training my own data issue

[Smoothing97]

@avanetten hey,thanks for your nice job!
I wanna train my own remote sensing datasethttp://captain.whu.edu.cn/DOTAweb/dataset.html using yolt3.There're 15 classes in my dataset,including
--large scales:roundabout,ground track field,soccerball field.They're large rectangles.
--small scales:planes,small veticles.They're Intensive small targets, neatly arranged.

Since objects of 15 classes have different scales,I try to cut the images to 416416 size,and many object boxes are splited.So CAN I CUT THE TRAINING IMAGES TO LARGER scales?I try 800800 but it seems that the input size should be 32*n??

If I want to train yolt3 with my dataset,should I must download the pretrain yolo weights in same input size?I visit https://pjreddie.com/darknet/yolo/ and find there're only 3 scales' yolo weights(320,416,608).So I must cut my images to (320,416,608) to match the pretrained weights?

Sorry but this is my first time using yolo, being explored and learning. I am trying to run yolt3 on my dataset. Also looking forward to some reply.：）