You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 13 Next »

Introduction:

This page tracks the current status and development  to support ONNX  on Mxnet. We currently support:

  • import of ONNX models into Mxnet.
  • export MXNet models to ONNX format.


ONNX-MXNET Import Module

We have two repositories for the import module:

  1. onnx-mxnet repository inside onnx https://github.com/onnx/onnx-mxnet  (deprecated)
  2. A newly created and refactored module inside Mxnet -> contrib. https://github.com/apache/incubator-mxnet/tree/master/python/mxnet/contrib/onnx/onnx2mx


MXNET-ONNX Export module

A new module has been added to MXNet repository to export MXNet models to ONNX model format. https://github.com/apache/incubator-mxnet/tree/master/python/mxnet/contrib/onnx/mx2onnx

 

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.

OperatorTest CoverageMxnet (Import)Mxnet (Export)
AbsYesOKOK
Acos


AddYesOKOK
AndYes 

OK


ArgMaxYesOKOK
ArgMinYes OKOK
Asin


Atan


AveragePoolYesOKOK
BatchNormalizationYesOKOK
CastYesOKOK
CeilYes OKOK
ClipYesOK
ConcatYesOKOK
ConstantYesOK
ConvYesOKOK
ConvTransposeYes OKOK
Cos


DepthToSpace
NOT OK, under development
DivYesOKOK
Dropout YesOKOK
EluYesOKOK
EqualYesOK
ExpYesOKOK
FlattenYesPARTIAL, only supports default axis=1OK
Floor YesOKOK
GRU
NOT OK , under development
GatherYesOK
GemmYesOKOK
GlobalAveragePoolYesOKOK
GlobalLpPool YesOK, In review
GlobalMaxPoolYesOKOK
Greater YesOK
HardSigmoid YesOK
Hardmax YesOK, In review
Identity


InstanceNormalization YesOK, In review
LRNYes OKOK
LSTM
NOT OK, under development
LeakyReluYesOKOK
LessYes OK
LogYesOKOK
LogSoftMaxYes OK
LpNormalization OK, In reviewOK
LpPoolYesOK, In review
MatMul


MaxYesOKOK
MaxPoolYesOKOK
MaxRoiPoolYes OK, In review
MeanYes OKOK
MinYesOKOK
MulYesOKOK
Multinomial


NegYesOKOK
NotYes OK
OrYes OK
PReluYesOKOK
PadYesOKOK
PowYesPartial, only supports default axis=1OK
RNN NOT OK, under development
RandomNormalYes OK
RandomNormalLikeYes OK
RandomUniformYes OK
RandomUniformLikeYes OK
ReciprocalYes OKOK
ReduceL1Yes OK
ReduceL2 YesOK
ReduceLogSum YesOK
ReduceLogSumExp YesOK
ReduceMaxYesOKOK
ReduceMeanYesOKOK
ReduceMinYesOKOK
ReduceProdYesOKOK
ReduceSumYesOK
ReduceSumSquareYes OK
ReluYesOKOK
ReshapeYesOKOK
SeluYesNOT OK, under developmentOK
Shape


SigmoidYesOKOK
Sin


Size


SliceYesPartial, supports axis=1OK
SoftmaxYesOKOK
SoftplusYesOKOK
Softsign YesOKOK
SpaceToDepth NOT OK, under development
SplitYesOKOK
SqrtYes OKOK
SqueezeYesOKOK
SubYesOKOK
SumYesOKOK
Tan


TanhYesOKOK
Tile YesOK
TopK


TransposeYesOKOK
Unsqueeze


Upsample


XorYes OK
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 GRUUnit NOT OK, under development
experimental GivenTensorFill NOT OK, under development
experimental If NOT OK
experimental ImageScaler YesOK, In review
experimental Loop


experimental LoopIndex


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

 

ONNX models coverage:

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

 

ONNX ModelsMxnet Support
bvlc_alexnetOK
densenet121OK
inception_v1OK, In review
inception_v2OK, In review
resnet50OK
shufflenetOK, In review
squeezenetOK
vgg16OK
vgg19OK
bvlc_googlenetOK
bvlc_caffenetOK
bvlc-rcnn-ilsvrc13OK

 

 

 

 

 

 

 

 

 

 

 

  • No labels