AttributeError: 'numpy.uintXX' object has no attribute 'encode'

[arnauddgbes]

Hi Team,

Found an issue similar to #476, but for sklearn.neural_network.MLPClassifier objects.
Any model trained using an unsigned integer output type fails to be converted. See below a script to reproduce the issue:

import numpy as np
from sklearn.neural_network import MLPClassifier
from skl2onnx import convert_sklearn
from skl2onnx.common.data_types import FloatTensorType

rettype = np.uint32
nbclasses = 4
nrattribs = 3
nrpts = 10

x_train = np.random.random_sample( (nrpts,nrattribs) ).astype( np.single )
y_train = (x_train[:,0].copy() * nbclasses ).astype( rettype )

model = MLPClassifier()
model.fit( x_train, y_train )

initial_type = [('float_input', FloatTensorType([None,nrattribs]))]
onnxmodel = convert_sklearn(model, initial_types=initial_type)

The conversion fails for rettype = np.uint8, np.uint16 and np.uint32, but not for signed integer types or floating-point types.
A workaround can be to modify the model output type after training, before calling convert_sklearn:
model.classes_ = model.classes_.astype( np.int32 )

Tested with Python 3.9, scikit-learn v1.0.1, skl2onnx v1.10.3