features = len(xgb_model_loaded.feature_names)
initial_type = [('float_input', data_types.FloatTensorType([1, num_features]))]
model_onnx = convert_xgboost(xgb_model_loaded, initial_types=initial_type)

model.tar.gz

[shadimari]

@xadupre I think this change has been introduced by you lately. would you please have a quick look?

[shadimari]

@xadupre is there a chance you can quickly confirm this bug?

[xadupre]

Hi,

I modified your example into the following:

import pickle
import xgboost
from onnxmltools.convert.common import data_types
from onnxmltools.convert import convert_xgboost
from onnxmltools.utils import save_model

xgb_model_loaded = pickle.load(open('xgboost-model', "rb"))
num_features = xgb_model_loaded.num_features()
initial_type = [('float_input', data_types.FloatTensorType([None, num_features]))]
model_onnx = convert_xgboost(xgb_model_loaded, initial_types=initial_type)

But it needs PR #480 to work.

[shadimari]

@xadupre many thanks for your support. The issue is fixed.