1
2 OpenVINO支持的ONNX模型
2-1 支持的公共模型
- bvlc_alexnet , bvlc_googlenet , bvlc_reference_caffenet , bvlc_reference_rcnn_ilsvrc13
- inception_v1, inception_v2
- resnet50
- squeezenet
- densenet121
- emotion_ferplus
- mnist
- shufflenet
- VGG19
- zfnet512
2-2 支持的Pytorch模型
Torchvision Models：
- alexnet,
- densenet121, densenet161, densenet169, densenet201,
- resnet101, resnet152, resnet18, resnet34, resnet50,
- vgg11, vgg13, vgg16, vgg19
Pretrained Models：
- alexnet,
- fbresnet152,
- resnet101, resnet152, resnet18, resnet34, resnet152, resnet18, resnet34, resnet50, resnext101_32x4d, resnext101_64x4d,
- vgg11
2-3 支持的PaddlePaddle模型
- fit_a_line
- recognize_digits
- VGG16
- ResNet
- MobileNet
- SE_ResNeXt
- Inception-v4
3 OpenVINO支持的ONNX层与其在Intermediate Representation (IR)中的对应关系
NUMBER | OPERATOR NAME IN ONNX* | LAYER TYPE IN THE INTERMEDIATE REPRESENTATION |
1 | Add | Eltwise(operation = sum) (added ‘axis’ support) |
2 | AveragePool | Pooling (pool_method=avg) |
3 | BatchNormalization | ScaleShift (can be fused into Convlution or FC) |
4 | Concat | Concat |
5 | Constant | Will be removed on constant propagation step |
6 | Conv | Convolution |
7 | ConvTranspose | Deconvolution (added auto_pad and output_shape attributes support)) |
8 | Div | Eltwise(operation = mul)->Power |
9 | Dropout | Ignored, does not apeear in IR |
10 | Elu | Activation (ELU) |
11 | Flatten | Reshape |
12 | Gemm | FullyConnected |
13 | GlobalAveragePool | Pooling (pool_method=avg) |
14 | Identity | Ignored, does not appear in IR |
15 | LRN | Norm |
16 | LeakyRelu | ReLU |
17 | MatMul | FullyConnected |
17 | MaxPool | Pooling (pool_method=max) |
19 | Mul | Eltwise(operation = mul) (added ‘axis’ support) |
20 | Relu | ReLU |
21 | Reshape | Reshape |
22 | Shape | Constant propagation |
23 | Softmax | SoftMax |
24 | Squeeze | Reshape |
25 | Sub | Power->Eltwise(operation = sum) |
26 | Sum | Eltwise(operation = sum) |
27 | Transpose | Permute |
28 | Unsqueeze | Reshape |
29 | Upsample | Resample |
30 | ImageScaler | ScaleShift |
31 | Affine | ScaleShift |
32 | Reciprocal | Power(power=-1) |
33 | Crop | Split |
34 | Tanh | Activation (operation = tanh) |
35 | Sigmoid | Activation (operation = sigmoid) |
36 | Pow | Power |
37 | ConvTranspose | |
38 | Gather | Constant propagation |
39 | Constant_fill | Constant propagation |
40 | ReduceMean | Reshape + Pooling(pool_method=avg) + Reshape |
41 | ReduceSum | Reshape + Pooling(pool_method=avg) + Power(scale=reduce_dim_size) + Reshape |
42 | Gather | Gather |
43 | Gemm | GEMM |
44 | GlobalMaxPool | Pooling (pool_method=max) |
45 | Neg | Power(scale=-1) |
46 | Pad | Pad |
47 | ArgMax | ArgMax |
48 | Clip | Clamp |
49 | DetectionOutput (experimental) | DetectionOutputONNX |
50 | PriorBox (experimental) | PriorBoxONNX |
51 | RNNSequence | TensorIterator(RNNCell) |
52 | GRUSequence | TensorIterator(GRUCell) |
53 | LSTMSequence | TensorIterator(LSTMCell) |
参考资料：
1 Converting a ONNX* Model
2 Supported Framework Layers
