Prefers resize to trained dimensions about pytorch_retinaface

[biubug6]

When training, Retinaface and Faceboxes all adopt multi-scale training methods(correspondingly resize scale 640 or 1024). Tests often yield good results.

[xsacha]

I recommend this code in general to calculate resize:

        resize = 1
        resizedelta = 0.25
        while img.shape[0] * resize < (640.0 * (1 - resizedelta / 2) ):
            resize += resizedelta
        if resize != 1:
            img = cv2.resize(img, None, None, fx=resize, fy=resize, interpolation=cv2.INTER_LINEAR)

I use the same on FaceBoxes but 1024.0 instead.