ONNX

Introduction:

This page tracks the current status and development to support ONNX on Mxnet. We currently support import of ONNX models into Mxnet.

ONNX-MXNET

We have two repositories for the import module:

onnx-mxnet repository inside onnx https://github.com/onnx/onnx-mxnet ( soon to be deprecated)
A newly created and refactored module inside Mxnet -> contrib. https://github.com/apache/incubator-mxnet/tree/master/python/mxnet/contrib/onnx

Operator Coverage:

This table keeps track of the status of all ONNX operators supported by Mxnet.
ONNX backend test script reports the coverage on the operators and attributes.

Operator	Test Coverage	Mxnet
Abs	Yes	OK
Add	Yes	OK
And
ArgMax		OK
ArgMin		OK
AveragePool	Yes	OK
BatchNormalization	Yes	OK
Cast	Yes	OK
Ceil		OK
Clip	Yes	NOT OK, under development
Concat	Yes	OK
Constant	Yes	NOT OK
Conv	Yes	OK
ConvTranspose	Yes (local)	OK
DepthToSpace		NOT OK, under development
Div	Yes	OK
Dropout		OK
Elu	Yes	OK
Equal	Yes	NOT OK
Exp	Yes	OK
Flatten	Yes	PARTIAL, only supports default axis=1
Floor		OK
GRU		NOT OK , under development
Gather	Yes	NOT OK, under development
Gemm	Yes	OK
GlobalAveragePool	Yes	OK
GlobalLpPool		NOT OK, under development
GlobalMaxPool		OK
Greater		NOT OK, under development
HardSigmoid		NOT OK, under development
Hardmax		NOT OK, under development
InstanceNormalization		NOT OK, under development
LRN	Yes	OK
LSTM		NOT OK, under development
LeakyRelu	Yes	OK
Less		NOT OK, under development
Log	Yes	OK
LogSoftMax		NOT OK, under development
LpNormalization		NOT OK, under development
LpPool		NOT OK, under development
LpPool	Yes	OK
Max	Yes	OK
MaxPool	Yes	OK
MaxRoiPool		NOT OK, under development
Mean		NOT OK, under development
Min	Yes	OK
Mul	Yes	OK
Neg	Yes	OK
Not		NOT OK, under development
Or		NOT OK, under development
PRelu	Yes	OK
Pad	Yes	OK
Pow	Yes	Partial, only supports default axis=1
RNN		NOT OK, under development
RandomNormal		OK
RandomNormalLike		OK
RandomUniform		OK
RandomUniformLike		OK
Reciprocal		OK
ReduceL1		NOT OK, under development
ReduceL2		NOT OK, under development
ReduceLogSum		NOT OK, under development
ReduceLogSumExp		NOT OK, under development
ReduceMax		OK
ReduceMean		OK
ReduceMin		OK
ReduceProd		OK
ReduceSum		OK
ReduceSumSquare		NOT OK, under development
Relu	Yes	OK
Reshape	Yes	OK
Selu	Yes	NOT OK, under development
Sigmoid	Yes	OK
Slice	Yes	NOT OK, under development
Softmax	Yes	PARTIAL, only supports default axis=1
Softplus	Yes	NOT OK, under development
Softsign		NOT OK, under development
SpaceToDepth		NOT OK, under development
Split	Yes	NOT OK, under development
Sqrt		OK
Squeeze		OK
Sub		OK
Sum	Yes	OK
Tanh	Yes	OK
Tile		NOT OK, under development
Transpose	Yes	OK
Xor		NOT OK, under development
experimental ATen		NOT OK, under development
experimental Affine		NOT OK, under development
experimental ConstantFill		NOT OK, under development
experimental Crop		NOT OK, under development
experimental FC		NOT OK, under development
experimental GRUUnit		NOT OK, under development
experimental GivenTensorFill		NOT OK, under development
experimental Identity		NOT OK, under development
experimental ImageScaler		NOT OK, under development
experimental MeanVarianceNormalization		NOT OK, under development
experimental ParametricSoftplus		NOT OK, under development
experimental Scale		NOT OK, under development
experimental ScaledTanh		NOT OK, under development
experimental ThresholdedRelu		NOT OK, under development
experimental Upsample		NOT OK, under development

ONNX models coverage:

ONNX backend tests include below models. The table lists current status on Mxnet:

ONNX Models	Mxnet Support
bvlc_alexnet	OK
densenet121	OK
inception_v1	NOT OK, precision difference due to issue
inception_v2	NOT OK, precision difference due to issue
resnet50	OK
shufflenet	NOT OK, precision difference due to issue
squeezenet	OK
vgg16	OK
vgg19	OK
bvlc_googlenet	OK
bvlc_caffenet	OK
bvlc-rcnn-ilsvrc13	OK

Page tree

ONNX

ONNX-MXNET