mmrotate support in mmdeploy

[hust-lidelong]

Could you support mmrotate in mmdeploy? For example, Oriented R-CNN or S2A-Net. Thanks very much.

[RunningLeon]

@hust-lidelong Hi, mmrotate is not supported by far. But you could post your request in this issue #170.

[lvhan028]

#349 is dealing with it.

[hust-lidelong]

#349 is dealing with it.

When to support a two-stage rotated detector in MMRotate? Thanks

[zytx121]

We will support two-stage rotated detector ASAP.

[Worromots]

Are [Rotated RetinaNet, Rotated FasterRCNN, Oriented R-CNN, Gliding Vertex, RTMDET-R] all the rotation detection models supported by mmdeploy? (I obtained this list from the mmdeploy official website). In addition, I have converted Rotated RetinaNet, Gliding Vertex, and Oriented R-CNN into fp16 TensorRT models and tested the mAP of both TensorRT and pth on DOTA1.0. The test results show a large fluctuation. Is this normal?

[RunningLeon]

@BUAA-Worromot hi, something is wrong. Could you test with fp32 precision? BTW, the converted models were not tested on dataset.

[Worromots]

Thanks for your reply. I tried to convert those three models mentioned above, but only gliding_vertex succeed, the other two throw :

[W] [TRT] Skipping tactic 0 due to Myelin error: autotuning: CUDA error 2 allocating 0-byte buffer: out of memory
[08/29/2023-16:28:00] [E] Error[10]: [optimizer.cpp::computeCosts::1981] Error Code 10: Internal Error (Could not find any implementation for node {ForeignNode[Unsqueeze_940...Slice_1018]}.)

even when I set max_workspace_size to 16G.
gliding_vertex with fp32 precision got mAP of 0.8110640048980713.
BTW, are [Rotated RetinaNet, Rotated FasterRCNN, Oriented R-CNN, Gliding Vertex, RTMDET-R] all the rotation detection models supported by mmdeploy?

[RunningLeon]

Could try with newer TRT. The supported model list is this https://mmdeploy.readthedocs.io/en/latest/04-supported-codebases/mmrotate.html#supported-models