Custom Operator for NMF Decomposition

NMF factorizes an input matrix into two matrices W, H of rank k so that . may be a binary matrix where i is a user and j a product he bought. The prediction function depends on whether or not the user needs a recommandation for an existing user or a new user. This example addresses the first case.

The second case is more complex as it theoretically requires the estimation of a new matrix W with a gradient descent.

Building a simple model

import os
import skl2onnx
import onnxruntime
import sklearn
from sklearn.decomposition import NMF
import numpy as np
import matplotlib.pyplot as plt
from onnx.tools.net_drawer import GetPydotGraph, GetOpNodeProducer
import onnx
from skl2onnx.algebra.onnx_ops import OnnxArrayFeatureExtractor, OnnxMul, OnnxReduceSum
from skl2onnx.common.data_types import FloatTensorType
from onnxruntime import InferenceSession


mat = np.array(
    [[1, 0, 0, 0], [1, 0, 0, 0], [1, 0, 0, 0], [1, 0, 0, 0], [1, 0, 0, 0]],
    dtype=np.float64,
)
mat[: mat.shape[1], :] += np.identity(mat.shape[1])

mod = NMF(n_components=2)
W = mod.fit_transform(mat)
H = mod.components_
pred = mod.inverse_transform(W)

print("original predictions")
exp = []
for i in range(mat.shape[0]):
    for j in range(mat.shape[1]):
        exp.append((i, j, pred[i, j]))

print(exp)

original predictions
[(0, 0, 1.8940507356352687), (0, 1, 0.10912372262184848), (0, 2, 0.3072453141962623), (0, 3, 0.3072453141962623), (1, 0, 1.014673742790866), (1, 1, 0.9848866016414943), (1, 2, 0.0), (1, 3, 0.0), (2, 0, 1.1066115912409111), (2, 1, 0.0), (2, 2, 0.19083752823558645), (2, 3, 0.19083752823558645), (3, 0, 1.1066115912409111), (3, 1, 0.0), (3, 2, 0.19083752823558645), (3, 3, 0.19083752823558645), (4, 0, 0.9470253678176344), (4, 1, 0.05456186131092424), (4, 2, 0.15362265709813114), (4, 3, 0.15362265709813114)]

Let’s rewrite the prediction in a way it is closer to the function we need to convert into ONNX.

def predict(W, H, row_index, col_index):
    return np.dot(W[row_index, :], H[:, col_index])


got = []
for i in range(mat.shape[0]):
    for j in range(mat.shape[1]):
        got.append((i, j, predict(W, H, i, j)))

print(got)

[(0, 0, 1.8940507356352687), (0, 1, 0.10912372262184848), (0, 2, 0.3072453141962623), (0, 3, 0.3072453141962623), (1, 0, 1.014673742790866), (1, 1, 0.9848866016414943), (1, 2, 0.0), (1, 3, 0.0), (2, 0, 1.1066115912409111), (2, 1, 0.0), (2, 2, 0.19083752823558645), (2, 3, 0.19083752823558645), (3, 0, 1.1066115912409111), (3, 1, 0.0), (3, 2, 0.19083752823558645), (3, 3, 0.19083752823558645), (4, 0, 0.9470253678176344), (4, 1, 0.05456186131092424), (4, 2, 0.15362265709813114), (4, 3, 0.15362265709813114)]

Conversion into ONNX

There is no implemented converter for NMF as the function we plan to convert is not transformer or a predictor. The following converter does not need to be registered, it just creates an ONNX graph equivalent to function predict implemented above.

def nmf_to_onnx(W, H, op_version=12):
    """
    The function converts a NMF described by matrices
    *W*, *H* (*WH* approximate training data *M*).
    into a function which takes two indices *(i, j)*
    and returns the predictions for it. It assumes
    these indices applies on the training data.
    """
    col = OnnxArrayFeatureExtractor(H, "col")
    row = OnnxArrayFeatureExtractor(W.T, "row")
    dot = OnnxMul(col, row, op_version=op_version)
    res = OnnxReduceSum(dot, output_names="rec", op_version=op_version)
    indices_type = np.array([0], dtype=np.int64)
    onx = res.to_onnx(
        inputs={"col": indices_type, "row": indices_type},
        outputs=[("rec", FloatTensorType((None, 1)))],
        target_opset=op_version,
    )
    return onx


model_onnx = nmf_to_onnx(W.astype(np.float32), H.astype(np.float32))
print(model_onnx)

