Pre-Release on the next version of YOLOv7

[lucasjinreal]

We have been massively experimented 2 kinds of models in the last few months:

• Large models no matter on latency;
• Small models which consider both latency and accuracy (mainly on CPU);

And here we made some breakthrough so far, we have made:

• YOLOTr-ConvnextTiny which get mAP 45.9 (new with aug) with very low latency on GPU;
• YOLO2Go series: yolo2go_mobilenetv2 baseline get mAP 30;
• YOLO2Go series: yolo2go_shuffletnetv2_lite_dsp get mAP 25 with minimal weights of 4M on fp16;
• YOLO2Go series: yolo2go_mobileone_lite_dsp get mAP 30.1 with the fastest speed compare above models;

We will release these new models very soon. Especially these small models, it was really fast and high accurate due to stable and rich components in yolov7 framework.

note: DSP is our developed high accuracy neck type. All small models above doesn't using Focus layer or SiLU activation, all using normal LeakyReLU or ReLU activation for more easy deployment!