the backend of torch supported by opencl
The project is based on artyom-beilis, link:https://github.com/artyom-beilis/pytorch_dlprim/tree/032a6933fed7e9e0cdcf0dbadf9c7f246804fa3b
how to install you can follow the above link.
I add some ops(as follows),but only forward, unsupport backward.
| op |
|---|
| arange |
| argsort |
| bmm |
| gelu |
| index_tensor_out |
| layernorm |
| maxtrix_power |
| max_dim |
| max_pool2d_with_indices |
| nonzero |
| UpsamplingNearest2d |
| filp |
| unique |
We test the partial of CSwin and yolov5 in 3090RTX, the result as follow:
| net | acc | time |
|---|---|---|
| CSwin tiny(our) | 80.2% | 46.17ms |
| CSwin tiny(cuda) | 80.2% | 40.04ms |
| CSWin_base_224(our) | 82.6% | 79.86ms |
| CSWin_base_224(cuda) | 82.5% | 60.70ms |
| yolov5n(our) | ----- | 50.44ms |
| yolov5n(cuda) | ----- | 50.68ms |
| yolov5l6(our) | ----- | 64.82ms |
| yolov5nl6(cuda) | ----- | 63.41ms |
CSwin test used 1K images, yolov5 only test time [have same result at 4 images].