v1.2.2a3

Support ONNX v1.2.1, 1.2.2, 1.3.0. onnx_chainer.export() method now supports opset_version option to specify the opset version of exported ONNX file.

v1.1.1a2

Add a description to README and fix the PyPI upload.

v1.1.1a1

This is the release of onnx-chainer v1.1.1a1. This version supports ONNX v1.1.1 and tested with MXNet 1.2.0 and TVM at this commit ID: ebdde3c277a9807a67b233cecfaf6d9f96c0c1bc.

Main Update

• Refactored export.py and switched to use chainer.FunctionHook to catch the arguments to every functions called in a network which is converted into ONNX.
• Added two compatibility tests with other framework. Now the exported ONNX files of VGG16 and ResNet50 written in Chainer are confirmed that the same outputs can be obtained after importing them using MXNet and NNVM/TVM.
• We added two examples of
  • Export Chainer models into ONNX and load it from MXNet
  • Export Chainer models into ONNX and compile it using NNVM

v1.0.0a1

Support ONNX v1

v0.2.1b4

Add some tests to compare the output of Chainer model and exported model using caffe2.

v0.2.1b3

ONNX-Chainer

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

Requirements

• onnx==0.2.1
• chainer>=3.1.0

Installation

See INSTALL.md

Quick Start

import numpy as np
import chainer.links as L
import onnx_chainer

model = L.VGG16Layers()

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

onnx_chainer.export(model, x, filename='VGG16.onnx')

Supported Functions

Currently 50 Chainer Functions are supported to export in ONNX format.

Activation

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

Array

• Cast
• Concat
• Depth2Space
• Pad ¹²
• Reshape
• Space2Depth
• SplitAxis
• Squeeze
• Tile
• Transpose

Connection

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

Math

• Add
• Absolute
• Div
• Mul
• Neg
• PowVarConst
• Sub
• Clip
• Exp
• Identity
• MatMul ⁴
• Maximum
• Minimum
• Sqrt
• SquaredDifference
• Sum

Noise

• Dropout ⁵

Pooling

• AveragePooling2D
• AveragePoolingND
• MaxPooling2D
• MaxPoolingND

Normalization

• BatchNormalization
• FixedBatchNormalization
• LocalResponseNormalization

---

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: Current ONNX doesn't support transpose options for matmul ops

5: In test mode, all dropout layers aren't included in the exported file

v0.2.1b1

Added many functions. Now it supports:

Activation

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

Array

• Cast
• Concat
• Depth2Space
• Pad
• Reshape
• Space2Depth
• SplitAxis
• Squeeze
• Tile
• Transpose

Connection

• Convolution2DFunction
• LinearFunction

Pooling

• AveragePooling2D
• MaxPooling2D

Normalization

• BatchNormalization
• FixedBatchNormalization

Math

• Add
• Sub
• Mul
• Neg
• Absolute
• Div

v0.2.1a2

Basic function support with naive implementation