onnx2chainer

This tool provides a way to convert ONNX models to Chainer models.

Requirements

• Python >= 3.7.2
• Chainer >= 5.1.0
• ONNX >= 1.3.0
• NumPy >= 1.16.0

The Python packages can be install via pip3 install chainer onnx numpy.

How to install

1. pip3 install -e .
2. python3 -c "import onnx2chainer"
3. python3 setup.py test

How to Use

import onnx
import onnx2chainer

model_onnx = onnx.load("path/to/onnx/model")
model_chainer = onnx2chainer.onnxToChainer(model_onnx)

x = a_xp_array
y = model_chainer(x)

See onnx2chainer_test.py for more details.

How to Test

The test script onnx2chainer_test.py

1. downloads ONNX models as well as test datasets from The ONNX Model Zoo (might be O(GiB)),
2. generate a Chainer model from each of them, and
3. verify the generated models by comparing output of the models and grand-truth outputs, as follows.

Testing bvlc_alexnet.tar.gz            with test_data_set_0: |y_chainer - y_true| / |y_true| = 5.27e-07
Testing bvlc_reference_caffenet.tar.gz with test_data_set_0: |y_chainer - y_true| / |y_true| = 5.24e-08
Testing bvlc_googlenet.tar.gz          with test_data_set_0: |y_chainer - y_true| / |y_true| = 9.00e-07
Testing resnet50.tar.gz                with test_data_set_0: |y_chainer - y_true| / |y_true| = 3.70e-08
Testing mnist.tar.gz                   with test_data_set_0: |y_chainer - y_true| / |y_true| = 5.78e-04
Testing vgg19.tar.gz                   with test_data_set_0: |y_chainer - y_true| / |y_true| = 9.68e-07

Note that the error of MNIST is relatively larger than others because the model does not have softmax.

Supported ONNX Operations

• activation
  • Relu
  • Softmax
• array
  • Concat
  • Split
  • Reshape
  • Squeeze
• connection
  • Conv
  • Gemm
  • MatMul
• math
  • Add
  • Sum
  • Sub
  • Mul
  • Abs
  • ReduceSum
• noise
  • Dropout
• normalization
  • LRN
  • BatchNormalization
• pooling
  • GlobalAveragePool
  • MaxPool
  • AveragePool