- ONNX 和 MLPerf 提供了 ONNX 模型, 可以直接使用
- 未提供 ONNX 模型的, 需要使用官方代码训练出 Tensorflow 或 PyTorch 模型
- 开源框架模型转换为 ONNX 模型
- 测试准确率
序号 | 模型 & Paper | 模型 ONNX/PT/PB | ONNX 转换 | 数据集 |
---|---|---|---|---|
1 | resnet50 | MLPerf resnet50 ONNX | imagenet2012 homepage, imagenet2012 subset | |
2 | Bert-Large | MLPerf Bert-Large ONNX | SQuAD v1.1 | |
3 | Bert-Base | Bert int8 ONNX | ||
4 | YOLO_v5 | YOLOv5 pt->ONNX, YoloV5 Official , YoloV5 pt models | ||
5 | VGG16 | VGG16, VGG16 int8 ONNX | ||
6 | Tacotron2 | Tacotron2 代码可以生成 ONNX 模型 | ||
7 | MaskRCNN | MaskRCNN ONNX | ||
8 | Transformer | nvidia Transformer .pt | ||
9 | TDNN | |||
10 | TextCNN | |||
11 | RNNT | RNNT ONNX | ||
12 | ||||
13 | seq2seq | |||
14 | DBnet | |||
15 | DLRM | DLRM ONNX | ||
16 | ViT | |||
17 | resnet18 | resnet18 ONNX | ||
18 | resnet34 | resnet34 ONNX | ||
19 | resnet101 | resnet101 pth | code & onnx convert | ImageNet1K |
20 | resnet101 | resnet101 ONNX | ||
21 | resNeXt | resnext pth | code & onnx convert | ImageNet1K |
22 | YOLO_v3_Tiny | Yolov3-Tiny ONNX | ||
23 | YOLO_v3 | Yolov3 ONNX | ||
24 | Inception_v1 | Inception_v2 | ||
25 | Inception_v2 | Inception_v2 | ||
26 | Inception_v3 | Inception_v3 | ||
27 | Inception_v4 | v4 tf | ||
28 | MobileNet | mobilenet_v1 tf | ||
29 | MobileNet_v2 | mobilenet_v2 pt | code & onnx convert | |
30 | MobileNet_v3 | |||
31 | ShuffleNet | pt | ImageNet1K | |
32 | R-FCN | tf | ImageNet1K | |
33 | FPN | pt | ImageNet1K | |
34 | SSD_mobilenetv1 | SSD_mb1 .pth | code & onnx convert (L26 enable) | COCO2017 |
35 | SSD | SSD ONNX | ||
36 | SSD-Resnext50 800x800 | SSD ONNX | openimages | |
37 | SSD-Resnet34 | SSD-Resnet34 | ||
38 | FasterRCNN-R50-FPN | FasterRCNN ONNX | ||
39 | FasterRCNN | pth | COCO2017 | |
40 | FasterRCNN-R50 | mmdetection | VOC2007 | |
41 | FasterRCNN | pth | COCO2017 | |
42 | SoloV2 | pth | COCO2017 | |
43 | deeplabv3 | pth | code & onnx converter | COCO2017 |
44 | PSPNet | pth | COCO2017 | |
45 | retinaface | pth_code, pth_googledrive | widerface | |
46 | SoloV2 | pth | COCO2017 | |
47 | 3D-UNet | 3D-UNet ONNX |
VGG16 模型结构代码 : 未使用BN, 5组(3x3)卷积之间用maxpooling, FC 使用(1x1)卷积替换
VGG16 ckpt 模型, Tensorflow ckpt 转 ONNX
使用 nvidia Transformer 代码进行推理, 下载上述 .pt 模型, 准备 en ->
序号 | 算子 | VGG16 | YoloV3 | YoloV5 | MaskRCNN |
---|---|---|---|---|---|
1 | Log | ✓ | |||
2 | Sub | ✓ | |||
3 | Floor | ✓ | |||
4 | Unsqueeze | ✓ | |||
5 | Scatter | ✓ | |||
6 | Conv | ✓ | |||
7 | ConvTranspose | ✓ | |||
8 | TopK | ✓ | |||
9 | Relu | ✓ | |||
10 | Resize | ✓ | |||
11 | Cast | ✓ | |||
12 | Expand | ✓ | |||
13 | Sigmoid | ✓ | |||
14 | Add | ✓ | |||
15 | NonMaxSuppression | ✓ | |||
16 | Gather | ✓ | |||
17 | MaxPool | ✓ | |||
18 | Mul | ✓ | |||
19 | Less | ✓ | |||
20 | Constant | ✓ | |||
21 | Shape | ✓ | |||
22 | Split | ✓ | |||
23 | ReduceMin | ✓ | |||
24 | And | ✓ | |||
25 | Gemm | ✓ | |||
26 | Equal | ✓ | |||
27 | Reshape | ✓ | |||
28 | Softmax | ✓ | |||
29 | Flatten | ✓ | |||
30 | NonZero | ✓ | |||
31 | Clip | ✓ | |||
32 | RoiAlign | ✓ | |||
33 | Not | ✓ | |||
34 | Concat | ✓ | |||
35 | Div | ✓ | |||
36 | Greater | ✓ | |||
37 | Exp | ✓ | |||
38 | Transpose | ✓ | |||
39 | Squeeze | ✓ | |||
40 | Slice | ✓ | |||
41 | ConstantOfShape | ✓ | |||
42 | Sqrt | ✓ |
序号 | 算子 | Bert | Tacotron2 | RNNT |
---|---|---|---|---|
1 | ||||
2 | ||||
3 | ||||
4 | ||||
5 | ||||
6 | ||||
7 | ||||
8 | ||||
9 | ||||
10 | ||||
11 | ||||
12 | ||||
13 | ||||
14 | ||||
15 | ||||
16 | ||||
17 | ||||
18 | ||||
19 |