onnx-chainer 1.4.1

Convert a Chainer model into ONNX

ONNX-Chainer

This is an add-on package for ONNX support by Chainer.

Tested environment

• Python 3.5.5, 3.6.7
• ONNX 1.4.0, 1.4.1
  • opset version 7, 8, 9
• Chainer stable, preview
• ONNX-Runtime 0.3.0

(You can still specify all opset versions <= 9, but please noted that opset versions <= 6 are not tested)

Installation

On Ubuntu 14.04/16.04

pip install onnx-chainer

Run Test

1. Install test modules

$ pip install onnx-chainer[test-cpu]

Or, on GPU environment

$ pip install cupy  # or cupy-cudaXX is useful
$ pip install onnx-chainer[test-gpu]

2. Run tests

$ CHAINER_TEST_GPU_LIMIT=0 pytest

Or, on GPU environment

$ pytest

Quick Start

First, install ChainerCV to get the pre-trained models.

import numpy as np

import chainer
import chainercv.links as C
import onnx_chainer

model = C.VGG16(pretrained_model='imagenet')

# Pseudo input
x = np.zeros((1, 3, 224, 224), dtype=np.float32)

onnx_chainer.export(model, x, filename='vgg16.onnx')

Supported Functions

Currently 75 Chainer Functions are supported to export in ONNX format.

Activation

• ClippedReLU
• ELU
• HardSigmoid
• LeakyReLU
• LogSoftmax
• PReLUFunction
• ReLU
• Sigmoid
• Softmax
• Softplus
• Tanh

Array

• Cast
• Concat
• Copy
• Depth2Space
• Dstack
• ExpandDims
• GetItem
• Hstack
• Pad ¹²
• Repeat
• Reshape
• ResizeImages
• Space2Depth
• SplitAxis
• Squeeze
• Stack
• Swapaxes
• Tile
• Transpose
• Vstack
• Where

Connection

• Convolution2DFunction
• ConvolutionND
• Deconvolution2DFunction
• DeconvolutionND
• EmbedIDFunction ³
• LinearFunction

Loss

• SoftmaxCrossEntropy

Math

• Absolute
• Add
• AddConstant
• BroadcastTo
• Clip
• Div
• Exp
• Identity
• LinearInterpolate
• LogSumExp
• MatMul
• Max
• Maximum
• Mean
• Min
• Minimum
• Mul
• MulConstant
• Neg
• PowVarConst
• Prod
• Sqrt
• Square
• Sub
• Sum

Noise

• Dropout ⁴

Normalization

• BatchNormalization
• FixedBatchNormalization
• LocalResponseNormalization
• NormalizeL2

Pooling

• AveragePooling2D
• AveragePoolingND
• MaxPooling2D
• MaxPoolingND
• ROIPooling2D
• Unpooling2D

Contribution

Any contribution to ONNX-Chainer is welcome!

• Python codes follow Chainer Coding Guidelines

---

1: mode should be either 'constant', 'reflect', or 'edge'
2: ONNX doesn't support multiple constant values for Pad operation
3: Current ONNX doesn't support ignore_label for EmbedID
4: In test mode, all dropout layers aren't included in the exported file