- '-' means use last result from top
- empty means no test results
- onnx convert based on mmdeploy
- model infer fps does not contain pre/post-process time cost
- trt & trt16 based on onnxruntime tensorrt EP
- input array located on cpu (io-binding done by onnxruntime itself)
- more info check onnx test codes
| Name | Attr |
|---|---|
| Sys | Ubuntu 20.04 |
| GPU | NVIDIA GeForce RTX 3080 10G |
| CPU | Intel® Core™ i9-10850K CPU @ 3.60GHz × 20 |
| MEM | 32G |
| Libs | onnxruntime-gpu=1.13 |
| MMPose | input shape | size | cpu infer fps | gpu infer fps | trt infer fps | trt16 infer fps |
|---|---|---|---|---|---|---|
| pvtv2-b2_static_coco | [1, 3, 256, 192] | 116.3m | 4.9 | 73 | 184 | 257 |
| [4, 3, 256, 192] | - | 2.5 | 47 | 106 | 178 | |
| hrnet_w48_dark+_dynamic_cocowholebody | [4, 3, 384, 288] | 254m | 2.9 | 31 | 39 | 83 |
| MMPose Post-process | ||||||
| gaussian_blur_k17 | [4, 133, 96, 72] | 0.19m | 7.9 | 119 | 147 | 142 |