ir_version: 7
opset_import {
  domain: ""
  version: 12
}
opset_import {
  domain: "ai.onnx.ml"
  version: 1
}
producer_name: "skl2onnx"
producer_version: "1.16.0"
domain: "ai.onnx"
model_version: 0
graph {
  node {
    input: "Ar_ArrayFeatureExtractorcst"
    input: "col"
    output: "Ar_Z0"
    name: "Ar_ArrayFeatureExtractor"
    op_type: "ArrayFeatureExtractor"
    domain: "ai.onnx.ml"
  }
  node {
    input: "Ar_ArrayFeatureExtractorcst1"
    input: "row"
    output: "Ar_Z02"
    name: "Ar_ArrayFeatureExtractor1"
    op_type: "ArrayFeatureExtractor"
    domain: "ai.onnx.ml"
  }
  node {
    input: "Ar_Z0"
    input: "Ar_Z02"
    output: "Mu_C0"
    name: "Mu_Mul"
    op_type: "Mul"
    domain: ""
  }
  node {
    input: "Mu_C0"
    output: "rec"
    name: "Re_ReduceSum"
    op_type: "ReduceSum"
    domain: ""
  }
  name: "OnnxReduceSum"
  initializer {
    dims: 2
    dims: 4
    data_type: 1
    float_data: 1.98630548
    float_data: 0
    float_data: 0.342542619
    float_data: 0.342542619
    float_data: 0.900879145
    float_data: 0.874432623
    float_data: 0
    float_data: 0
    name: "Ar_ArrayFeatureExtractorcst"
  }
  initializer {
    dims: 2
    dims: 5
    data_type: 1
    float_data: 0.896955
    float_data: 0
    float_data: 0.557120502
    float_data: 0.557120502
    float_data: 0.448477507
    float_data: 0.124793746
    float_data: 1.126315
    float_data: 0
    float_data: 0
    float_data: 0.0623968728
    name: "Ar_ArrayFeatureExtractorcst1"
  }
  input {
    name: "col"
    type {
      tensor_type {
        elem_type: 7
        shape {
          dim {
          }
        }
      }
    }
  }
  input {
    name: "row"
    type {
      tensor_type {
        elem_type: 7
        shape {
          dim {
          }
        }
      }
    }
  }
  output {
    name: "rec"
    type {
      tensor_type {
        elem_type: 1
        shape {
          dim {
          }
          dim {
            dim_value: 1
          }
        }
      }
    }
  }
}

Let’s compute prediction with it.

sess = InferenceSession(
    model_onnx.SerializeToString(), providers=["CPUExecutionProvider"]
)


def predict_onnx(sess, row_indices, col_indices):
    res = sess.run(None, {"col": col_indices, "row": row_indices})
    return res


onnx_preds = []
for i in range(mat.shape[0]):
    for j in range(mat.shape[1]):
        row_indices = np.array([i], dtype=np.int64)
        col_indices = np.array([j], dtype=np.int64)
        pred = predict_onnx(sess, row_indices, col_indices)[0]
        onnx_preds.append((i, j, pred[0, 0]))

print(onnx_preds)

[(0, 0, 1.8940508), (0, 1, 0.10912372), (0, 2, 0.3072453), (0, 3, 0.3072453), (1, 0, 1.0146737), (1, 1, 0.9848866), (1, 2, 0.0), (1, 3, 0.0), (2, 0, 1.1066115), (2, 1, 0.0), (2, 2, 0.19083752), (2, 3, 0.19083752), (3, 0, 1.1066115), (3, 1, 0.0), (3, 2, 0.19083752), (3, 3, 0.19083752), (4, 0, 0.9470254), (4, 1, 0.05456186), (4, 2, 0.15362266), (4, 3, 0.15362266)]

The ONNX graph looks like the following.

pydot_graph = GetPydotGraph(
    model_onnx.graph,
    name=model_onnx.graph.name,
    rankdir="TB",
    node_producer=GetOpNodeProducer("docstring"),
)
pydot_graph.write_dot("graph_nmf.dot")
os.system("dot -O -Tpng graph_nmf.dot")
image = plt.imread("graph_nmf.dot.png")
plt.imshow(image)
plt.axis("off")

(-0.5, 1654.5, 846.5, -0.5)

Versions used for this example

print("numpy:", np.__version__)
print("scikit-learn:", sklearn.__version__)
print("onnx: ", onnx.__version__)
print("onnxruntime: ", onnxruntime.__version__)
print("skl2onnx: ", skl2onnx.__version__)

numpy: 1.23.5
scikit-learn: 1.4.dev0
onnx:  1.15.0
onnxruntime:  1.16.0+cu118
skl2onnx:  1.16.0