freeze metrics

Dmitry Kurtaev 2017-08-15 12:52:38 +03:00
parent c7a6ae1e5e
commit 67755b5039

@ -21,26 +21,28 @@ Best time of single image forward pass (in milliseconds):
#### CPU
All calculations are done in float32.
| Model | DNN, C++ | DNN, Halide | Intel-Caffe, MKLDNN | TensorFlow | Torch w. MKL |
|----------------:|----------:|--------:|---------:|---------:|---------:|
| GoogLeNet | 19.50 | 33 | **9.6** | | |
| AlexNet | 14.68 | 22.4 | 13 | | |
| ResNet-50 | 42.56 | 74 | **24.5** | | |
| SqueezeNet v1.1 | 4.96 | 6.4 | **3.35** | | |
| Inception-5h | 19.58 | 32 | | **17.9** | |
| ENet @ 512x256 | 45.79 | **41** | | | 240 |
Actual measurements in progress: https://github.com/opencv/opencv/pull/9313
> | Model | DNN, C++ | DNN, Halide | Intel-Caffe, MKLDNN | TensorFlow | Torch w. MKL |
> |----------------:|----------:|--------:|---------:|---------:|---------:|
> | GoogLeNet | 19.50 | 33 | **9.6** | | |
> | AlexNet | 14.68 | 22.4 | 13 | | |
> | ResNet-50 | 42.56 | 74 | **24.5** | | |
> | SqueezeNet v1.1 | 4.96 | 6.4 | **3.35** | | |
> | Inception-5h | 19.58 | 32 | | **17.9** | |
> | ENet @ 512x256 | 45.79 | **41** | | | 240 |
#### GPU (OpenCL 2.0):
All computations in float-32.
| Model | DNN, Halide| clDNN | clCaffe |
|----------------:|------------:|----------:|-------------:|
| GoogLeNet | 93.56 | **28** | 181 |
| AlexNet | 49 | **14.65** | 27 |
| ResNet-50 | 189 | | **132** |
| SqueezeNet v1.1 | 15.8 | **8.0** | 18.2 |
| Inception-5h | 89.5 | | |
| ENet @ 512x256 | 39.67 | | |
Actual measurements in progress: https://github.com/opencv/opencv/pull/9313
> | Model | DNN, Halide| clDNN | clCaffe |
> |----------------:|------------:|----------:|-------------:|
> | GoogLeNet | 93.56 | **28** | 181 |
> | AlexNet | 49 | **14.65** | 27 |
> | ResNet-50 | 189 | | **132** |
> | SqueezeNet v1.1 | 15.8 | **8.0** | 18.2 |
> | Inception-5h | 89.5 | | |
> | ENet @ 512x256 | 39.67 | | |
#### References
* OpenCV's deep learning module, https://github.com/opencv/opencv_contrib/tree/master/modules/dnn.