Two ways to implement a converter

There are two ways to write a converter. The first one is very verbose (see ada_boost.py for an example). The other is less verbose and easier to understand (see k_means.py).

The first way is used in Implement a new converter. This one demonstrates the second way which is usually the one used in other converter library. It is more verbose.

Custom model

It basically copies what is in example :ref:`l-plot-custom-converter.

from skl2onnx.common.data_types import guess_proto_type
from onnxconverter_common.onnx_ops import apply_sub
from onnxruntime import InferenceSession
from skl2onnx import update_registered_converter
from skl2onnx import to_onnx
import numpy
from sklearn.base import TransformerMixin, BaseEstimator
from sklearn.datasets import load_iris


class DecorrelateTransformer(TransformerMixin, BaseEstimator):
    """
    Decorrelates correlated gaussian features.

    :param alpha: avoids non inversible matrices
        by adding *alpha* identity matrix

    *Attributes*

    * `self.mean_`: average
    * `self.coef_`: square root of the coveriance matrix
    """

    def __init__(self, alpha=0.):
        BaseEstimator.__init__(self)
        TransformerMixin.__init__(self)
        self.alpha = alpha

    def fit(self, X, y=None, sample_weights=None):
        if sample_weights is not None:
            raise NotImplementedError(
                "sample_weights != None is not implemented.")
        self.mean_ = numpy.mean(X, axis=0, keepdims=True)
        X = X - self.mean_
        V = X.T @ X / X.shape[0]
        if self.alpha != 0:
            V += numpy.identity(V.shape[0]) * self.alpha
        L, P = numpy.linalg.eig(V)
        Linv = L ** (-0.5)
        diag = numpy.diag(Linv)
        root = P @ diag @ P.transpose()
        self.coef_ = root
        return self

    def transform(self, X):
        return (X - self.mean_) @ self.coef_


data = load_iris()
X = data.data

dec = DecorrelateTransformer()
dec.fit(X)
pred = dec.transform(X[:5])
print(pred)

Out:

[[ 0.0167562   0.52111756 -1.24946737 -0.56194325]
 [-0.0727878  -0.80853732 -1.43841018 -0.37441392]
 [-0.69971891 -0.09950908 -1.2138161  -0.3499275 ]
 [-1.13063404 -0.13540568 -0.79087008 -0.73938966]
 [-0.35790036  0.91900236 -1.04034399 -0.6509266 ]]

Conversion into ONNX

The shape calculator does not change.

def decorrelate_transformer_shape_calculator(operator):
    op = operator.raw_operator
    input_type = operator.inputs[0].type.__class__
    input_dim = operator.inputs[0].type.shape[0]
    output_type = input_type([input_dim, op.coef_.shape[1]])
    operator.outputs[0].type = output_type

The converter is different.

def decorrelate_transformer_converter(scope, operator, container):
    op = operator.raw_operator
    out = operator.outputs

    # We retrieve the unique input.
    X = operator.inputs[0]

    # In most case, computation happen in floats.
    # But it might be with double. ONNX is very strict
    # about types, every constant should have the same
    # type as the input.
    proto_dtype = guess_proto_type(X.type)

    mean_name = scope.get_unique_variable_name('mean')
    container.add_initializer(mean_name, proto_dtype,
                              op.mean_.shape, list(op.mean_.ravel()))

    coef_name = scope.get_unique_variable_name('coef')
    container.add_initializer(coef_name, proto_dtype,
                              op.coef_.shape, list(op.coef_.ravel()))

    op_name = scope.get_unique_operator_name('sub')
    sub_name = scope.get_unique_variable_name('sub')
    # This function is defined in package onnxconverter_common.
    # Most common operators can be added to the graph with
    # these functions. It handles the case when specifications
    # changed accross opsets (a parameter becomes an input
    # for example).
    apply_sub(scope, [X.full_name, mean_name], sub_name, container,
              operator_name=op_name)

    op_name = scope.get_unique_operator_name('matmul')
    container.add_node(
        'MatMul', [sub_name, coef_name],
        out[0].full_name, name=op_name)

We need to let skl2onnx know about the new converter.

update_registered_converter(
    DecorrelateTransformer, "SklearnDecorrelateTransformer",
    decorrelate_transformer_shape_calculator,
    decorrelate_transformer_converter)


onx = to_onnx(dec, X.astype(numpy.float32))

sess = InferenceSession(onx.SerializeToString())

exp = dec.transform(X.astype(numpy.float32))
got = sess.run(None, {'X': X.astype(numpy.float32)})[0]


def diff(p1, p2):
    p1 = p1.ravel()
    p2 = p2.ravel()
    d = numpy.abs(p2 - p1)
    return d.max(), (d / numpy.abs(p1)).max()


print(diff(exp, got))

Out:

(1.0531815934911037e-06, 0.0003497831114202993)

Let’s check it works as well with double.

onx = to_onnx(dec, X.astype(numpy.float64))

sess = InferenceSession(onx.SerializeToString())

exp = dec.transform(X.astype(numpy.float64))
got = sess.run(None, {'X': X.astype(numpy.float64)})[0]
print(diff(exp, got))

Out:

(4.440892098500626e-16, 2.2783365782956412e-16)

The differences are smaller with double as expected.