[Example] support more ONNX image classification models

[huningxin]

The existing example (https://github.com/intel/webml-polyfill/tree/master/examples/squeezenet) only supports ONNX SqueezeNet model for image classification. This task is to make the example more general to support more ONNX image classification models hosted here:

https://github.com/onnx/models#image_classification

It requires:

1. refactor the example to load, compile and execute multiple models
2. identify the unsupported ops if the model is not supported

[pinzhenx]

o - Not supported by current impl. x - Not supported by current impl. & Not supported directly by our API

Image Classification

Model	BatchNormalization	Add	Mul	Reshape	LRN	AveragePool	Gemm	Unsqueeze	Sum	Transpose
MobileNet v2-1.0	x	o		o
GoogleNet				o	x	o	x
AlexNet				o	x		x
DenseNet-121	x	o	o			o		x
R-CNN ILSVRC13				o	x		x
CaffeNet				o	x		x
Inception v1				o	x	o	x
Inception v2	x	o	o	o		o	x	x
VGG w/ bn	x						x
ZFNet-512				o	x		x
ResNet-152v1	x	o					x
ResNet-152v2	x	o		o			x
ShuffleNet	x			o		o	x		x	x

Other Applications

Model	BatchNormalization	Add	Sub	Mul	Div	Reshape	Gemm	Sum	MatMul	LeakyRelu	ImageScaler	Identity	PRelu
CNTK on MNIST		o				o			x
Tiny YOLOv2	x									x	x
DUC (ResNet101)	x					o		x
ArcFace (ResNet100)	x	o	x	o			x					x	x
Emotion FERPlus		o	x		x	o			x

[huningxin]

@pinzhenx, thanks for the investigation! As the next step, please help investigate the possibility to fuse the batchnormalization op into conv op in JavaScript.

[huningxin]

open https://github.com/intel/webml-polyfill/issues/240

[huningxin]

@pinzhenx, could you please create a wiki page to capture the required ops of ONNX image classification as TFLite one: https://github.com/intel/webml-polyfill/wiki/Tflite-Model-for-Image-Classification-and-the-Ops-They-Include. Thanks!

[SharanSMenon]

Will densenet be supported?