Train Yolov7 with non square images without padding and resizing

[shokohigol]

I want to use Yolov7 to train an object detection model with my custom dataset. Can the model be trained with rectangular images without padding? I have a dataset of FHD images, that I want to train Yolov7 with the original size (not the padded one) because, in the padded state, the model input images will be big. The padded part is not useful and inference time will increase. Consider I don't need any augmentation and all images in my dataset are the same size(1080x1920). So, I want to feed images with 1080x1920 size to model and inference with the same size.

[YannickGibson]

I think the only way is to change the source code.