OpenVINO之四：转换TensorFlow模型

1

2 OpenVINO支持的TENSORFLOW模型

• Inception v1 、 Inception v2、 Inception v3 、Inception V4 、 Inception ResNet v2
• MobileNet v1 128、 MobileNet v1 160 、MobileNet v1 224
• NasNet Large 、 NasNet Mobile
• ResidualNet-50 、ResidualNet-101 、ResidualNet-152
• VGG-16 、VGG-19

3 OpenVINO支持的TENSORFLOW层与其在Intermediate Representation (IR)中的对应关系

NUMBER	OPERATION NAME IN TENSORFLOW	LAYER NAME IN THE INTERMEDIATE REPRESENTATION
1	Transpose	Permute
2	LRN	Norm
3	Split	Split
4	SplitV	Split
5	FusedBatchNorm	ScaleShift (can be fused into Convolution or FullyConnected)
6	Relu6	Clamp
7	DepthwiseConv2dNative	Convolution
8	ExpandDims	Constant propagation
9	Slice	Split
10	ConcatV2	Concat
11	MatMul	FullyConnected
12	Pack	Reshapes and Concat
13	StridedSlice	Constant propagation and several cases when StridedSlice can be expressed with Splits
14	Prod	Constant propagation
15	Const	Constant propagation
16	Tile	Tile
17	Placeholder	Input
18	Pad	Fused into Convolution or Pooling layers (not supported as single operation)
19	Conv2D	Convolution
20	Conv2DBackpropInput	Deconvolution
21	Identity	Ignored, does not appear in the IR
22	Add	Eltwise(operation = sum)
23	Mul	Eltwise(operation = mul)
24	Maximum
25	Rsqrt	Power(power=-0.5)
26	Neg	Power(scale=-1)
27	Sub	Eltwise(operation = sum) + Power(scale=-1)
28	Relu	ReLU
29	AvgPool	Pooling (pool_method=avg)
30	MaxPool	Pooling (pool_method=max)
31	Mean	Pooling (pool_method = avg); spatial dimensions are supported only
32	RandomUniform	Not supported
33	BiasAdd	Fused or converted to ScaleShift
34	Reshape	Reshape
35	Squeeze	Reshape
36	Shape	Constant propagation (or layer generation if the “–keep_shape_ops” command line parameter has been specified)
37	Softmax	SoftMax
38	SpaceToBatchND	Supported in a pattern when converted to Convolution layer dilation attribute, Constant propagation
39	BatchToSpaceND	Supported in a pattern when converted to Convolution layer dilation attribute, Constant propagation
40	StopGradient
41	Square Constant propagation
42	Sum	Pool(pool_method = avg) + Eltwise(operation = mul)
43	Range	Constant propagation
44	CropAndResize	ROIPooling (if the the method is ‘bilinear’)
45	ArgMax	ArgMax
46	DepthToSpace	Reshape + Permute + Reshape (works for CPU only because of 6D tensors)
47	ExtractImagePatches	ReorgYolo
48	ResizeBilinear	Interp
49	ResizeNearestNeighbor	Resample
50	Unpack	Split + Reshape (removes dimension being unpacked) if the number of parts is equal to size along given axis
51	AddN	Several Eltwises
52	Concat	Concat
53	Minimum	Power(scale=-1) + Eltwise(operation = max) + Power(scale=-1)
54	Unsqueeze	Reshape
55	RealDiv	Power(power = -1) and Eltwise(operation = mul)
56	SquaredDifference	Power(scale = -1) + Eltwise(operation = sum) + Power(power = 2)
57	Gather	Gather
58	GatherV2	Gather
59	ResourceGather	Gather
60	Sqrt	Power(power=0.5)
61	Square	Power(power=2)
62	Pad	Pad
63	PadV2	Pad
64	MirrorPad	Pad
65	ReverseSequence	ReverseSequence
66	ZerosLike Constant	propagation

参考资料：
1 Converting a TensorFlow* Model
2 Supported Framework Layers