[Feature] Improve inference speed

[dahhei]

What is the feature?

Currently, it appears as though the inference speeds for multiple models that I have tested are slower than other libraries. In an attempt to speed up inference, I have gone down a few avenues and found that:

• MMdeploy's Docker runs on an older version of ubuntu, which is no longer supported on most HPCs
• Batching is not supported in inference for both topdown and bottomup

Any other context?

I test a model from another library DeepLabCut (dlc) which is widely used in the animal tracking field, a bottom-up network, and various top-down networks within mmpose. I ran these models on a 1-hour video of a single mouse that I am tracking the nose and tail-base keypoints. DeepLabCut ended up being twice as fast as any mmpose model that I tested.

I have adjusted the bottomup_demo.py and topdown_demo_with_mmdet.py to exclude writing out the labeled video file to save on time, as I am solely trying to benchmark the inference speed. I also set the visualizer to None in the topdown script to account for an error that I get in regard to the number of keypoints (default is 17, I have 2 in my model).

Thanks for the great library to test a lot of different models easily, and thanks in advance for any input.