Make the pre-processing consistent with yolov5

[zhiqwang]

🚀 The feature and motivation

As mentioned in #265 (comment), the YOLOv5.load_from_yolov5() strategy will introduce minor differences with ultralytics, now there are two options to solve this problem:

• Make the pre-processing behave the same with yolov5. The advantage of this approach is that there is no need to write code on the C++ backend.
• Just use YOLO.load_from_yolov5() and implement the pre-precessing in the C++ backend. This requires to modify the example we provide. And the main difference between this strategy and the official yolov5 is that we don't need to write the post-processing (nms) module on the C++ backend.

In fact, the main ops needed for preprocessing is interpolate, yolov5 uses the cv2.resize operator, we use torch.nn.functional.interpolate in yolort. (And the ops in cv2 can't be traced or scripted.) The implementation of the interpolate operator is not entirely consistent across frameworks.

UPDATED: Fortunately, the interpolation operator of PyTorch is aligned with that of OpenCV for the float type, however, PyTorch's interpolate does not currently support integers type: pytorch/pytorch#5580 , a little loss of precision will occur here.

https://github.com/zhiqwang/yolov5-rt-stack/blob/949fa0b08dd336c21dc051594941af457a312b21/yolort/v5/utils/augmentations.py#L126

https://github.com/zhiqwang/yolov5-rt-stack/blob/949fa0b08dd336c21dc051594941af457a312b21/yolort/models/transform.py#L286-L293

Additional context

https://pytorch.org/blog/introducing-torchvision-new-multi-weight-support-api/#associated-meta-data---preprocessing-transforms

Originally posted by @zhiqwang in #271 (comment) cc @mattpopovich .

[zhiqwang]

Hi @mattpopovich ,

Here is a GOOD news! I think we've finally resolved the issue in #293 , and I'll update the comparison documentation in a follow-up PR.

I haven't done a complete test on the speed, but since we are using basically the same operators both before and after the change, it' s likely that the code will not be slower after the change. We will pay more attention to the performance part in subsequent updates. We would welcome additional testing on different devices here.

As such I'm closing this thread, we'd be happy to hear more from you about yolort, and feel free to create another tickets if you have more questions.

[zhiqwang]

FYI, we've added the doc for comparison between yolov5 & yolort:
https://zhiqwang.com/yolov5-rt-stack/notebooks/comparison-between-yolort-vs-yolov5.html